Résumé

• Author adversarial evaluation sets and reinforcement learning environments for frontier LLMs, writing domain-specific scenarios, questions and rubric-based reward functions; over 50% stumping rate on multi-step quantitative reasoning tasks.
• Build and curate labeled evaluation datasets, refining prompts against model responses and reasoning traces to separate genuine reasoning gaps from prompt ambiguity.
• Score and analyze outputs against calibrated rubrics that serve as RL reward signals, surfacing systematic weakness patterns for the next round of environment design.

• Analyzed 1M+ underwriting and litigation records using SQL and Python to surface data inconsistencies and workflow bottlenecks, improving decision accuracy by 10%.
• Built optimized data management pipelines and reporting across multiple systems, improving data accessibility and operational efficiency.
• Validated an AI-assisted analytics tool with business stakeholders, improving processing efficiency by 60%.

• Processed and analyzed 700+ structured and unstructured forensic artifacts per run.
• Designed data ingestion, management and validation pipelines for downstream analysis.
• Documented schemas and processing logic while testing outputs, improving reliability and investigative efficiency by 30%.

• Developed SQL and PL/SQL reporting workflows across 20+ enterprise modules.
• Analyzed 50K+ monthly transactional records, improving system performance by 15%.
• Partnered with cross-functional teams to gather requirements and validate reporting outputs.

Toki · On-device AI meeting assistant

• Fully offline meeting intelligence chaining ASR, speaker diarization and a 4-bit quantized Qwen3-4B for note generation; zero cloud inference cost, no data egress.
• GPU inference pipeline for a 0.6B ONNX ASR model and 4B LLM on a 4GB-VRAM consumer GPU; live transcription latency from 21s to ~1s, diarization at 0.014 RTF.
• Sole engineer from model selection to a packaged Windows release; 9.2% WER, retrieval precision@1 from 82.9% to 87.5%.

GlassBox · Mechanistic interpretability

• Mech-interp system for medical LLMs on Gemma 3 4B/2B, surfacing 15+ SAE feature activations with real-time uncertainty and harmfulness tracking.
• FastAPI and RunPod GPU microservices with Claude agents, Arize Phoenix and Sentry for live interpretability monitoring.

ExoSeeker · AI prediction platform

• Exoplanet detection on Kepler data, over 90% accuracy, with PyTorch and scikit-learn pipelines.
• Best Use of NASA Data, NASA Space Apps Challenge (2,000+ participants); represented Cincinnati at the Global Challenge.

NeedleHelp · MedTech

• End-to-end IoT + ML system for real-time robotic control, over 90% system accuracy.
• 1st Place, Best Project Overall, Ohio's largest hackathon (140+ teams, 800+ participants).