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, 40.0% instance-level (12/30, single attempt) on the cheapest model tier (deepseek-v4-flash) — +6.7pp over the previous version in a same-model A/B (33.3% → 40.0%). Approaching Claude 3.5 Sonnet-level agents at ~1/100 the cost. Full methodology, sampling seed, cost accounting, and per-engine breakdown 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.

Research interests:

• Agent reasoning paradigms and evaluation methodology
• Evidence-constrained architectures that treat LLM output as claims to be verified
• Static analysis + semantic IR for deterministic code understanding
• Cost-efficiency in production LLM applications

Currently exploring:

• Cross-language semantic analysis (Python, Go, TypeScript, Rust)
• MCP protocol integration for agent tool ecosystems
• Benchmark design for agent reliability measurement

Seeking: AI Agent / LLM Infrastructure engineering roles (2027 grad, available from July 2027)

What I bring:

• Reproducible evaluation methodology (SWE-bench, upstream bug verification)
• Production-focused architecture design (caching, observability, evidence gates)
• Cross-stack engineering (Python/Go/C++/TypeScript, static analysis, distributed systems)

• GitHub: @xianyu-sheng
• Email: xianyu.sheng.research@gmail.com
• Technical blog: 掘金 @闲余生
• Location: Chongqing / Chengdu, China (open to relocation)

_{💡 Open-source philosophy: Every claim should be reproducible. Every benchmark should publish its methodology. Every "smart" system should show its work.}