I build retrieval-augmented systems, and lately most of my time goes to the Jev / TypeSafe ecosystem — decision models that answer typed questions with calibrated probabilities instead of generating text. Rank #17 contributor to WeKnora, Tencent's open-source LLM knowledge platform (30k+ stars).
Jev ecosystem — three public repos, one loop: the pattern library supplies the questions, the benchmark measures them, the skill chains them into a workflow.
- awesome-jev-prompts — question patterns, anti-patterns & calibration notes for typed decisions. 51 entries / 9 categories (incl. embodied & multimodal), bilingual, every entry sourced under a four-level verification policy
- jev-rag-zh-benchmark — reproducible Chinese reranking benchmark: Jev Score nDCG@5 0.546 vs BM25 0.413 on T2Reranking hard negatives, pinned model versions (1.13 and 1.14 both measured), full logs — DOI 10.5281/zenodo.23322277, mirrors on HF & ModelScope
- jev-pipeline-skill — one workflow for question design: find → fit → draft → lint → spread → threshold → cascade; installs as an agent skill on five channels
minimind-jev (public) — a minimind-style, from-scratch tiny decision model with a full Chinese tutorial (open the box, not chase SOTA). First checkpoint already out: cnfeichen/minimind-jev-v0.1 — a 3-minute trained head, statistically tied with hosted Jev-1.13 on our rerank benchmark; the v0.2 distillation arc is a documented negative result, teacher judgments released as a dataset. Also learning embodied AI the honest way on a single 8GB laptop: robot-learning-journey — imitation learning, quadruped MPC vs RL, VLA fine-tuning, failure notes included.
Selected work
- Awesome RAG Frontier — rule-based, machine-refreshed tracker of RAG papers, projects & waves; it caught the decision-layer wave before any other list
- WeKnora — contributor, rank #17
- deer-flow · Octop — contributor
Focus areas: RAG in production, knowledge platforms, decision-layer models (Jev / TypeSafe ecosystem), agentic retrieval, question design & calibration.
中文：RAG 与知识库的工程落地，WeKnora 贡献者（第 17 位）。现在主线在 Jev 决策模型生态：问题模式库（51 条双语）→ 中文重排基准（带 Zenodo DOI）→ 问题设计流水线技能 → minimind-jev（从零训练小决策模型 + 中文教程，首个检查点已发布），另有 8GB 笔记本上的具身智能学习笔记（robot-learning-journey）。