I'm a student learning ML and backend systems by building real, working things — not tutorials. The thread running through most of what's below: don't let an LLM's confident answer stand in for a verified one. Whether that's forcing a fine-tuned model to be structurally incapable of invalid output, or checking every AI-claimed fact against a real source URL before trusting it, I keep running into the same problem from different angles and building a different fix each time.
Fraud_tool_LLM — a small LLM fine-tuned and served by a custom-built inference engine, built from scratch as a free, local alternative to a hosted LLM API for a real fraud-detection system's tool-planning decision.
- Manual KV-cache management, batching, and from-scratch grammar-constrained
JSON decoding — no
model.generate(), no external constrained-decoding library - 99.3%→100% exact-match accuracy against the real production decision logic (precision-dependent), every claim checked against held-out data rather than assumed
- Caught two real regressions before they shipped: int8 quantization looked like a clean speed/size win until a direct accuracy check showed it wasn't, and a memory-constrained deployment target OOM-killed the container under fp32 until fp16 fixed it with no accuracy cost
- Containerized, deployed to AWS (authenticated, private VPC), and integrated into the target pipeline in shadow mode
| Project | What it is |
|---|---|
| Creator_track | An agentic CRM that discovers real events, researches contacts, and drafts outreach emails — with human approval required before anything sends, idempotent writes throughout, and every AI-claimed source URL checked against what a real search actually returned |
| groundtruth_mcp | An MCP server with one job: find real things matching a schema you define, and verify every result's source in code rather than trust the model's word for it |
| Git_mcp | An MCP server that helps Claude search, inspect, and compare open-source GitHub repos — live and public, no signup required |
| Flight-Price-Prediction | Regression pipeline (Random Forest / XGBoost / Gradient Boosting, tuned via RandomizedSearchCV) predicting flight prices from 10,000+ real bookings, served through a Gradio demo |
| Hexo-Frames | My photography portfolio site — wildlife, automotive, concerts, and street work, shot on a Sony α6400 |
