June 20, 2026 at 5:29 PM

cool things i've worked on

a quick index of the things i'm proud of. the platforms below are written out in full, click a title to jump straight to that section.

  • ๐Ÿ“ž core voice ai platform: human-in-real-time voice agents. cut turn latency 60% (p90), ~100k calls / day.
  • ๐Ÿค ai sales copilot: a full-stack ai sales OS for account executives.
  • ๐Ÿ› ๏ธ nustack: reliable enterprise software, ai-native.
  • ๐Ÿงช anneal (side project): a web qa engine. an agent writes the tests, plain code replays them.
  • ๐Ÿ† hackathons: 3ร— nurix hackathon winner.

๐Ÿ“ž core voice ai platform

the voice platform is a core part of nurix's offering

what cool things i worked on:

  • improved voice-conversation quality and user-interruption handling so the agent stops talking when you do, like a person would. asks you to repeast if it could not hear you properly and ask you to hold if an API/Tool call is taking too long. i wrote about how we built the "can you repeat that" bit.
  • cut end-to-end turn latency by 60% at p90 and 20% at p50.
  • owned and built the campaign manager that orchestrates outbound calling and manages concurrency. scale ~100k calls / day.

stack: livekit-agents ยท python ยท campaign manager in java spring boot.

๐Ÿค ai sales copilot

a full-stack ai-powered sales OS platform for anyteam.com, built for account executives.

the problems worth solving turned out to be:

  • meeting prep, automated research and context aggregation so an AE walks in already knowing the room
  • structured summaries from past meetings
  • concrete next-step recommendations from past meetings
  • all this is generated automatically.

what i owned:

  • two backend services (node.js)
  • the data ingestion + consolidation pipeline (fastapi) that powers the whole thing
  • led a pod of 3 junior engineers through it

๐Ÿ› ๏ธ nustack

a bet that you can generate enterprise-grade software with ai without sacrificing reliability. instead of letting an llm write greenfield code and hoping, nustack constrained it to compose vetted primitives and synthesize only the genuinely novel ~10%, so reliability became a property of the system, not of any single generation. i owned the autonomous qa layer, which generated and ran the test, performance, and security batteries that gated generated code before it shipped.