Software Engineer ยท Backend / Infrastructure & Applied ML
I build the systems that sit next to models and have to stay reliable: containerized CI pipelines, PostGIS-backed services, and LLM/retrieval stacks.
A queryable spatial-graph reasoner over captured 3D scenes. It builds a typed graph of spatial relations and prunes it before retrieval, so only the relevant subgraph reaches the LLM. The router can answer, return a grounded empty, abstain, or say unknown, so it never fabricates certainty the graph can't support. Inputs are hash-pinned and pipeline stages isolated, so any failure attributes to a single stage.
Most people meet a bird by ear long before they ever see one. This map works backwards from that: sweep a region to hear what lives there, or describe the shape of a sound you half-remember and let the field narrow until a name surfaces. Hundreds of species across a full seasonal cycle, assembled by an ingest pipeline that pulls open occurrence and media data, then clips, normalizes, and classifies every recording into static JSON. No API keys, no server, no tile provider.
A daily digest of the LLM ecosystem: scrapes arXiv (cs.CL / cs.LG / cs.AI), clusters and filters for relevance with an LLM summarizer backend, and surfaces what's actually moving in research and tooling so the signal doesn't get lost in the volume.
I'm always excited to collaborate on projects that make a difference. Whether you have an interesting problem to solve or just want to chat about technology and innovation, I'd love to hear from you.