Projects
Production systems from my enterprise and client work, followed by products and open-source work I run myself. Client names are omitted.
Enterprise RAG assistants, LLM chatbots, and Whisper large-v3 transcription running on self-hosted LLaMA and Qwen models on GPU Kubernetes. Retrieval runs over FAISS and Chroma, and transcripts feed NLP extraction and analytics, all without data leaving the client's infrastructure.
Impact: data-sovereign generative AI for privacy-critical enterprise environments.
A multi-stage pipeline that transcribes meetings with Whisper, summarizes them with LLMs, and uses LangChain agents to extract action items and draft follow-ups for enterprise teams.
Impact: post-meeting documentation cut from hours to minutes.
Dense retrieval with sentence-transformers and FAISS, followed by re-ranking, to match candidate resumes to job requirements across recruitment workflows handling thousands of applications.
Impact: deployed in government hiring workflows.
Analysis of recorded interviews combining facial recognition, sentiment analysis, and document processing, with AI-driven fraud detection to flag inconsistent or impersonated candidates.
Impact: added automated integrity checks to enterprise HR workflows.
Visual search on CLIP-style transformer embeddings with FAISS approximate-nearest-neighbor retrieval, CNN-based auto-tagging, and recommendation pipelines mixing collaborative filtering with embedding retrieval.
Impact: replaced manual catalog tagging and improved product discoverability.
OpenFPL Scout AI is a live Fantasy Premier League scout built on an ensemble of Ridge, XGBoost, CatBoost, and MLP models over time-series features, trained on historical form and refreshed with official FPL data, and validated on a full held-out season. It forecasts player points, optimizes a full 15-player squad, and ranks captaincy picks before every gameweek deadline.
Impact: a free public app, rebuilt every gameweek from live data.
Designed multi-server resource loading for get_resources in LangChain's MCP adapters: fetch resources from every connected server concurrently when no server is named, with clear validation errors. The LangChain team merged the feature upstream in #380, building on my PR #360. I've also published 15+ reproducible Kaggle notebooks on PyTorch, NLP, and ML fundamentals.
Impact: open-source tooling and learning resources for the ML community.