โง ๐๐๐๐พโด๐๐พ๐๐ ๐๐๐ปโด๐๐น๐ ๐ถ๐ โฏ๐๐น๐โฏ๐๐ ๐๐๐๐ถ๐ โง
Building verifiable, reliable AI in the open โ from constrained generative models for science (ICDM '26) to LLM post-training, observable agents, and practical AI apps.- PDexFlow โ Official code for Escaping the Mirror Trap: Parity-Harmonized Flow Matching on Riemannian Manifolds for Peptide Design (ICDM 2026).
- LoopPrism โ A local-first session inspector tracking tool calls, context usage, timing, and estimated costs across Claude Code, Codex, OpenCode, DeepSeek Harness, and WorkBuddy.
- dsh-effort-pilot โ Adaptive per-step thinking-effort plugin for DeepSeek Harness, alongside dsh-quota-meter, svg-animation-pipeline, and handwriting-pdf plugins/skills.
- ecom-posttrain โ LLM post-training experiments with SFT, DPO, GRPO, and LoRA alongside local quantized GPU serving (paired with nano-posttrain).
- microduck-qwen โ Simulation research combining Qwen vision grounding, natural-language navigation, and post-training baselines for a bipedal robot agent.
- OfferPilot โ A TypeScript AI job-search app for graduates combining hard-constraint filtering, weighted ranking, and LLM-assisted resume workflows.



