Building a large-scale benchmark dataset for hardware e-waste detection and recovery in data centers. Scraped more than 1.5 million images with agent-driven web scraping pipelines, and am using it to train efficient vision models for identifying recoverable hardware.

Black-box search framework that clones a target voice onto a frozen TTS model without retraining it, built for a social robot. Raised speaker-similarity fitness from 84.95 to 96.61 in 2,000 evaluations, and now runs in production on the robot.

Multiparty perception for a social robot. Fuses microphone-array direction of arrival into camera video to resolve who is speaking and who is being addressed in group conversation, with a geometry-grounded benchmark for evaluation.

Built a multi-agent job search tool where an orchestrator dispatches specialist sub-agents for search, scoring, and drafting, each scoped to one task through its own agent skill. Wrote a custom MCP server normalizing eleven job boards behind a single tool surface. Kept dealbreaker filtering in deterministic Python so an untrusted posting cannot prompt-inject past it, and enforced a human-in-the-loop gate with no code path to auto-submit.

Built a face authentication system with liveness detection that runs on a standard RGB camera instead of depth or infrared hardware. Uses blink detection to reject photo and screen spoofing attempts, and stays accurate when the user is wearing a mask.

Built a vision tool that identifies electronic waste from a photo and returns its carbon footprint, recoverable value, and disposal action. Implemented in Streamlit using Gemini 3 Flash for vision with strict JSON output, plus a heuristic decision engine with safety overrides.

Built an NLP system to classify sentiment in high-volume financial news. Benchmarked classical and transformer models across two datasets, with FinBERT at 0.73 accuracy and 0.74 macro-F1 on Alpha Vantage, and BERT-uncased at 0.86 and 0.86 on Yahoo Finance, validated against real ticker movements.

Built a Streamlit tool for interactive backtesting of forecasting strategies with reliable horizons up to two months, paired with a chat interface for querying 10-K filings in plain language.