AI systems that hold up in production.

• When a promising demo has to become something people rely onTurn LLM prototypes into systems teams can trust: structured outputs, cost control, production hardening.
• When nobody can say whether it’s actually workingBuild the evaluation harness: judges checked against hand labels, controlled A/B runs, so “better” is a number.
• When the inputs are messy: documents, specs, voice, scattered dataTurn unstructured inputs into dependable, checkable decisions.
• When you need someone embedded with the problem, from model to interfaceWork alongside the people who use it, scope what matters, then build and own the result.

1. Measure the measurement

   Run the baseline more than once before trusting any before-and-after comparison. See it in practice

2. Test the obvious fix

   The intuitive improvement is sometimes the one that hurts. Check it with a controlled comparison first. See it in practice

3. Find the waste

   Quantify what a pipeline repeats or discards before adding more model calls. See it in practice

• Team of 4

  Teaching a small model to say “I don’t know”

  Invented answers cut from 38% to 0.03% by fine-tuning a sub-1B model.

• Team of 5

  Plan, don’t generate

  An AI that directs a narrated explainer video but never writes the code that renders it.

• Small team

  Vāṇī: a voice AI engine where every part swaps

  Real-time conversation with natural interruption, in the cloud or fully offline.

All projects