MCA ’26 · Anna University. AI/ML Engineer and Full-Stack Developer based in Tamil Nadu. Building and deploying AI-powered applications.
DistilBERT fine-tuned on 17,880 EMSCAD listings, ONNX INT8-quantized to 268MB — 87.57% F1 on a 4.8:1 class-imbalanced dataset. A model router hot-swaps between BertPredictor and LinearSVC at runtime without restart. Paired with a 10-signal URL fraud scorer and token-level prediction explainability. |
Point-by-point WebSocket sync with sub-100ms latency via FastAPI. Board state replays fully on reconnect — late joiners see the complete canvas. Supports 9 brush types, 64-step undo/redo synced across all clients, and auto-reconnect with exponential backoff (500ms → 8s). |
Concurrently scrapes multiple e-commerce sources with tenacity-backed retries, deduplicates results, then passes top candidates to Gemini 2.0 Flash for ranked AI recommendations. Disk-cached with a 30-minute TTL — keeps p95 response times acceptable on Render's free tier. |
Building things, learning along the way. 🚀
