M.S. in Computer Science @ CQUPT (2027). I build AI Agent infrastructure — and I verify it with reproducible benchmarks and upstream contributions, not adjectives.

My design philosophy, shared across everything I build:

Let the model think. Let the program decide. LLM output is a claim; only a tool result is evidence.

An open-source agent runtime built around an online Evidence Runtime: a deterministic, zero-LLM verification chain that runs vertically through the whole task lifecycle (plan → pre-tool → execution → patch-binding → delivery). It treats every LLM statement as a claim to be verified against real tool results — catching "pasted a diff but never wrote the file" hallucinations at delivery time with zero false-kills.

• SWE-bench_Lite official harness: 30 instances, 36.7% instance-level (11/30, single attempt) on the cheapest model tier — approaching Claude 3.5 Sonnet-level agents at ~1/100 the cost. Full methodology, sampling seed, cost accounting, and no-patch attribution are committed and reproducible (report).
• 7 inference paradigms (direct / ReAct / Plan-Execute / Reflection + combinations) over one engine abstraction; MCP-native tool layer; 2000+ unit tests + CI on every commit.
• Hardened via a systematic boundary-probing methodology (malformed input / path-traversal / corrupted-file probes on every public entry point) that surfaced and fixed real defects, including a session path-traversal and a permission-gate fail-open.

A code-diagnosis workbench that splits responsibility deliberately: language frontends and deterministic analyzers own source facts; the LLM may only propose hypotheses; resolvers and validators decide whether a claim binds back to real operations, types, and control flow. Under-supported conclusions stay unknown/abstained — they are never promoted to findings.

• Real upstream bugs found and verified before submission — e.g. stunner #89 (TURN connection leak) confirmed fixed upstream; issues/PRs to sniproxy, Robyn, qscan.

Merged into DeepSeek-Reasonix (DeepSeek's official agent framework):

• #7086 — preserve advisory diagnostics without weakening the evidence gate (merged 2026-07-31)
• #6618 — honour finish_reason=stop on reasoning-only final answers (merged 2026-07-17)

Open / under discussion:

• NousResearch/hermes-agent #80155 — gate finalization on active todos; surfaced a boundary that drew maintainer triage and a "facing the same problem" from the community.
• volcengine/veadk-python #716 — Xenon Skill + MCP interoperability docs.
• deepseek-ai/awesome-deepseek-agent #301 — listing Xenon as a DeepSeek V4 terminal agent.

AI Agent / LLM application engineering roles (秋招 2027 届). I care about the layer between "the model said something" and "the system can trust it" — evaluation, evidence, reliability, and cost.

📫 2107644732@qq.com · 掘金 @闲余生 · 小红书/抖音「手搓 AI Agent」