KaitlynChen

🏆 1st place at Amazon's internal hackathon, out of 190+ engineers competing. Cleared a cohort-wide peer vote, then a panel of Amazon L8 directors named ours the best build in the field.

Amazon flies in hundreds of interns to cities where they know nobody, then seats them on heads-down teams for ten weeks. We surveyed 45+ of them first: 70% had never met peers sitting nearby, and 58% ate lunch alone weekly. That pointed the work at the last mile: not scoring matches, but getting two people to actually sit down.

• A 16-question survey feeds a recommender that scores by Gower similarity and auto-weights each question by its entropy, so questions everyone answers identically carry no signal.
• Gumbel-top-k sampling and MMR re-ranking keep your slate genuinely varied instead of five clones of you, with your own team hard-gated out.
• Serverless on real AWS behind internal SSO, not localhost: Lambda, API Gateway, single-table DynamoDB, Bedrock icebreakers, and a 93-chunk RAG onboarding bot guaranteed to cite a real doc.
• Mine specifically: the two-sided invite protocol, with strict sender/recipient role separation and cross-device sync so neither side can accept its own invite, plus read receipts, calendar booking, and the match-filtering layer.
• After the win I added crowdsourced map spots, geocoded server-side and signed with SigV4 I hand-rolled in pure Node crypto, because the build sandbox blocks network access and the AWS SDK client wasn't available.

→ 1st of 190+ engineers. 50 of the team's 109 commits and 138 of its 157 tests, the most of anyone on either count.

Face recognition that knows when it's being fooled: a liveness model that tells a real face from a photo or a phone screen, and gates recognition behind it.

I wrote the first version as a webcam face-recognition tool and it worked, until I held up a photo of myself and it greeted me by name. The matcher wasn't wrong. "Is this the right face" and "is there a person here" are different questions, and answering only the first produces a system that feels secure and isn't.

• Trains a small PyTorch model to detect presentation attacks, scored with the ISO/IEC 30107-3 metrics the PAD literature reports: APCER per attack type, BPCER, ACER, EER.
• Liveness runs before recognition, so a rejected face is never embedded, enforced by a test with a call counter. The UI never prints "SPOOF, matched Kaitlyn", which would tell an attacker their spoof found the right target.
• Trust fails closed: with no liveness model loaded it reports UNVERIFIED, never TRUSTED, because a check that defaults to pass is worse than no check at all.
• A train/serve preprocessing mismatch was enrolling on an aligned crop and matching on an unaligned one, so embeddings landed in different regions of the space and matching degraded silently, with nothing ever erroring.
• --seed silently didn't work: augmentation built an unseeded RNG per call, making every metric unverifiable, including to me. Now seeded from (seed, epoch, index), so two runs produce bit-identical weights.
• dlib is gone, so there's no CMake or C++ toolchain to install, and CI proves a clean install across Linux/macOS/Windows × Python 3.10 and 3.13.

→ Liveness at 0.44 ms median on a 1.0 MB, 226K-param model, about 2% of the frame budget. 5,600 lines, 304 tests. Accuracy on real PAD data is deliberately unpublished: the synthetic fixture is trivially separable and reports 0.00% on everything, so quoting it would be meaningless.

Data-driven steering controller for an autonomous vehicle, built on Gaussian Process regression.

Trained a GP model on a 678k-point system-identification pipeline to control steering, and used the GP's own variance estimate as a live fault detector, flagging when the model was operating outside its trusted region.

→ 0.23 m RMSE tracking error.

Offline speech-to-text and translation across 20+ languages, found and adopted by strangers on the open-source ecosystem.

Transcription and translation that runs entirely on your own machine: no cloud, no API keys, nothing leaving the device. I open-sourced it under MIT expecting nobody to notice, and people found it on their own.

• 15 stars and 5 forks from developers I've never met, none of them classmates, teammates, or anyone I asked.
• Users I don't know opened issues asking for features, which meant maintaining something real: triaging requests, deciding what belongs in scope, and answering people depending on it.
• Runs fully offline on Vosk for speech recognition and Argos for translation, which is the point: it works on a machine with no internet and sends nothing anywhere.

→ Open-sourced under MIT, adopted by real users with zero promotion.

Open-source tool that diagnoses why ML models break or slow down when deployed to edge hardware.

Profiles a model against a target device and surfaces the real bottlenecks (quantization mismatches, unsupported ops, memory pressure) instead of leaving you to guess.

Full-stack LLM app serving generated content with safety guards.

FastAPI + PostgreSQL backend serving LLM-generated content, with structured-output validation and fallback guards so a bad model response never breaks the experience.