Hey there! I'm Kenny. 👋

I build end-to-end AI agent systems — multimodal LLMs that reason, call tools, and remember — and I ship them to production with tests, telemetry, and cloud deployment.

I'm currently pursuing my M.S. in Electrical & Computer Engineering at New York University, based in New York City.

I'm actively seeking ML/AI engineering opportunities with teams building real agent products. Feel free to reach out to me!

Selected Work

VideoSense

Flagship · AI Agent

Independent project — end-to-end conversational agent for video understanding

Ask your video library anything in plain English. A self-designed probe-and-step agent loop on Gemini 2.5 watches the raw footage, reasons across a multi-turn conversation, and answers with the playable clips and charts that prove it — each reply carrying its own cost receipt. Runs in production on Google Cloud Run, with an MCP tool layer, cross-session memory, and a τ²-bench-style suite of 128 graded evaluation tasks.

Visit videosense.work ↗ Watch the live demo ↗ GitHub ↗

A finished VideoSense answer — 3 of 12 clips are pure freefall, shown as labelled clip cards, a freefall-vs-canopy mini chart, and a footer with steps, time and cost

SocialLens

Independent project — consumer-insights dashboard

Short-form video in, decisions out. An interactive insights dashboard — sentiment drivers, momentum, personas, clip-level receipts — plus the real Gemini pipeline that generates it from raw footage.

Live demo ↗ GitHub ↗

SocialLens dashboard — topic KPIs at a glance

DINOv2-GraphNAV

Personal project — visual navigation

Explore once, navigate forever. Builds a topological map from what a robot sees — DINOv2 features over a k-NN graph — then localizes in real time with FAISS and routes with A*/Dijkstra at sub-100 ms per query.

GitHub ↗

DINOv2-GraphNAV — navigate from a feature graph

face_detect

Personal project — computer vision

Real-time face recognition in video streams: YOLOv8 detection, MTCNN alignment, and FaceNet embeddings trained with triplet + cross-entropy loss.

GitHub ↗

face_detect — know every face in the frame