g/ A BUILDER'S TIMELINE @cryptoSUN2049

Experience & direction

  1. EARLIER01 / SYSTEMS

    Cloud systems engineering

    Tencent Cloud · OS team

    Worked on cloud operating systems and infrastructure. The systems layer came before the model layer.

    • Cloud infrastructure
    • Systems engineering
  2. FROM 202302 / WEIGHTS

    LLM training & post-training

    xDAN AI · Founder & model training engineer

    Led our team's work on the xDAN model family: training, alignment, and the engineering around model releases.

    • SFT / DPO
    • MoE
    • Reward modeling

    Selected team releases ↴

    2023xDAN-L1-Chat-RL-v17B chat model · SFT + DPO 2024xDAN-L3-MoE141B-parameter MoE release 2025xDAN-L2 · Qwen2.5-32B RM32B release in the xDAN model collection 2026APUS-OpenJev-v1Decision models · 4B / 9B / 35B-A3B · co-author
  3. NOW CURRENT FOCUS03 / AGENTS

    Building Agent Native

    xDAN AI · Founder

    Building products and systems around agents: tools, execution, feedback, and real-world tasks. Turning model capabilities into useful work.

    • Agent infrastructure
    • Tools & execution
    • Evaluation

    2026 / JEV REPRODUCTION SERIES

    APUS-OpenJev-v1

    Co-author · APUS AI-LAB · 4B / 9B / 35B-A3B

    An open-source model series reproducing the Jev decision-model approach, with extensions for dynamic-depth computation. All three models outperform the recorded Jev API baseline on our shared 1,000-question development panel: 4B 80.5%, 9B 81.1%, and 35B-A3B 82.2%, versus 77.0%. Built for browser agents and business workflows.

    35B-A3B · 1,000-question panel
    82.2%
    Above the recorded Jev API result
    +5.2 pp
    9B · local HTTP median latency
    ~26 ms

    Dynamic-depth computation. Jointly trained early and full-depth paths let applications select a smaller or larger compute budget through effort="low" or "high". Candidate scoring keeps decisions within the supplied action space.

    Same 1,000-question development panel: APUS-OpenJev 35B-A3B 82.2%, 9B 81.1%, 4B 80.5%; Jev API 77.0%; Laya typed full input 50.5%. APUS uses full-depth checkpoint-5949 base plus adapter.
    Valid and correct decisions out of all 1,000 questions. Full-depth base + adapter; an expanded development panel, not a sealed final test. View full-size chart

    Latency comes from a separate 9B serving run on one RTX PRO 6000: ~26 ms median local HTTP response time. It is not a matched-hardware speed comparison with the Jev API.

    Other current work includes OpenFinClaw, a financial agent project.

    xDAN-DSH-uni-agentWork on a uni-agent fork
  4. NEXT04 / LEARNING LOOPS

    Self-improving models

    Research direction · metaRSI

    I want to close the loop: agents that don't just execute, but turn experience into better models.

    How can agent experience become measurable gains in model capability?

    An exploration direction. The next thing to build.