KaitlynChen

🏆 1st place at Amazon's internal hackathon — 190+ engineers competing, cleared a cohort-wide peer vote, then a panel of Amazon L8 senior leaders 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 built the bridge, and deployed it on real AWS behind Amazon's internal SSO instead of leaving it on localhost. We surveyed 45+ interns first: 70% had never met peers sitting nearby and 58% ate lunch alone weekly, which 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 candidates by Gower similarity, auto-weights each question by its entropy so questions everyone answers identically carry no signal, hard-gates out your own team, then applies Gumbel-top-k sampling and MMR re-ranking so your slate is genuinely varied instead of five clones of you. Serverless throughout: Lambda, API Gateway, single-table DynamoDB, a Bedrock foundation model writing icebreakers, and a 93-chunk RAG onboarding bot on Titan embeddings that's guaranteed to cite a real doc. Mine specifically — the two-sided invite protocol (strict sender/recipient role separation and cross-device sync, so neither side can accept its own invite), read receipts that only fire when a human actually opens the thread, the calendar-booking flow, searchable org/city/building pickers, the match-filtering layer, and 138 of the 157 tests. After the win I added crowdsourced map spots: server-side geocoding via Amazon Location Service, 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, behind an admin approval queue so nobody can invent a place.

→ 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 just different questions, and answering only the first produces a system that feels secure and isn't. So the project changed shape. It now trains a small PyTorch model to detect presentation attacks, scores it with the ISO/IEC 30107-3 metrics the PAD literature reports (APCER per attack type, BPCER, ACER, EER), and reports a face as trusted only when it is both recognized and verified live. Liveness runs before recognition, so a rejected face is never embedded — enforced by a test with a call counter — and 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 the tool reports UNVERIFIED, never TRUSTED, because a liveness check that defaults to pass is worse than none at all. Three correctness bugs came out of the rewrite. 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. A train/serve preprocessing mismatch had enrollment using an aligned crop while the live path used an unaligned one, so embeddings landed in different regions of the space and matching quietly degraded with nothing ever erroring. And --seed silently didn't work: augmentation built an unseeded RNG per call, so two runs at the same seed produced different models — which makes every metric unverifiable, including to me. Augmentation is now seeded from (seed, epoch, index) and two runs produce bit-identical weights.

→ Liveness at 0.44 ms median (p95 0.46 ms) on ONNX Runtime CoreML, 226K params in a 1.0 MB model — about 2% of the frame budget, where detection is the real cost at 18 ms. 5,600 lines, 304 tests. Accuracy on real PAD data is deliberately not published yet: 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 + translation pipeline.

End-to-end transcription and translation that runs locally. Open-sourced and picked up by real external users.

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.