Projects
Selected work spanning AI/ML pipelines, agent systems, and full-stack applications.
Projects
InterviewIQ
A PEACE-framework forensic interview AI that replaces rigid script-based questioning with an adaptive LLM agent. Fine-tuned Phi-4 models via LoRA achieve 10/10 PEACE compliance at ~750ms per turn — 7.2x faster than baseline. Powered by a 4-node LangGraph pipeline with Neo4j knowledge graph for real-time context tracking and quality-driven strategy selection.
- 4-node LangGraph agentic pipeline with 7 interview strategies
- LoRA fine-tuned Phi-4 (3.8B) — 353MB adapter, production-deployed
- Neo4j knowledge graph tracking 7 CIRs and 43 EEIs in real time
10/10
PEACE Score
vs 0–1 baseline
~750ms
Avg Latency
7.2× faster
AdFlow
An intelligent automation system on Google Cloud that handles end-to-end ad trafficking. Monitors email for insertion orders, validates campaigns against an OMS, generates standardized trafficking sheets, and updates the ad server — all orchestrated by Vertex AI agents with human-in-the-loop approval.
End-to-End
Pipeline
fully automated
GCP Native
Stack
serverless
ConvoStack
A real-time voice agent platform powered by Deepgram for speech-to-text and text-to-speech. Built with a FastAPI backend handling voice stream processing and a React frontend for live conversation UI. Features low-latency bidirectional audio streaming, intelligent turn-taking, and LLM-driven conversational responses.
Full-Stack
Architecture
end-to-end
Deepgram
Voice Engine
real-time STT/TTS
Open Source
8
Repos
24
Stars
340+
Contributions
15
Followers