A four-service streaming pipeline — Deepgram STT to Groq inference to Cartesia TTS to a lip-synced Simli avatar — running at ~387ms end-to-end over WebSocket and WebRTC. The engineering was backpressure, stream teardown and keepalive handling, not prompting. Latency is measured by an instrumented endpoint, not estimated.
Makes BLAS, LAPACK and ScaLAPACK codebases queryable in natural language. The hard part was chunking: naive line windows split a SUBROUTINE in half and return an algorithm with no signature, so chunks are cut on routine boundaries instead. Python, 15-test suite covering the chunker and retrieval.

AI-powered talent retention platform that uses predictive analytics to identify at-risk employees and provide actionable retention recommendations. Built to demonstrate ethical AI that enhances human employment rather than replacing it.