Systems I've built end to end.

Multi-agent risk-intelligence platform: hybrid BM25 + Milvus retrieval feeds a RAG pipeline over live risk data, with RAGAS and Langfuse wired in for real-time hallucination monitoring and a finetuned sLLM generating compliance reports.

Multi-agent fake-news detection engine — queries 3 local LLMs (Qwen, Phi-3, LLaMA) in parallel and aggregates verdicts via a consensus layer to suppress single-model hallucination bias, evaluated with RAGAS across political and scientific claim sets.

Four-stage request pipeline — prompt reformulation → security filtering → task triage → token optimization — routing to a local Ollama model first, cloud LLM only on cost-threshold fallback. Shipped as a backend service and Chrome extension.

Hybrid BM25 + FAISS retrieval with cross-encoder reranking and hallucination-detection fallbacks. Answers compliance questions grounded in RBI regulatory filings, with per-query Langfuse tracing on latency and reranker scores.

Retrieval-augmented question answering grounded in Airbus Flight Crew Operating Manual documentation — a domain-grounding test case for RAG in a setting where wrong answers aren't an option.

A local-first coding assistant built on Ollama — keeps inference and code context on-device rather than routing every completion through a cloud API.

Image enhancement engine for automated quality upscaling and denoising — integrated into the FamioX content pipeline alongside the vision-tagging service.

CNN-based image classifier distinguishing healthy potato leaves from early and late blight — an agri-tech application of the same computer-vision fundamentals used in the FamioX vision pipeline.

A full-stack platform for reporting and tracking local civic issues — built to connect citizens with municipal follow-up rather than let reports disappear into a form.

A network scanning and reconnaissance tool for surfacing open ports and service fingerprints — infrastructure-adjacent tooling outside the ML stack.