• Built an automated eval-driven release gating system for LLM serving changes, enforcing promotion/rollback decisions via side-by-side evaluation of quality, latency, and behavioral drift metrics under production load.

• Designed and built a real-time semantic observability service for LLM drift detection, comparing online logprob distributions against calibration baselines using non-parametric statistical methods to flag silent hardware faults and behavioral changes.

• Built a low-latency LLM inference platform using SGLang on Kubernetes, leveraging sticky routing to optimize KV-cache hit rate and prefix overlap across a multi-cloud GPU federation.

• Designed a forecast-driven proactive scaling control plane to align H200 GPU capacity with projected demand, virtually eliminating cold starts to meet conversational latency SLAs (400ms p90) — enabling the company to scale from 100k to 2M phone calls per month.

• Parallelized PyTorch inference runtime by decoupling tokenization from model execution, improving GPU utilization and increasing throughput by 25% ($100k/month savings).

• Scaled BERT-based NER inference from thousands to billions of clinical notes by redesigning pipelines for distributed execution using Spark.

• Replaced ad-hoc manual inference runs with production-grade Airflow pipelines, enabling reliable reprocessing of billions of clinical notes with full auditability.

• Optimized database snapshot infrastructure using delta-based updates and columnar indexing, enabling 1000x faster reads and supporting hundreds of concurrent queries.

Trained and deployed a surgery duration prediction model into Epic EHR, improving scheduling accuracy by 17% MAE over the baseline and saving hundreds of nurse hours per week — system remains in production.

Teaching Assistant for Artificial Intelligence class under Dr. Ansaf Salleb-Aouissi (co-taught by Dr. Tony Dear)

• Taught ML concepts (SVMs, random forests) and search algorithms (heuristic, adversarial, backtracking) in Python during weekly office hours.

• Graded student work, both coding and conceptual.