I build LLM-powered agentic systems — and I'm increasingly pulled toward the layer beneath them: inference, model serving, and the systems that make agents actually run.
Engineer with a habit of shipping end-to-end.
I work mostly on agentic AI — multi-agent pipelines, RAG, LLM orchestration — but I like owning the whole stack, from the model to the FastAPI backend to the deploy. Two-time hackathon track winner, six-plus builds shipped under pressure.
A polyglot microservices security platform — a Python FastAPI AI engine plus three Spring Boot services over an event-driven Kafka pipeline (scan-jobs / scan-results). Sub-second HTTP responses for async scans, an AST-based parser feeding a Claude review agent with enforced JSON tool-calling, fully Dockerised and CI/CD'd on AWS.
Fine-tuned Qwen2.5-7B on Stanford's LegalBench with LoRA + DPO via TRL for contract-clause classification, statutory reasoning, and legal QA. Built an eval pipeline benchmarking it against GPT-4o mini and Claude Haiku on accuracy, cost, and latency — trained on 2× T4s with 4-bit quantization, served via FastAPI on AWS.
An agentic marketing workspace — multi-step AI agents that handle campaign planning, content generation, and social scheduling end-to-end, with human-in-the-loop approval flows. Hackathon-winning build.