Notes on AI, software, and research
Long-form writeups, research notes, and thoughts on the systems I build and study.
- Warren: the plan
What I'm building, why now, and the design lessons it's reverse-engineered from. Notes from the plan rather than a post-mortem. In active development.
- Scheduled Partial-Credit RL for Reliable Code Generation with Small Language Models (WIP)
A reliability-first RL framework for SLM code generation. Partial-credit functional reward, binary-to-partial curriculum, on DeepSeek-Coder-1.3B and APPS+. LCTES 2026.
- Match Your Loss to Your Cost: Asymmetric Losses and Conformal Capacity Bands for Backbone Traffic Forecasting
Backbone operators care about SLA violations and over-provisioning, not RMSE. A drop-in cost-aware loss and a conformal wrapper that cut realized cost up to +76% on Abilene at equal forecast accuracy. Submitted to CNSM 2026 (June 1, 2026).
- Fixing performance bugs through LLM explanations
Training an LLM to explain a performance bug, not just classify it, produces a stronger detection signal. 490-bug dataset, fine-tuned GPT-4o-mini, IEEE AITest 2025.
- A citation-grounded RAG for technical documents at Morgan Advanced Materials
Capstone build for the Penn State Learning Factory, sponsored by Morgan Advanced Materials. Offline RAG that refuses to answer without a citation.
- TruthCast: multi-agent fact-checking with on-chain provenance
A fact-checking pipeline that decomposes claims, weights evidence by source credibility, debates ambiguous cases, and writes verdicts to a Solana ledger. HackPSU Spring 2026, Solana track winner.
- Shift: personalized sustainability with radical transparency about AI's energy cost
12-hour hackathon build that delivers one tailored climate action per day while disclosing the carbon cost of every inference. GDG Solution Challenge winner.
- Hello World
Welcome to my new blog.