AI Research Engineer | Post-Training & Agentic Systems
DPO · RLVR · ExGRPO · LoRA/QLoRA · Autonomous Reinforcement Learning Environments
Python · Go · TypeScript · PostgreSQL
Anthony Lewallen
Bridging theoretical loss landscapes with distributed infrastructure. I build autonomous agents from the metal up—from architecting from-scratch transformer models, to engineering deterministic evaluation harnesses, to deploying self-healing RLVR training loops.
Projects
DV Eval Harness
LLM evaluation harness for hardware design verification. Drives agents through 5-step debug trajectories against broken RTL, scores reward across 5 components (root cause, evidence quality, tool use, fix plausibility, hallucination), and emits DPO-ready preference pairs. Adapter pattern supports Icarus, Cocotb, Questa, and VCS simulators.
Binary Architecture Classifier
Go inference engine classifying random binary fragments by CPU architecture via hex word n-gram TF-IDF + LinearSVC. Solved the Praetorian ML Binaries challenge 11 consecutive times — each run reaching 500 correct predictions with 0 wrong. Ported the trained scikit-learn inference pipeline to pure Go and deployed a React/TypeScript live dashboard on AWS EC2 streaming predictions in real time.
RLHF Eval
Data quality pipeline for RLHF training data. Seven detectors flagged 7.9% of 160,800 preference pairs from Anthropic's HH-RLHF dataset. Trained competing reward models on clean vs. unfiltered data to measure impact.
ScratchLM
GPT-2 (124M) architecture built from scratch in PyTorch — custom LayerNorm, GELU activation, causal multi-head self-attention, and transformer blocks with pre-norm residual connections. No high-level abstractions.
Education
University of Pennsylvania
Master of Science in Engineering -- Artificial Intelligence
University of Pennsylvania
Master of Applied Science in Computer Science -- Software Systems Concentration
American Public University System
B.S. Mathematics -- Operations Research · Summa Cum Laude