AI Resume Agent — Conversational AI
Production AI agent with 10 backend services, 5 LLM-as-a-Judge evaluators,
graceful degradation, SSE streaming, and full observability.
- Raw HTTP/2 to Gemini — no SDK, async httpx connection pooling + thread-safe TTLCache.
- 5 automated evaluators (Hallucination, Relevance, Conciseness, Helpfulness, Toxicity)
via Langfuse at 100% sampling.
FastAPIGeminiReactSSEMem0Langfuse
FunctionGemma — Hybrid AI Router
2nd Place / $5K at Cactus × Google DeepMind Hackathon. 3-tier hybrid inference:
0.99 F1 at 548ms, 70% on-device ratio.
- FunctionGemma-270M local + Gemini 2.5 Flash Lite cloud with lexical pre-router.
- Intelligent routing between local and cloud models for optimal cost/latency.
PythonGeminiLLM
RoutingHackathon
PathGuard — Spatial Safety Intelligence
Real-time hazard detection pipeline: Grounding DINO + SAM2 + Depth Anything V2 with
VLM scene understanding. UMD × Ironsite Hackathon.
- VLM generates scene-specific prompts dynamically — zero-shot object detection from
images.
- State machine + telemetry pipeline, deployable on RPi 5, iPhone, or Android.
PythonVLMComputer
VisionHackathon
English2SQL — NL → SQL Assistant
Production-style NL→SQL over Postgres. Achieved 9% accuracy lift via schema-aware
prompting.
- Evaluated multiple LLMs on accuracy, latency, and cost.
- Hardened execution with schema-aware prompting + validation.
FastAPIPostgresLLM
EvalDocker
AlphaFoundry — FF5 Factor Strategy
Rolling-window factor modeling. Backtested strategy outperformed SPY benchmark (Sharpe
0.95 vs 0.85).
- FF5 + XGBoost Learning-to-Rank with walk-forward evaluation.
- Production FastAPI inference + reproducible runs.
PythonXGBoostFastAPIBacktesting
Domain-Specific LLMs Research
NeurIPS 2023 Workshop paper. Proposed novel evaluation methodologies for faithful
LLMs.
- Co-authored research on LLM alignment and domain adaptation.
- Published at NeurIPS Muslims in ML Workshop.
NeurIPSResearchLLM