Backend and machine learning engineer.
I build backend services and the machine learning systems behind them, from the API layer down to the infrastructure they run on.
Next.js 16 React 19 TypeScript Perfect Corp API Vitest
Selfie in, scores across 11 skin concerns and a matched AM/PM routine out. The planner is deterministic code and the LLM only rewrites wording, so the output schema cannot break and 120 Vitest tests cover the planning logic directly.
Next.js 16 FastAPI Python PyMuPDF
Paste a job posting, get the keywords your resume is missing ranked from critical down to nice to have. It runs on your own key across six providers, four of them free, and the key is passed per request and never stored, so hosting it costs nothing.
FastAPI PyTorch AWS S3 Docker Terraform Prometheus Grafana
EfficientNet-B0 served behind FastAPI, with model weights pulled from S3 at runtime rather than baked into the image, so the container stays small and the model can be swapped without a rebuild. Terraform provisions the AWS side and the compose stack brings up Prometheus and Grafana alongside the API.
Node.js TypeScript Apollo Server v5 GraphQL Redis Kafka PostgreSQL Kubernetes CI
WebSockets for delivery, GraphQL for history. Redis Pub/Sub fans messages across instances so any server can reach a client connected to any other, and Kafka carries async events to decoupled consumers. Auth is JWT with bcryptjs, DataLoader batches the resolvers, and the repo ships GitHub Actions CI alongside Kubernetes manifests for the app, Postgres, and Redis.
TensorFlow EfficientNet-B0 MTCNN dlib Grad-CAM
EfficientNet-B0 with spatial attention reached 89.33% at 0.9611 AUC on FaceForensics++, and the best ensemble reached 95.33%. Tested cross-dataset on Celeb-DF v2 it falls to 75.00%, which says more about how well it generalizes than the headline figure does.
Python Ridge Regression scikit-learn
Which failure mode actually breaks a brain-computer interface: electrodes dying, or the neural signal drifting over months? On 256-electrode intracortical recordings from an ALS patient (Card et al., NEJM 2024), drift is by far the more damaging of the two.
Python Google ADK Gemini DataHub
Ask a data catalog a question in plain English, get an answer traced through real lineage and ownership with every asset named by its URN. The agent holds one write tool out of the twelve DataHub exposes, because catalog text is untrusted input reaching a model that has write access.
JavaScript Node.js CI · 8 stars · zero dependencies
Pulls fresh job postings into a spreadsheet from the same public listings a logged-out visitor sees. Most tools in this space want your session cookie, which is exactly the automation LinkedIn restricts accounts for.
Other projects
| Project | Stack | What it does |
|---|---|---|
| weather-forecast-project | LSTM · ARIMA · Streamlit | Ensemble weather forecasting dashboard, best RMSE 2.61°C |
| slippage-impact-model | Python · NumPy · SciPy | Nonlinear market impact from live order book data, with Lagrange-optimized trade scheduling |
| ofi-feature-construction | Python · pandas | Order flow imbalance features from limit order book data, including cross-asset OFI |
| music-recommender | Python · Flask | Collaborative filtering over a user similarity graph, walked to surface unheard tracks |
Languages: Python, TypeScript, JavaScript, SQL
Backend: FastAPI, Node.js, GraphQL, REST, WebSockets, JWT, Next.js
Machine learning: PyTorch, TensorFlow, scikit-learn, OpenCV, EfficientNet, LSTM
Infrastructure: AWS, Docker, Kubernetes, Terraform, Redis, Kafka, PostgreSQL, Prometheus, Grafana