I'm Bishal Upadhyaya. I started in neuroscience — recording electrical signals in living neural circuits — and now I build AI systems that ship: research on one end, products people actually use on the other.
What I do
I work end-to-end: data pipelines, models, infrastructure, and the product on top. My bar is that the system has to feel effortless to the person using it — not just impressive in a benchmark table.
- Agents and LLM systems that survive contact with production: tool-calling with structured outputs, multi-model routing, eval harnesses in the loop, and honest failure reporting — the demos on my homepage run on these systems, live.
- Retrieval and model efficiency: agentic RAG, hybrid + semantic search over vector stores, and models distilled to a fraction of their size without giving up recall — my production fingerprinting model is 27.7M parameters doing work most teams throw 300M+ at.
- Infrastructure a CTO can sign off on: Terraform-managed AWS and GCP, least-privilege credentials, SOC2-minded logging, and cost governance with receipts.
Current focus
I co-own and build Vibeset — AI music tooling across three products: Curation (setlist generation), Cue (music synced to picture, live at cue.vibeset.ai), and Choon (audio fingerprinting and provenance).
Music is just where these systems ship today. The same research-to-production spine has gone into early Alzheimer's detection at UCSF and model-efficiency research at Thermo Fisher, alongside NVIDIA and HPE — the domain changes, the rigor doesn't.