Recent work:
- OpenCodeIntel (https://opencodeintel.com): code search for AI coding agents. Web app, REST API and a 12-tool MCP server; 94% Hit@1 across 14 open-source codebases on a 665-query research eval. Solo, MIT.
- Overhear (https://github.com/DevanshuNEU/overhear): grades every call a voice agent takes against the clinic's database. Code decides the facts, an LLM judge scores the rest; caught 23 of 23 planted failures on a 39-call golden set.
- CallBudget (https://github.com/DevanshuNEU/callbudget): a voice agent that finds scarce drugs in fewer calls (4.3 to 2.3 expected) and abstains instead of guessing (false "in stock" 10% to 0%, simulated pharmacist on real data).
- Saar (https://getsaar.com): Chrome extension on the Web Store that meters Claude.ai tokens and cost, entirely client-side.