AI Agent / LLM Application Developer
I build agent systems for cross-border commerce - profit intelligence, multi-store operations, CRM workspaces - and send every bug I hit back upstream as a pull request.
- ⚙️ Agent Runtime — bounded tool loops, capability leases, durable workers, and approval before any action
- 📚 RAG & Evidence — hybrid retrieval,
no_answerrefusal, chunk citation checks, and reproducible retrieval evals - 🔗 LLM Tooling — native function calling, read-only MCP servers, and deterministic fallback when no API key is present
- 🌏 Cross-border Commerce — profit attribution across FX, platform fees, refunds, and settlement
- ⚡ ProfitOps-AI ·
PythonLangGraphFastAPIDeepSeek— computes margin drift from the real ledger, attributes it to accounting items, then proposes approval-gated simulations with a full audit trail - 🏬 Liteclaw ·
TypeScript— organises support, sourcing, and ads into role-scoped agents with auditable capability boundaries - 🗂️ ai-native-crm ·
TypeScriptMCPPostgres— keeps contacts and deals as the agent's state model, exposing tools, approvals, and data changes on every run
Updated 2026-11-03 · 4 merged · 21 in review · 14 issues filed
| Project | Contribution | PR |
|---|---|---|
| agno-agi/agno | Correct FlexAI DeepSeek image-input model id in cookbook README | #11401 |
| bytedance/deer-flow | Report video generation failures with a nonzero exit status | #7034 |
| crewAIInc/crewAI | Keep runtime input prompts on stderr | #8297 |
| helsome/folio | 校验规则保存的 IPC 输入 | #276 |
| Project | Contribution | PR |
|---|---|---|
| agentscope-ai/agentscope | Validate session ownership before returning run status | #3187 |
| agentscope-ai/agentscope | Preserve missing final newlines in file diffs | #3244 |
| agentscope-ai/agentscope | Match newline-containing filenames in glob patterns | #3386 |
| assafelovic/gpt-researcher | Image generation config read with wrong (uppercase) attribute names - feature... | #2246 |
| assafelovic/gpt-researcher | Config-file REASONING_EFFORT silently reset to medium by _set_llm_attributes | #2247 |
| assafelovic/gpt-researcher | DOC_PATH and REPORT_SOURCE env overrides ignored by doc path setup | #2248 |
| browser-use/browser-use | Fix tab/space mix in langchain example snippet so it parses | #6259 |
| browser-use/browser-use | SchemaOptimizer RecursionError on recursive pydantic models | #6261 |
…and 13 more open PRs - see them on GitHub
Also reported 14 issues upstream (mostly crewAI, agentscope, MetaGPT).
- Agent runtime: tool loops, state, approvals, and failure recovery
- Retrieval quality: citation traceability and refusal when evidence is thin
- Cross-border profit analysis: FX, fees, refunds, and settlement
- Python and TypeScript services that stay debuggable without an LLM
This page updates itself daily from public GitHub data.