No matter how much code agents write for me today, I’m still the directly responsible individual for every bug. So when things go wrong, I still have to crack open my laptop².

But unlike the old days, I don’t know every line of code! I’m more like a manager on call – yikes.

On call with AI: slower, more stressful

1. Get paged in Slack
2. Ack the page
3. Immediately copy+paste the page into my coding agent with “omg what is happening!!”
4. Agonize as it methodically investigates the issue
5. Steer random context into the chat as it goes

make no mistakes

Generally with enough tokens and time the agents resolve the outage, but:

• Outage comms are worse. I’m just pasting random hypotheses from my agent into the Slack thread. I don’t have as good of a sense of the ETA to resolution.

• The agent doesn’t work in outage mode by default. The default coding agent behavior is to investigate, then make a speculative forward fix. That’s the wrong approach during an outage! We should mitigate first with the lowest-risk change – roll back! turn off the flag! – rather than shipping more new code with less testing.

Source: Generic Mitigations by Jennifer Mace, O’Reilly Radar

After a few sweaty incidents, we’ve changed how we build to make being on call easier with agents.

Feature flag everything

The simplest tip I got from Alon to ship faster was to feature flag everything and test in prod. When starting a new project I build feature flag support right away. Feature flagging is an acknowledgment that I know so little about the code that I need a big red switch to manage it.

This was mostly a mindset change to remember to prompt “…and put this behind a flag”. Beyond that, a line in AGENTS.md like “we use PostHog for feature flags” and a file called features.ts are enough.

We don’t ask the agents to decide what to feature flag themselves yet. Do you? Should we?

Write a doc/skill explaining how you do release management

We built the Napkin Math iOS and Android apps in the YC Spring 2026 batch using Expo with React Native.

Expo’s magic trick is the “over-the-air update” – a React Native app is built with a native layer and an update layer of JavaScript and assets that can be updated over-the-air (OTA) without the user needing to visit the App Store.

Source: How EAS Update works, Expo Docs

OTAs can be super useful to mitigate an outage if done right!

After I explained this release management system to agents mid-outage a few times, it was helpful to just write a doc into the repo. I’d recommend you do the same! Everyone’s release process is a different special snowflake, so it’s probably not just In The Weights.

Recursive self-improvement: postmortems and runbooks

I confess: we haven’t done this one yet! Have you?

A few times now we’ve “fixed” a bug, only to have it crop up again. The agents fixed some narrower version of the issue or didn’t test the right thing before declaring victory.

In a repeat outage it’s stressful to watch the agent plod down the same investigation/discovery path, wasting tokens and time. I don’t want them to root-cause the same issue twice!

I’m thinking of adding a simple recursive self-improvement loop for the repo where we have:

docs
├── postmortems
│   └── 2026-08-18-photo-upload-failed-....md
└── runbooks
    └── expo-updates.md

I’m not sure what threshold for notability I’d use to warrant a postmortem or runbook except “when I say so”. But I’d want these to be primarily agent-readable (not explicitly designed to be human-readable) docs.

Is this cope?

How much should we change how we build software to adapt to coding agents, and how much should we just wait for the next model release?

My thought is:

• if it’s generically making the agent “code better”, it’s probably gonna be in the weights. That’s why I won’t write a generic “on-call” skill which would basically be “follow the Google SRE book”.
• if it’s about how your specific team/company does things, might as well make it legible to agents by writing it down

I still hold tightly (too tightly?) to some aspects of software development.

Maybe someday soon AI will replace my SRE job and we’ll delete all the docs and walk into the sunset building freely :)

Related reading

I read these after writing the above post. These writers are thinking along similar lines!

Compound Engineering: How Every Codes With Agents A four-step engineering process for software teams that don’t write code https://every.to/chain-of-thought/compound-engineering-how-every-codes-with-agents

AI demands more engineering discipline. Not less If you lived through the shift from handcrafted server pets to immutable infrastructure, you should sense something oddly familiar about what's happening now. https://charity.wtf/p/ai-demands-more-engineering-discipline

Stop trying to review AI's code faster: bet on rollback instead | Rootly https://rootly.com/blog/stop-trying-to-review-ais-code-faster-bet-on-rollbacks-instead

For example, I’ll go to develop on jep, and run into:

% npm run dev

> dev
> react-router dev

Port 3000 is in use, trying another one...
  ➜  Local:   http://localhost:3001/
  ➜  Network: use --host to expose
  ➜  press h + enter to show help

Hey! Who is using localhost:3000 ?!

How to use whoport

whoport is my new small, sharp tool to solve this. Pass in a port and it’ll tell you what command (and therefore what path -> project + worktree) is using it:

% whoport 3000
PORT   PID      COMMAND
3000   61981    node ~/conductor/workspaces/echo/rome-v3/target/debug/api

whoport can look up multiple ports at once:

Check it out here on GitHub 🐈‍⬛

GitHub - cmnord/whoport: Who's using this port? Who's using this port? Contribute to cmnord/whoport development by creating an account on GitHub. https://github.com/cmnord/whoport

After you’ve found the process using your port, you can obviously kill -9 it 🔪 or otherwise shuffle ports around.

Why? ps + grep falls apart with coding agents

My previous tool for this was good ol’ ps with some text searches. ps (“process status”) lists information about processes running on your computer. The typical command is ps aux to list all running processes. I used this in combination with a search tool like grep (or ag or, lately, rg) like so:

 Atuin v18.19.0                                                               <esc>: exit, <tab>: edit, <enter>: run, <ctrl-o>: inspect                                                        history count: 18631
  Search │ Inspect

  3 111ms   3mo ago ps aux | rg 'expo start'
  2 101ms   3mo ago ps aux | rg 'dev:'
  1 128ms   2mo ago ps aux | rg 'expo'
  > 123ms   2mo ago ps aux | rg 'expo '
 [     GLOBAL     ] 'ps aux'
 ps aux | rg 'expo '

shoutout to atuin

ps aux | rg 'npm ' worked when I had a hunch about the name of the command and I could search them. But now I have like 20–50 node processes running on my computer at any time…

% ps aux | rg 'node' | wc -l
      53
% ps aux | rg 'npm ' | wc -l
      28

translation: count how many running processes begin with 'node ' or 'npm '

Moreover, the commands agents kick off are completely unreadable…

% ps aux | rg 'npm ' | head -n 2
cnord            21752   0.1  0.4 446355296 149152   ??  S     8:02PM   0:00.22 node /Users/cnord/.nvm/versions/node/v22.22.3/bin/pnpm -C packages/editor exec vitest run --environment jsdom src/note/index.test.ts src/plugins/doc-change-listener.test.ts

↳ Okay, seems like it’s running Vitest, that’s fine.

cnord             6522   0.0  0.0 435341104   5136   ??  S     7:46PM   0:00.01 /usr/bin/open -n -W -o target/debug/example-dev.stdout.log --stderr target/debug/example-dev.stderr.log --env RANDOM_CONFIG_VAR=foo --env VITE_API_URL=https://api.example.com --env YOUR_SECRET_KEY=hunter2 --env YOUR_RANDOM_OAUTH_CLIENT_ID=hunter2 --env PATH=./my/whole/hecking/path/list/.bin:/......... target/debug/Example.app

↳ What in the world is happening in this 3,000-character command which has my whole env and PATH in it???

which of you is hogging my port?

whoport under the hood

whoport glues together ~two other tools: stalwart ps, plus lsof to list open files.

The core lsof command is:

lsof -t -nP -iTCP:"$port" -sTCP:LISTEN

Brief explanation of the command:

So with just this command we can get a newline-separated list of process IDs (PIDs).

whoport turns this into a comma-separated list to hand to ps.

Heroic ps returns, instead filtering by PID instead of search on the commands!

% ps -o pid=61981,command=COMMAND -p 61981
61981 COMMAND
61981 node ~/conductor/workspaces/echo/rome-v3/target/debug/api

So now I know that the rome-v3 echo worktree is using the port! And from there I can *ahem* terminate it or move aside. localhost:3001 isn’t so bad, anyway.

Filling out your tool belt

I keep little scripts like this in my dotfiles repo, which documents my favorite terminal commands and setup across laptops and machines. 🛠️

You should too! Some influential blogs on this topic:

• Your Dotfiles Are Meant to be Forked, Zach Holman
• Managing Your Dotfiles, Anish Athalye

Read more about tiny personal programs here:

• Some tiny personal programs I’ve written, Julia Evans

And to learn more about the command line, ps, etc., I highly recommend The Missing Semester of Your CS Education.

Happy coding :)

:earth_americas: Play a game: https://whatis.club

:computer: GitHub repository:

GitHub - cmnord/jep: 🌎 Jep! multiplayer trivia online 🌎 Jep! multiplayer trivia online. Contribute to cmnord/jep development by creating an account on GitHub. https://github.com/cmnord/jep

If you’re curious to learn more, this blog post explains further rules and how it’s built.

• What it does
  • Real-time gameplay
  • Upload your own games
• How it’s built
  • Room logic (state machine)
  • Storing room state
  • Timeout actions + serverless functions
  • Real-time updates
• Designing Jeopardy! without a host
  • Check your own answers
  • Say your guess, then check privately
  • Keeping up momentum
• Conclusion

What it does

Real-time gameplay

When you click on a clue on the board, it pops up on the screen for every player at the same time.

You all get a chance to read the clue, then race to buzz in. The fastest buzzer gets to answer the clue and gets points if they’re right. If they’re wrong, the buzzers reopen for all other players.

Upload your own games

I made a JSON file format so people can upload their own games. The format allows for an arbitrary number of rounds, categories per round, and clues per category. I hope this flexibility is fun for constructors!

The upload help page describes the format and has a sample file.

Biiig board

Games can also be unlisted or private so constructors can play their games or playtest draft games without sharing them widely.

Edit game visibility

How it’s built

Jep! is built in React using the Remix web framework. The backend is not a persistent server but rather a collection of serverless functions that respond to HTTP requests. All persistent data is stored in a database. The frontend uses websockets to subscribe to real-time updates to the database.

Room logic (state machine)

Choosing a game from the home page creates a new “room”. Every player joins the room from their own web browser. Each player sees the same initial room state (round not yet started, all clues are hidden, everyone’s score is 0, etc.).

Players then take actions in the room that modify the state. For example, a player chooses a clue, then buzzes in to answer.

The room state machine is implemented as a reducer in React.

Simplified room state machine.

Each room state also has a corresponding client-only mini state machine.

Simplified state machine for a client in the “reading clue” room state.

Storing room state

Instead of storing the room state itself in one database row, we store the list of actions that have occurred in the room.

• Reads: Fetch all the actions for the room, then apply each action one by one to the initial state to get the current state.
• Writes: When a player takes a new action, the client sends an HTTP POST request to the server, which appends it to the list of actions in the database.

This is called an event sourcing architecture. Writes are fast because they append to the list of actions without worrying about the current state or any other actions. It also makes it easy to replay the actions to compute the state at any point in time.

Each client has a local state machine which computes the current state from past actions.

Example actions table

const state: State = {
  type: RoomState.RevealAnswerToAll,
  activeClue: [0, 0],
  boardControl: 456,
  buzzes: { 123: 489, 456: 351 },
  isAnswered: [
    [{ isAnswered: true, answeredBy: 456 } /* ... */],
    // ...
  ],
  players: {
    123: {
      name: "player 1",
      score: 0,
    },
    456: {
      name: "player 2",
      score: 200,
    },
  },
  round: 0,
  // ... more state
};

State computed from example actions above

Timeout actions + serverless functions

Some actions aren’t really player-initiated, but have to be modeled as such because there’s no persistent server running on the backend. For example, what if no players buzz in within 5 seconds to answer a clue? There’s no server-side timer tracking this, so all the players would be stuck waiting for someone to buzz in.

To fix this, each client submits a “5 seconds elapsed” action after the timeout. Upon receiving any “5 seconds elapsed” message the state machine advances the room state to the next step.

Real-time updates

If Player 1 chooses a clue, how does Player 2 see that clue pop up on their screen at the same time? In other words, how do they read one anothers’ writes in real time without refreshing the page?

The client subscribes to real-time updates from the “actions” table. This opens a websocket connection to the Supabase realtime server. When the actions table receives a new row, the server broadcasts the row to all clients subscribed to that action’s room ID.

The client then receives that action and applies it to its local state machine. This way all clients have a consistent view of the room state.

Example use case from the Supabase realtime page

Designing Jeopardy! without a host

The biggest difference between the show and Jep! is that there’s no host to communicate what’s happening at any given time, keep up the momentum, and check answers. The first two can be addressed with good design, but the last one is trickier.

Check your own answers

I didn’t want to require one player be the host, or make any player check other players’ answers. I also didn’t want to try to make a ~virtual host~ for a couple reasons:

1. There are often multiple correct answers to a question, like “aubergine” vs “eggplant” or “Queen Elizabeth” vs “Elizabeth I”. I don’t have an automated way to check this for all questions.
2. I don’t want players to have to type, which means the virtual host would have to use speech recognition. I find speech recognition annoying and error-prone.

Having each player check their own answers keeps the game fast-paced and doesn’t put the hosting burden on one person in particular.

Groups also develop house rules for which answers are “close enough” to be correct. Self-checking helps account for more and less strict play styles. However, it’s not as suited to a competitive tournament setting.

Say your guess, then check privately

We keep players honest by having them say their guess out loud to others before checking it. This requires players to join a voice call or the same physical room together.

Here’s an example transcript:

PLAYER 1: Ferns for $600.

**Text: FERNS FIRST EMERGE IN THE FOSSIL RECORD IN THIS PERIOD**

PLAYER 1: [Buzzes in]
PLAYER 1: Hmm... what is the Jurassic period?
PLAYER 1: [Privately sees the answer]
PLAYER 1: Drat!
PLAYER 1: [Marks "incorrect"]

**Buzzers reopen**

PLAYER 2: [Buzzes in]
PLAYER 2: Ah, it's me! What is the Devonian period?
PLAYER 2: [Privately sees the answer]
PLAYER 2: Yep!
PLAYER 2: [Marks "correct"]

**Text: DEVONIAN**

Example game transcript with honest players

With this scheme, you can tell later if someone lied. If they said “Jurassic” was correct, you’d know they lied when the game later showed the answer was “Devonian”.

This design is clunkier than having a host: it’s hart not to blurt out the answer as soon as you learn you were wrong! It adds more cognitive burden during an emotional moment in the game.

Keeping up momentum

I tried to make the visuals as clear as possible, but it’s still sometimes confusing who’s up next to choose the clue or who won the buzz-in. This confusion slows the game momentum and keeps it from moving as snappily as a 20-minute TV show.

I also love using my phone for everything and tried to make the game mobile-friendly, but not everything can fit on phone screens smaller than mine. There’s also a design tension between showing all possible information and controlling what’s shown to keep the stakes high.

These problems are related! I definitely hit the limits of my UI design skills. I would love to get specific feedback or designs that would improve this. For now, I’m happy the game generally works and am ready to move on to other projects.

Conclusion

A major inspiration for Jep! was Down for a Cross, a collaborative crossword website. I wanted to make something similar for Jeopardy! to hopefully give back to the trivia fan community.

I implemented a minimal set of features from Down for a Cross, plus the Jeopardy! game logic. Some fun additions could be:

• Timing the game with a clock
• An in-browser puzzle creator instead of just JSON files
• Coryat score tracking
• …And more!

Thank you to all the playtesters who gave great feedback! Your insight into my fundamental design problems kept me awake at night. Thanks in particular to Alex, Jynnie, Tommy, and Neil, as well as the NCSJ, JJSR, MJKT+, and GACJI playtest groups.

I hope people enjoy playing Jep! and I’d love to hear any feedback or ideas for improvements. Play a game here and check out the code here!

Source: Fast (Motion) album cover + my edits

What is a sport?

For my purposes, a sport is a physical competition you can win or lose based on some skill.

This means non-physical skilled competitions like chess are not sports under this definition.

Leisure activities vs. sports

Many outdoor leisure activities are not sports. Kayaking alone is not a sport because you can’t win or lose. You would need to add competitive rules to make it into a sport, like speed kayaking.

However, the rules must be designed carefully to reward skill more than chance. Coin tossing would obviously not be a sport because it only depends on chance. If fishing is made into a sport via a fishing tournament, the rules should control for the unpredictable fish by evaluating fishers in the same area of a river for the same time period.

Under this definition, I claim that all sports can be categorized as Measure, Precision, Spectacle, Combat, or Play.

:runner: Measure

“Measure” sports are ones where the goal is to min/max a certain value. The most common measure sports are races. You win a measured sport by running the fastest, lifting the most, throwing the farthest, etc.

The objective is to get the record, and there is no “best” performance or ceiling other than the current world record.

Examples: track & field, car racing, running, cycling, rowing, bobsledding, speed-climbing, power lifting

:dart: Precision

Precision sports are like Measured sports, but with the goal of accuracy. The goal is to hit a target, not a min/max with no bound. You win a precision sport by being the most precise with your movement.

Examples: pistol, archery, darts, cornhole, golf

:dancer: Spectacle

Spectacle sports are subjectively judged and aesthetics matter. They often have panels of judges.

For example, figure skating is judged on “interpretation of music” among other more technical criteria. Almost any physical activity could be turned into a spectacle sport if a point system is designed around it to determine who wins and who loses.

This is not to say that all non-Spectacle sports are perfectly objectively scored. For example, many Play sports have referees who make subjective calls for fouls. However, individual judges have fairly wide subjective leeway in Spectacle sports. Would a “perfect” judge with all information always make the same calls?

Because Spectacle sports require subjective judging, casual enjoyment of the activity is not a sport. Skating in Central Park with your friends is not a sport unless you count how many triple lutzes you each do.

Examples: cheerleading, figure skating, gymnastics, artistic (synchronized) swimming, skateboarding, parkour, surfing

:muscle: Combat

Combat sports encompass most martial arts or fighting sports, where you win when the other person loses. They involve some direct physical strength contest or grappling, so fencing is a combat sport but tennis is not. These are often 1-v-1 sports, except maybe tug-of-war.

What makes fencing a Combat sport and tennis a Play sport? There’s something visceral about Combat sports, like they are one layer closer to actual fighting. Please provide a counterexample if you disagree.

Examples: boxing, fencing, judo, taekwondo, karate, arm wrestling, tug of war

:soccer: Play

Play is the leftover, catch-all category which encompasses all sports that don’t fit the other definitions. The top 9 most popular sports in the world are all Play sports, and they are perhaps the most natural expression of “play” even as kids. You can complete the sentence “I play ___” with any Play sport.

Play sports are hard to define on their own and not in contrast to the other four sports categories. Most Play sports have more complex rules than Measure, Precision, Spectacle, or Combat sports. Play sports are often, but not always team sports — both singles and doubles tennis are Play sports.

Examples: basketball, football, soccer, tennis, cricket, hockey, volleyball, ping pong, baseball, rugby

Limitations

This blog post was a fun thought exercise for me, but I’m curious what others think. Would you categorize anything differently? Would you propose a new category? Do you have a better definition for any category?

Photo by Ancaro Project on Unsplash

Object detection algorithms can solve problems difficult for humans, like identifying vehicle models from satellite images, counting the number of people in a crowd, or rejecting unripe fruits from a conveyor belt of fresh produce. The last example is what’s known as a sporadic visual search problem, where a huge stream of repetitive data contains anomalies only infrequently and sporadically. Humans struggle with these types of problems due to cognitive fatigue. Object detection algorithms, however, excel at this.

A perfect example of the sporadic visual search problem is X-ray baggage screening at museums. At the entrance to a museum, security operators may scan passengers’ bags with X-ray machines to detect threats like knives or handguns. Baggage screeners see thousands of bags a day, but still need to identify the single bag that contains a threat.

An X-ray scanner at the entrance to the Louvre museum.

At Synapse Technology, we use an object detection model to tackle this problem. A model in this context is an algorithm trained on a large number of labeled examples to solve a task. Let’s train an object detection model, then run it against each X-ray image the bag screener inspects.

A straightforward example: the knife stands out in this bag.

Detection can get much more difficult, though, depending on the bag contents and arrangement.

This bag contains a knife. Can you spot it?

It’s hard to tell whether this bag contains a threat, even with a very good object detection model. That’s why we use multiple models in ensemble to do our detections. If two models both detect the same object, there’s a much higher chance that something is actually there. Using multiple models can also improve detection rate and lower false alarms by making the whole system more robust towards uncertain/difficult cases. Ensemble detection also allows us to compare different model architectures against each other to continuously improve our product. Finally, it allows us to scale the number of threat classes we support: each model supports some set of threats, and ensemble detection makes it easy to incorporate a new model’s set of threats into the ensemble.

Implementing ensemble detection presents additional challenges beyond single object detection. Here, I’ll describe how we approached these challenges.

How can a multi-model ensemble be better than a single model?

Suppose we have two object detection models, Model A and Model B. Each of them has detected some objects in the image, and each detection has an associated confidence score.

The two different models have detected a threat. Note that is the same image as the previous one.

Model A (red) and Model B (yellow) have both detected a knife in the top right corner of the image, but neither model has a high enough confidence score alone for us to conclude that there is a threat in the image. If we consider the models as an ensemble, though, we observe that since two of their detections overlap, they probably refer to the same object. Given that they refer to the same object, there is a much greater probability that something is actually there. For example, if the confidence scores are the probability that each detection is a true positive (rather than a false positive), then the probability that the ensemble is a true detection is 0.7, which is greater than either detection alone (see calculations below).

prob_threat_modelA = 0.4
prob_threat_modelB = 0.5

prob_nonthreat_ensemble = (1 - prob_threat_modelA) * (1 - prob_threat_modelB)
#                         = (1 - 0.4) * (1 - 0.5)
#                         = 0.6 * 0.5
# prob_nonthreat_ensemble = 0.3

prob_threat_ensemble = (1 - prob_nonthreat_ensemble)
#                      = 1 - 0.3
# prob_threat_ensemble = 0.7

Ensemble detection can improve the outcome of this example, but also even more complex examples. Let’s formalize the problem and address further extensions.

Input: For a single image, each detection model independently returns a set of inferences, which are composed of a bounding box, a label, and a confidence score. The detection model will only generate one inference for each object it thinks is in the image. In the image above, Model A has returned one inference and Model B has returned one inference. Model A’s inference has the label “knife” and a confidence score of 40%, meaning that it’s 40% sure that the inference is correct.

Output: After ensemble detection, return a list of inferences, one per perceived threat in the image (no disagreements or redundancies). These inferences must have high confidence scores.

How can you implement multi-model ensemble voting for object detection?

Let’s consider a hypothetical example with the same two models, but more inferences:

Model A and Model B have detected more objects in this hypothetical image. Some of these perceived objects overlap.

In this case, we can’t just group overlapping inferences as one “ground truth object” like in the first example because two of Model B’s inferences overlap. Model B thinks that there are two different overlapping objects in the lower half of the image, but Model A thinks that there is one. How do we group them? What if there were even more inferences than this? Let’s model this as a graph problem.

Graph model for inference cliques

Create a vertex for each inference. Let the vertex’s color indicate the model it came from (e.g. red for Model A, yellow for Model B). Draw an edge between two inferences if their bounding boxes overlap. Weight this edge by how much they overlap, in particular, their intersection over union (IoU) score.

A graph representation of our inferences, where weighted edges join nodes whose bounding boxes intersect.

Now we have a weighted undirected graph. Our goal is to partition each disjoint subgraph into cliques such that no two vertices in a clique share the same color. We also want the cliques to have the maximum sum of edge weights (be as overlapping as possible). We take a greedy approach to this. First, we create a clique for each vertex. Then, we sort the edges by edge weight (IoU score). In descending order of edge weight, merge the two cliques the edges would connect if their sets of vertex colors are disjoint. A pseudocode implementation of this is below.

# Create a single-element clique for each vertex
vertex_to_clique_map = {v: clique(v) for v in graph.vertices}

edges = []

for every pair of colors color1, color2:
    for every pair of vertices from color1 and color2 v1, v2:
        # if v1 and v2's bounding boxes overlap, they have an edge between them
        if v1.bounding_box.overlap(v2.bounding_box):
            edges.append(edge(v1, v2))

# sort edges by IoU score, descending
edges = sorted(edges, key=lambda edge: -edge.iou_score)

for edge in edges:
    clique1 = vertex_to_clique_map[edge.v1]
    clique2 = vertex_to_clique_map[edge.v2]
    # if no 2 vertices in the cliques share a color, they can merge
    if clique1.colors.isdisjoint(clique2.colors):
        clique1.update(clique2)
        vertex_to_clique_map[v2] = clique1

# return the remaining cliques
return set(vertex_to_clique_map.values())

We’ve grouped the detections into three cliques. The clique on the left contains two inferences.

Voting on cliques

Now we have cliques of inferences that each refer to one object in the image. Which of these cliques should we accept? The fact that only Model A detected an object near the top of the image (Clique 3) means that Model B did not think there was an object near the top of the image. Here is where we can introduce voting to increase the accuracy of our ensemble detection. A model votes for a clique if one of the inferences in the clique comes from that model, and only accept a clique if majority of voters vote for it. Let’s define a voting threshold as the minimum percentage of models that must vote for a clique for us to accept it.

Suppose we had a voting threshold of 0.5, so both models need to vote for a clique for us to accept it. Then we’d reject Clique 2 and Clique 3 because they each only have one vote, but we’d accept Clique 1 because it has two votes.

Results of ensemble detection after voting: only accept Clique 1.

After voting, we only accept Clique 1. We finally compute the combined score of Clique 1 as a function of its inferences, and output a representative bounding box for this clique with that combined score as its confidence score.

Takeaways

Through effective grouping and voting schemes, multiple object detection models in ensemble can perform better than one alone. In the context of X-ray baggage screening this means more accurate detections, lower false positive rates, and ultimately a safer world.

We hope other computer vision teams can apply this technique to make more accurate and robust detection systems.

Some cool detections we’ve made at Synapse Technology, including handguns, gun parts, ammunition, razor blades, and fixed and folding knives.

I designed and implemented this system over the course of my one-month internship at Synapse Technology (acquired) this January. Thanks to my mentor Steven for thoughtful design documents and whiteboard sessions. Thanks to Sims for encouraging me to share my work, as well as the rest of the engineering team for a fun and productive internship.

Thanks to Noah Moroze, Simanta Gautam, Bruno B. Ferrari Faviero, Ian Cinnamon, and Sanjeevani Lakshmivarahan for feeback.

I broke my jaw at the end of August by standing up and fainting—way to go, right? I got up out of bed in the morning, felt dizzy, blacked out, and fell. I woke up within a few seconds and noticed that my jaw felt out of place. The left side felt stuck higher up, so the right side of my bite didn’t close. I also had a scrape on my chin and my ear was bleeding.

My mom rushed me to the dentist, where they did a panoramic dental x-ray to find that my jaw was fractured in two places.

The original panoramic x-ray result. The white line across my lower middle teeth is my permanent retainer!

The two fractures: the body fracture on my right and the condylar fracture high up on my left.

The doctors identified two fractures: one along the body of my jaw and one high up on the left condyle [diagram] and recommended I go to the hospital for surgery. This blog post recounts my experiences from injury to full recovery, and is basically everything I wish I’d known before it happened.

• The Hospital
• The Surgery
• Immediately Post-Op
• 1 Day Post-Op
• 2–3 Days Post-Op
• 4 Days Post-Op
• Week 1
• Week 2
• Weeks 3 + 4
• Weeks 5 + 6
• Arch Bar Removal
• Weeks 7+
• Why did I fall?

The Hospital

After I was admitted to the hospital I waited and napped for a long time. It was overall fine and relaxing, but my IV for fluids was annoying to move around with. Throughout the day I met with various doctors and nurses: oral & maxillofacial surgeons for the surgery, and internal medicine doctors to investigate why I fell.

I could talk alright before the surgery, but couldn’t open my mouth very much, chew, or brush my teeth. I never felt much pain while waiting — I didn’t really feel the fracture along the body at all, and the condylar fracture only hurt when I tried to open my jaw wide. The doctors and nurses were very concerned about my ear bleeding, but that turned out to just be a surface cut. I had a CT scan at the hospital to better understand the injury and decide which surgery to do.

The Surgery

For the fracture on the jaw body, the doctors decided to do an “internal fixation”, where they put a small titanium plate across the break to fixate it. The plate looks like this:

The plate goes from a bit before the corner of my jaw on the right to a little past the curve at the front of my chin. The plate is 10cm long (about the length of my index finger), 1.7mm thick, and held on by 7 titanium screws. The plate had to be so long because the fracture was sort of parallel along the bone. It’s made to stay in for life.

With the condylar fracture, the head of the condyle actually broke off of the bone and is “floating” in that area. Because it’s so high up in the joint, the surgeons decided not to open up the joint area and fixate it with a plate — it’s an invasive surgery and could have caused problems with articulating my jaw. Instead, they decided to temporarily lock my jaw shut with arch bars and elastic bands.

Arch bars (source: AO Foundation)

Arch bars are like a row of braces above your teeth with little metal hooks like a coat rack for hooking the elastics. The bars are held on by wire twist-ties “lassoed” around each tooth. They decided to use elastic bands instead of wires because my jaw could have locked if the condyle healed without any movement. The elastics were moved and adjusted as I healed, allowing for more and more guided movement.

To put in the plate, the surgeons made an incision inside my lip (like if you pulled down your lower lip all the way, and cut at the line between lip and gum). They stitched it up with dissolvable sutures.

For the surgery itself, I changed into a hospital gown and socks, peed extensively so I wouldn’t need to during surgery, and talked with the main oral surgeon and anesthesiologist. They gave me the KO medication through my IV and I was completely out for the 3–3.5hr surgery.

The plate and arch bars after the surgery.

Immediately Post-Op

I was SUPER PUFFY and could not feel most of my lower lip and jaw area, but still wasn’t really in any pain. I couldn’t move my jaw or open my mouth to speak. My throat was a bit scratchy because they had to put a tube up my nose and down my throat for anesthesia. I drank chicken broth and rested for the rest of that day and night.

1 Day Post-Op

I was still just as puffy the day after the surgery. I kept drinking water and broth and getting IV fluids. I also showered, which was a bit scary because I was afraid of falling again (except naked this time), but it went OK. I used Vaseline to keep my lips from getting chapped for the next month, because I couldn’t open my mouth to lick my lips.

I was given 5 medications:

• Antibiotic (amoxicillin) every 8hrs for a week. Sweet pink liquid.
• Prescription painkiller (hydrocodone) as needed. I only took this once, and just used ibuprofen from then on, because I wanted to avoid opioids.
• Non-prescription painkiller (children’s ibuprofen) as needed. I took this every 8hrs with my antibiotic. White/clearish and sweet, but stung my throat a bit.
• Strong mouthwash (chlorohexidine) rinse after every meal because I couldn’t open my mouth to brush my teeth. Medical-grade alcohol-free mouthwash.
• Anti-nausea and anti-constipation meds to combat possible side effects from the hydrocodone. I never used these.

At the end of the day I finally had my IV taken out and went home!

2–3 Days Post-Op

For the next few days I just took medicine and rested. I used a bed wedge at home to rest with my head elevated and ordered one for school too. I could only drink extremely smooth things because I could only sip through the space between my upper and lower teeth. Example liquids:

• Juice
• Cream of {Chicken, Mushroom, etc.} soups (strained)
• Broths
• Ensure/Premier Protein
• Smoothies (strained only, but this was still very difficult and didn’t really work)

I flew to Boston for my fall semester of school 3 days after my surgery. The flight was uneventful — the TSA inspected my ice packs, but I was convincing enough that I needed them (I was puffy as heck, couldn’t speak, and had a plastic bag full of medicine!).

4 Days Post-Op

I was fortunate to be able to stay with my boyfriend’s family in the Boston area for a day before moving back into my dorm. At first I used a text-to-speech app to talk, but then I started to become able to speak in a garbled voice without moving my jaw or lips much.

I had a lot of angst over meeting my boyfriend and his family because I was really embarrassed with how I looked. He, his family, and all my friends were really kind and supportive, though, and it meant a lot to me.

I was still very puffy, but my chin scab started to heal. The skin on my face peeled a little (making room for puffier skin), and so did the skin on my lips. My mouth felt fuzzy and gross from not brushing… pretty much from then until I could brush my teeth again.

Week 1

It was very difficult to talk. My mouth had a bit of a lingering taste of blood, and my spit was always red/pink. Keeping up with antibiotics and pain meds every 8 hours was annoying, but I’m glad I was never really in pain.

Once I was in Boston, I made weekly appointments with Dr. Somi Kim at Massachusetts General Hospital (MGH). Dr. Kim was really nice! She adjusted the locations of my elastics so that my bite was back to where it naturally rests, which was a huge relief. It had been a bit out of place since the surgery, with my bottom front teeth pushing on my top front teeth. She also told me not to do any contact sports for 3 months, and that I was absolutely not allowed to chew for the next 6 weeks.

The most surprising thing she said was that it’s possible to remove the plate! I had expected to have to keep it in forever, but it can be removed after 6 months. It’s still worth a discussion whether the removal is worth it.

As far as swelling, Dr. Kim said that 50% of the swelling goes down in the first 6 weeks, and the next 50% goes away in the next year. For me most of the swelling went away after 2.5 weeks, though.

Me (left) struggling to smile normally at HackMIT 2017 (2.5 weeks post-op). You can see my arch bars peeking out, and my right side is still a little puffy.

My lower right lip and cheek near the plate were numb for a long time after surgery because the nerve that passes through that area was skeletonized. I was still pretty numb even after the swelling went away. At 3 months post-op I had about 90% of the feeling back, and at 4 months now I have 95%+ back, but I was worried at the beginning because the numbness receded so slowly.

Week 2

During week 2 talking got easier, but I still couldn’t move my jaw at all. The bruising/swelling started turning yellow. One of my elastics broke at this time, and I was TERRIFIED of putting it back in. I was afraid of tearing my stitches, and didn’t know exactly which hooks the elastic was supposed to go on. It took me like 30 minutes and I almost cried from stress. I got way better at putting these back on by the end.

At this week’s doctor’s appointment I was put in lighter “guided elastics” that let me open my mouth. I had an elastic band on each side of my mouth in the following positions:

I remembered the positions as 5–2–3–4, like room 5-234 for HackMIT meetings!

I had to take the elastics out 3 times a day and stretch open my jaws as far as I could. This sounded scary based on my last elastic replacement experience, but it was way easier when my stitches were healed (at around 3.5 weeks in!). I didn’t really get around to doing the jaw exercises 3 times a day, but I did do 2 times (morning and night). I also stopped taking ibuprofen for pain around this time.

Guiding elastics in my mouth IRL.

Other miscellaneous notes:

• Using a waterpik would have been fine as long as I didn’t blast my stitches. I was ultimately too afraid of this to use it, though.
• Sleeping on my side was fine as long as it didn’t hurt (ordering a bed wedge was a bit of overkill I guess, haha). I was very concerned about my bite alignment slipping out of place when I relaxed, but Dr. Kim reassured me it would go back to normal (and it did). The condyle on the left side is slowly regrowing to match the right, so over time it’s becoming more and more even.
• I read online that chlorohexidine (the strong mouthwash) could stain your teeth, but Dr. Kim reassured me that a regular cleaning at the dentist’s would fix it. I didn’t end up having noticeable stains. I switched to a regular alcohol-free mouthwash when I ran out of chlorohexidine.

I only had trouble sleeping for one night in the whole recovery period, 13 days post-op. For some reason my jaw just hurt a lot being in the aligned position, but I was afraid to let it rest in the askew position while I slept in case it stayed there. I just took some ibuprofen and stressed myself out until I fell asleep at 4am. My worry ended up being unfounded — my bite got more and more even, and I had a comfortable and fairly even bite by week 4.

Weeks 3 + 4

By week 3 I looked pretty much normal and got a lot better at speaking without moving my jaw. I spoke on the mic in the arena at HackMIT without sounding very distorted, and I didn’t need to preface conversations with new people by explaining that I broke my jaw. I just looked like a slightly less symmetric version of myself with weird braces.

Having room to open my mouth also made eating so much easier! My eating pattern was to take miscellaneous smoothie ingredients in a green box from the dining hall up to my wing’s main lounge, blend them, and supplement the mix with Ensure or Premier Protein. I got tired of Ensure’s distinct flavor (vanilla tootsie rolls) and started drinking Soylent, which I liked a lot.

Throughout all of this, eating took much, much longer than usual. Drinking food instead of chewing it probably takes twice as long.

Another breakthrough I had during this time was to use orthodontic wax! Before I used it I couldn’t really laugh or smile without being lacerated with all the wires. The wax made my mouth bulkier, but was definitely worth it.

Maybe my favorite milestone of week 4 was being able to take out my elastics and brush all of my teeth with a manual toothbrush!! :heart: :heart: It felt SO GOOD.

Weeks 5 + 6

I only had to wear my elastics at night for these last 2 weeks. I was told to do jaw exercises more frequently (5x/day for 5min each), but I didn’t remember this very well and just tried to practice whenever I remembered.

It was a bit tough to remember not to chew while my mouth was totally free, but I stuck with it. I could mush things in my mouth like oatmeal, cream of wheat, chili, soup, cottage cheese, soft pasta, and scrambled eggs. I got pretty good at swallowing small soup ingredients like kidney beans or chickpeas without chewing them. Instead of going up to the main lounge and blending things, I would drink cups of soup and milkshakes in the dining hall.

I started using my electric toothbrush and flossing too, which felt amazing :heart:

Arch Bar Removal

I was so excited to have the arch bars gone!! There were 3 possible anesthesia methods to remove my arch bars:

• Swish lidocaine in my mouth: tastes bad, but it’s easy. Maybe not as numbing as the other options.
• General anesthetic: going totally under. It’s kind of a big deal and takes a lot of time for how short the operation is.
• Injections: I’d need 7–8 shots to numb enough of the lip area, so it’s almost as much work and pain as the removal itself.

I decided to go with the lidocaine. I was worried because Dr. Kim said it was “uncomfortable”, but it was really fine. It took about 10 minutes, and felt like… getting wires pulled out from between your teeth, which is what it was. She untwisted the wires, snipped them short, then pulled them around and out. I grimaced for a few of the pulls, but it was fairly quick and I really wanted it all gone.

Imprint of the removed arch bar!

Weeks 7+

My first meal after I could chew was a super lame sandwich. :stuck_out_tongue: Chewing felt weird, but good. Putting my clear retainer back on my top teeth hurt really bad at first, but it did fit and went back to normal.

Food gets stuck in the lower right side of my mouth sometimes, probably because the stitches made the pocket a bit smaller. It’s a bit annoying but I think it’s getting better, or I’m getting used to it. I had to work up to chewing really crunchy things, but it only took maybe a week.

At my last dentist appointment everything checked out, and I feel 100% normal now :blush:

Why did I fall?

Tests in the hospital ruled out any serious brain or heart issues. I still feel faint sometimes when standing up quickly, but I’m careful now to wait a bit before getting up, or I just sit down on the floor carefully as I start to feel it. I feel more dizzy when I’m sleep deprived, so I think I fainted because I went to bed too late the night before I fell. Let this be a warning that getting good sleep is extremely important, and lack of sleep can have drastic effects on your health and body! I’m still learning this, and feeling the plate under my skin is a firm reminder to keep improving and building good habits.

I’m very thankful to Dr. McLean at the dentist’s office for taking my x-ray on such short notice, Dr. Pokorney and the Beaumont OMFS staff for doing my surgery so well, Dr. Kim at MGH for being so patient with all my followup questions, my boyfriend and his family for hosting me in Boston while I recovered, my friends for all being so accommodating and understanding, and of course my family for helping me through everything and keeping me company in the hospital.

When I first got hurt, I worried a lot about what was going to happen and if I was ever going to be okay again. Reading about other people’s experiences helped me get a feel for what I should expect, so I hope that someone can benefit from reading about my experience.

Thanks to Noah Moroze and Kimberli Zhong for feedback.

HKN eligible Stef Ren ’19 reviews a student’s resume.

What is HKN Tutoring?

Tutoring is the second-largest branch of HKN after the Underground Guide, with 83 student tutors in fall 2017. Tutors for HKN submit the classes and number of hours they’re willing to tutor, and they are manually matched with tutees by the HKN tutoring officers. The tutor and tutee typically meet for an hour once a week, discussing anything from pset questions to exam practice problems to topics from lecture in more depth.

The most popular classes for tutoring are large classes like 6.006, 6.046, 6.036, and 6.004. 205 students were matched with tutors this semester, which means over 20% of students in the department are involved in HKN tutoring (source). HKN member Justin Xiao ’17 discusses tutoring statistics in more detail here.

Why request a tutor?

Office hours for these large Course 6 classes tend to be very crowded, which means TAs have to focus on immediate needs like psets or exams without reviewing concepts. Unlike office hours, lectures, or group psetting, tutoring time is solely focused on you and your understanding of the subject. You’re paired with the same tutor all semester, so you can build a personal relationship with them beyond transactional Q&A. Your tutor knows you, and can encourage you and adapt to your learning style.

Tutoring can be especially helpful for classes like 6.036, which switched to a flipped classroom style for the first time this semester. HKN eligible Sooraj Boominathan ’19 notes that “[tutoring] is the first time students might be getting the subject explained to them by a real person, and not just a video.”

Why tutor for HKN?

“It was p fun” ~Anthony Liu ’19

Aside from being paid $15/hour by the EECS department or earning HKN volunteer hours, tutoring is a great chance to sharpen your understanding of past class material, meet new people, learn about other MIT students’ study habits, and give back to the Course 6 community. HKN eligible Anthony Liu ’19 says that tutoring 6.046 helped him keep up with algorithms, which was good for interviewing. “My tutee would usually have some topics in mind and I would explain them, like a super personal recitation”, Liu says. “Before tests my tutee often wanted to do an extra hour, and we’d do more problems. I’d give test taking advice too — it was pretty fun.”

If you’ve considered TAing a class, tutoring is a great lower-commitment way to try it out without being responsible for course management (grading, writing problems, discipline/fairness, etc.). In fact, Sooraj ’19 decided to TA 6.036 because of how much he enjoyed tutoring. “I think I’ve enjoyed teaching/explaining things for a long time now, and I like helping friends out with psets on classes I’ve taken before”, he says. Previous tutoring experience can also strengthen your TA application.

In Summary

Peer teaching is a key part of the MIT undergraduate experience! Take advantage of this opportunity and sign up to tutor or be tutored here. Tutors are in high demand, so be sure to sign up early if you request tutoring for a popular class.

If you want announcements for HKN events, add yourself to hkn-interest@ here. If you have ideas for how to improve tutoring or anything else Course 6, contact HKN at hkn-officers@mit.edu, or the tutoring officers in particular at hkn-tutoring@mit.edu.

Thanks to Nalini Singh and Noah Moroze for feedback.

The goal for this tutorial.

The Samsung Gear S3 uses an open-source operating system called Tizen that’s based on the Linux kernel. You can program apps for the watch in C (“Native”) or with HTML/CSS/JS (“Web”). My project needs fast real-time signal processing, so we chose Native.

The source tutorial for this post is “Creating Your First Tizen Wearable Native Watch Application” by Tizen Developers. The info below should fill in all the gaps in the tutorial.

This project is up-to-date as of 20 March 2018. It uses Tizen Studio 2.2 and Wearable 3.0.

Installing Tizen Studio

1. Make sure you have enough space to download and install the Tizen Studio IDE. It’s ~5GB!
2. In the Tizen Studio package manager, install everything for the 3.0 wearable. This was my first big stumbling block. Also be sure to install the Samsung Certificate Extension and the Samsung Wearable Extension under “Extras”.

Once you’ve installed Tizen Studio, you can go through the “Creating a Project”, “Managing the Application Configuration”, and “Building Your Application” sections of the tutorial.

Running the Emulator

The emulator is basically a little watch that runs on your computer. It takes a long time to start, but you can tell it’s successfully starting up by the boot messages at the top of the emulator watch face.

This is a good sign.

Start the emulator from the Tizen Emulator Manager window. You know it’s ready when it shows up in the Tizen Studio Connection Explorer window. If you have issues running the emulator on OSX, you may need to install Intel HAXM and follow these instructions to install it. TL;DR: run sudo kextload -bundle-id com.intel.kext.intelhaxmand then restart Tizen Studio. After this, you need to set up security certificates.

Security Certificate Profiles

This is the mistake that took me the most time to figure out, and it is completely glossed over in the tutorial. Essentially, you need to go to Tools>Certificate Manager and click “+” to make new certificate. You have to make a Tizen certificate for the emulator to work, and a Samsung certificate for the app to run on the actual watch. For the Tizen certificate, choose “Partner” privilege level.

You need the one on the right too!!

It should be fairly straightforward to follow the steps to set up both certificates from there. You can follow the rest of the Tizen Developers tutorial to run the app on the watch emulator with the Tizen certificate profile.

Huzzah!

Running on the real watch

Follow along with the tutorial section “Running on a Target Device”. In particular, be sure to turn on the device WiFi and Bluetooth, connect your computer to the same WiFi network, and make sure both your laptop and device have the correct date and time. Also be sure to check your device IP address each time you connect to WiFi, because it may change.

When I ran ./sdb connect <watch IP> I never actually got a notification on my watch saying “Allow Gear to read log data, copy files,…”. I just got a notification that said “RSA key fingerprint:…” and clicked the check mark. I had to reset my watch once to get a notification to pop up at all.

Like it says in the tutorial, you know the watch is correctly connected when you see it in the Tizen Device Manager window. When I connect via the command line interface it says device unauthorized. Please approve on your device. each time, but it still works :sweat_smile:

Another thing they don’t mention in the tutorial: you may need to copy the device-profile.xml file to the watch. This file is located at /home/cnord/SamsungCertificate/my\_certificate\_name/device-profile.xml for me and needs to be copied to home/developer/ on the watch. Copy it over to the watch by right clicking on the device name in Tizen Device Manager and selecting “Permit to install applications”.

Now, you should be able to select the Samsung security profile, then continue the tutorial with Run As>Tizen Native Application, select the watch face, and see your beautiful new app!

Feel free to email me (cnord@school) if you have any questions, comments, or tips about Samsung Gear development.