Skip to content

Portfolio

Portfolio

Things I've designed and built, end to end.

AI Agent · TestingCase studyCode on request

Prooftrace

An AI tester that proves every promise a feature makes.

  • 8 of 8 criteria proven
  • 29.7 s run
  • $0.028 in AI cost

TypeScript · Node.js · Playwright · SQLite · LLM agents

CAFE
Edge AI · ACM/IEEE SEC 2026PaperOpen source

CAFE: facial expression recognition at the edge

A two-stage YOLOv12n pipeline made structurally prunable, then compiled to run 3.33× faster and 2.12× smaller on a 15 W Jetson Orin Nano.

  • 3.33× faster
  • 2.12× smaller
  • 88 configurations

PyTorch · TensorRT · YOLOv12 · Jetson Orin Nano

AI Security · IEEE SoutheastCon 2025PaperOpen source

ReAL: reflective attack detection for LiDAR

Real-time detection of reflective attacks on LiDAR, running on a Jetson Orin: up to 99.97% accuracy at 0.083 ms per inference.

  • 99.97% accuracy
  • 0.083 ms latency

Python · Machine learning · LiDAR · Jetson Orin

AI Security · Physical testbedCode on request

Adversarial patch removal for stop signs

A 3.35M-parameter dual-head U-Net that erases adversarial stickers before detection, lifting stop-sign detection from 48% to 76% on a Quanser QCar2, with no loss on clean signs.

  • 48% → 76% detection

PyTorch · U-Net · ONNX · TensorRT · QCar2

GenAI · Full stackOpen source

AI Voice Compliance Auditor

Upload or record a call and get a transcript, speaker turns, sentiment, toxicity, intents and a 0 to 100 compliance score, with custom rules, alerts and reports.

Whisper · FastAPI · React · TypeScript · Transformers

AI Security · IEEE SoutheastCon 2025PaperOpen source

GNAP: attacking and defending face detection on edge devices

A paired attack and defense for edge face detectors: GNAP drops mean detection confidence from 0.99 to 0.82, and the GNAG defense restores it to 0.98.

  • 0.99 → 0.82 → 0.98

Python · YOLO · Adversarial ML · Edge AI

Vision-language · BenchmarkCode on request

VLM object detection baseline on BDD100K

An honest, reproducible zero-shot benchmark of Qwen2.5-VL on BDD100K driving scenes, served with vLLM and scored with COCO mAP.

Python · Qwen2.5-VL · vLLM · Docker · FastAPI

Geospatial ML · Full stackOpen source

Nashville crime hotspot analysis

DBSCAN hotspots from Metro Nashville 911 data, weighted risk scores, safe-route planning and proximity alerts on an interactive map.

Python · FastAPI · React · TypeScript · DBSCAN

Anneal
Data PlatformReturning soon

Anneal

Enterprise visual data pipeline platform with governance, lineage and data quality built in.

Data · Governance · Pipelines

NLP · ToolOpen source

Academic Text Summarizer

A lightweight DistilBART summarizer for papers and technical text, with a Gradio UI that also runs offline.

Python · Hugging Face · PyTorch · Gradio

Recommender · AppOpen source

Movie Recommender System

Pick a film and get five similar ones with posters, from a content-based model on the TMDB 5000 dataset.

Python · Streamlit · TMDB API