SSW.

Projects.

Research & applied engineering case studies

ConGrs: Graph-Based Response Synthesis for Factual LLM Generation
FeaturedPublic
click to expand

ConGrs: Graph-Based Response Synthesis for Factual LLM Generation

Consensus Graphs (ConGrs): a DAG-based data structure that synthesizes epistemic signals across multiple sampled LM responses into a single, more reliable response. Accepted to COLM 2026.

LLM ReasoningTest-Time ScalingFactuality
Reliable Micro-Benchmarking for LLM Evaluation
FeaturedPublic
click to expand

Reliable Micro-Benchmarking for LLM Evaluation

A meta-evaluation measure for LLM micro-benchmarking that reveals when small evaluation subsets fail to reliably preserve model rankings from full benchmarks. ICLR 2026 Oral.

Efficient EvaluationLLM EvaluationMeta-Evaluation
Assembly Copilot: Computer-Vision Assembly Verification for Fortune 500 Manufacturing
FeaturedPrivate
click to expand

Assembly Copilot: Computer-Vision Assembly Verification for Fortune 500 Manufacturing

Real-time, camera-based assembly verification deployed to Fortune 500 manufacturing lines, achieving 98% detection accuracy while cutting manual annotation overhead by 80% through synthetic data generation.

Object DetectionActivity RecognitionSynthetic DataComputer VisionPython
Ergo Copilot: Computer-Vision Ergonomics Risk Assessment
Private
click to expand

Ergo Copilot: Computer-Vision Ergonomics Risk Assessment

A computer-vision ergonomics risk assessment module for continuous, objective monitoring of workplace safety and regulatory compliance on manufacturing floors.

Computer VisionErgonomics AnalysisPose EstimationPython
Edge-Optimized Recommender & Model Compression Pipeline
Private
click to expand

Edge-Optimized Recommender & Model Compression Pipeline

Two tightly-related production workstreams: knowledge distillation and TensorRT optimization for low-latency edge inference, and a scalable cloud recommender system driving conversion.

Knowledge DistillationTensorRTRecommender SystemsEdge AICloud
GenAI Title-Performance Explainability & Forecasting
Private
click to expand

GenAI Title-Performance Explainability & Forecasting

A production-grade GenAI explainability system for entertainment content performance, combining multi-platform social signal forecasting with ensemble LLM-as-a-judge attribution for non-technical stakeholders.

LLM-as-JudgeForecastingPythonData VisualizationGenAI

Portfolio in progress. I'm actively open-sourcing projects to GitHub and adding new case studies here over time.