A Python framework that auto-tunes serverless AI/ML workflows across cost, latency, and accuracy. Profile a workflow once, then answer what-if questions across memory tiers and model variants with zero redeployment or re-profiling.

• Sweeps AWS Lambda + Step Functions configurations from 128 MB to 10 GB, benchmarked across six workflows spanning ResNet, YOLO, and BERT variants
• Published as a documented, semantically versioned package with unit and integration test suites, GitHub Actions CI, Prometheus monitoring, and Docker-based reproducible environments

• Python
• AWS Lambda
• Step Functions
• Prometheus
• Docker
• GitHub Actions

A voice agent that phones rental listing agents on a renter's behalf, built at the OpenAI-sponsored AI Tinkerers Agentic Hackathon.

• Architected the backend: a FastAPI service whose webhook tools are the only writer to one session store, with a Next.js page polling that same state so a live phone call and the screen never disagree
• Designed the transport seam every tool call routes through, so the whole product ran over a text harness while telephony was still being wired, and swapping scripted calls for real Twilio ones is a single environment variable
• Owned inbound session handling (one session id per call so every webhook landed in the same session, plus an SMS link to a live shortlist), listing ingestion, and the GitHub Actions deploy to a VM serving the app behind FastAPI

• FastAPI
• Next.js
• Twilio
• Webhooks
• GitHub Actions