genx
multitenant enterprise AI platform: 12 microservices, agent runtime, hybrid RAG and MCP connectors.
github profileread postmultitenant enterprise AI platform: 12 microservices, agent runtime, hybrid RAG and MCP connectors.
github profileread posthand detection as input: turn any camera into a touchless gesture controller.
PalmRecognitionPublicread postlegal document intelligence: six-stage OCR, parsing and knowledge-graph pipeline.
github profileread postopen source contributions to Meta's synthetic dataset generator for LLM fine-tuning.
meta-llama/synthetic-data-kitread postmost of the work is platform: a 12-service agent runtime, a six-stage document pipeline from scan to knowledge graph, and computer-vision work that turns a webcam into a touchless controller.
some of it ships upstream. i contribute to meta's synthetic-data-kit, and use it in my own PEFT and LoRA experiments.
How do we make agent runtimes reliable across OpenAI, Anthropic, Gemini and local TGI models?
How do we keep retrieval quality high with dense + BM25 fusion, reranking and per-tenant isolation?
How do we generate and evaluate high-quality synthetic data for fine-tuning?
I like building systems that other engineers can extend: clear service boundaries, typed interfaces, tests around the risky parts, and documentation that treats the next engineer as the user.