Start Here
A short map of the main themes on this site.
This site is about the practical side of ML and AI systems: how models, agents, training pipelines, evaluations, and product interfaces behave when they leave the paper and enter a real engineering loop.
The fastest entry point is to pick a path below, then follow the linked topic hub when you want the full archive.
Reading Paths
LLM Agents
Tool use, agent runtime design, evaluation, context, and production patterns for systems that act across tools and environments.
- From Tool Calls to Policy Updates: A Reproducible Agent RL Stack
A reproducible blueprint for Agent RL: environment snapshots, verifier contracts, credit assignment, three-policy semantics, asynchronous generation and training, partial rollouts, and runnable smoke tests.
- A Repair Can Save a Trajectory Without Teaching the Agent
An eight-seed counterfactual-replay experiment found that high-value one-action repairs reliably rescue failed agent trajectories—but a repair-value-dominant SFT selector did not beat simpler rules and recovered only 25% of matched-RL improvement.
- Beyond 49 ms: Where VM Resume Latency Actually Goes in Agentic RL
A measurement-first look at fast sandbox resume, speculative tool execution, and action-observation co-speculation—and why time-to-useful-work matters more than boot time.
Evaluation
Practical approaches to measuring model and agent capability with deterministic checks, rubrics, trajectories, and verifiable outcomes.
- From Tool Calls to Policy Updates: A Reproducible Agent RL Stack
A reproducible blueprint for Agent RL: environment snapshots, verifier contracts, credit assignment, three-policy semantics, asynchronous generation and training, partial rollouts, and runnable smoke tests.
- A Repair Can Save a Trajectory Without Teaching the Agent
An eight-seed counterfactual-replay experiment found that high-value one-action repairs reliably rescue failed agent trajectories—but a repair-value-dominant SFT selector did not beat simpler rules and recovered only 25% of matched-RL improvement.
- Do Verifier Errors Grow Superlinearly with Horizon? A Three-Stage Experiment
A controlled long-horizon experiment found a clear horizon effect but no preregistered evidence of superlinear verifier error—and exposed why token measurement and false-negative costs matter before scaling RL evaluations.
Post-Training
SFT, RLHF, preference optimization, instruction following, reasoning traces, and data pipelines for shaping model behavior after pretraining.
- From Tool Calls to Policy Updates: A Reproducible Agent RL Stack
A reproducible blueprint for Agent RL: environment snapshots, verifier contracts, credit assignment, three-policy semantics, asynchronous generation and training, partial rollouts, and runnable smoke tests.
- A Repair Can Save a Trajectory Without Teaching the Agent
An eight-seed counterfactual-replay experiment found that high-value one-action repairs reliably rescue failed agent trajectories—but a repair-value-dominant SFT selector did not beat simpler rules and recovered only 25% of matched-RL improvement.
- Beyond 49 ms: Where VM Resume Latency Actually Goes in Agentic RL
A measurement-first look at fast sandbox resume, speculative tool execution, and action-observation co-speculation—and why time-to-useful-work matters more than boot time.
RLHF and Preference Optimization
Engineering notes and research synthesis on PPO, DPO, GRPO, reward modeling, preference data, and model behavior optimization.
- From Tool Calls to Policy Updates: A Reproducible Agent RL Stack
A reproducible blueprint for Agent RL: environment snapshots, verifier contracts, credit assignment, three-policy semantics, asynchronous generation and training, partial rollouts, and runnable smoke tests.
- A Repair Can Save a Trajectory Without Teaching the Agent
An eight-seed counterfactual-replay experiment found that high-value one-action repairs reliably rescue failed agent trajectories—but a repair-value-dominant SFT selector did not beat simpler rules and recovered only 25% of matched-RL improvement.
- Beyond 49 ms: Where VM Resume Latency Actually Goes in Agentic RL
A measurement-first look at fast sandbox resume, speculative tool execution, and action-observation co-speculation—and why time-to-useful-work matters more than boot time.
Generative UI
How AI systems can produce, steer, and execute user interfaces with structured representations and practical product constraints.
- Ad Formats in LLM Products: What's Live vs. What's Research
A survey of advertising formats in LLM products—separating what's deployed in production from what remains in research.
- Generative UI Doesn't Move the Needle—Steering Does
After shipping multiple generative UI features, I've concluded that the sophistication of AI-generated interfaces often doesn't translate to user benefit—but steering does.
- UI Representation and Action Execution for Generative UI
Exploring structured UI representation using JSON Schema, and how to implement action handlers for generative UI systems.