Block or report alanxurox
Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.
You must be logged in to block users.
Contact GitHub support about this user’s behavior. Learn more about reporting abuse.
Report abuseI build AI systems that survive contact with production: explicit tools, approval gates, observability, deterministic evaluation, durable state, and verified delivery.
Based in Prague. Native Mandarin, professional English. Former GenDigital Senior Software Engineer (AI Focus); Babson + Olin background in technology entrepreneurship, analytics, and engineering.
- Production agent systems — led engineering for an enterprise email-delivery agent at GenDigital; authored 13/13 PRs from prototype to production across guided workflows, persistent chat state, retrieval, preflight, and Langfuse observability
- Agent reliability and evals — queue/drain behavior, memory, permissions, scheduling, send acknowledgement, rubric leakage, and human-grounded evaluation
- MCP and enterprise integration — template retrieval, Jira, BigQuery/Postgres, enterprise search, and WeChat content tooling
- Product systems — production campaign platforms, audience targeting, experimentation, and translating customer incidents into bounded requirements
| Project | What it demonstrates |
|---|---|
| mission-control | Zero-dependency coordination for agent fleets using Bash + SQLite. Task queues, health checks, and fleet state. 36 stars / 6 forks |
| mcp-weixin-spider | MCP server for crawling WeChat public-account articles with browser backends and anti-bot handling |
| oh-my-claudecode PR #839 | Merged standalone-hook reliability guard in a 37K+ star multi-agent coding framework |
| better-glean | Precise enterprise-search patterns across 40+ data sources |
Some work is private or enterprise-internal; these descriptions intentionally omit internal URLs, customer data, and infrastructure identifiers.
- Production email agent — guided end-to-end delivery workflow, human approval, persistent chat state, catalog retrieval, targeting/QA preflight, and observability
- WeCom conversation archive — sole-authored async ingest, RSA decryption/key rotation, scoped query API, Postgres schema, and media preservation; 30 commits
- AI-assistant reliability program — reproduced failures across queueing, durable memory, timers, permissions, rendering, and egress; converted evidence into issues and SOPs
- better-litellm — internal zero-dependency Node.js proxy; six releases, setup reduced from roughly 30 minutes to 2 minutes, command-injection paths fixed
- Generated is not delivered. Completion requires transport, acknowledgement, and user-visible proof.
- Prompts cannot repair missing state. Persistence, identity, scheduling, and routing need explicit system contracts.
- A valid eval needs the right oracle. Capability smoke tests are not evidence of persona or workflow alignment.
- Claims need evidence. I distinguish shipped work, measured outcomes, hypotheses, and work in progress.
AI systems: LangGraph, Model Context Protocol, Langfuse, LiteLLM, human-in-the-loop workflows, deterministic evaluation
Languages: Python, TypeScript/JavaScript, SQL, Bash
Application: FastAPI, Node.js, React, Playwright/Puppeteer
Data: PostgreSQL, SQLite, Chroma, BigQuery
Infrastructure: GCP, AWS Bedrock, Vertex AI, Docker, Kubernetes, TeamCity
- GenDigital — Frontend Developer → Software Engineer → Senior Software Engineer (AI Focus), October 2023–May 2026. Norton, Avast, and AVG; the brands reach 500M+ users
- Babson College — BS, Technology Entrepreneurship + Business Analytics (STEM)
- Olin College of Engineering — cross-registered in Design & Computing
Pinned Loading
If the problem persists, check the GitHub status page or contact support.