# mcp/github_mock.py # GitHub mock — 20+ endpoints @app.tool(name="create_pr") async def create_pr(repo, title, body): auth = verify_token(scope="repo") pr = db.prs.insert({...}) return {"number": pr.id} 200 OK · p95 14 ms · tests 327/327
MCP Mock Servers for LLM Training
20+ high-fidelity mock platforms used as controlled training environments for frontier LLMs.
Led a 5-engineer team building mocks of GitHub, HubSpot, Google Drive, PayPal and Slack — faithfully replicating data models, API behaviours and permission systems so LLMs could be trained on realistic tool-calling without hitting production.
- Replicated layered authorization: fine-grained tokens, OAuth, app tokens, repo / org scopes.
- Built a 150+ RLHF evaluation dataset for Microsoft LLMs.
- Designed SFT/CIF alignment for ServiceNow's Apriel-1.5-15B — measurable gains in reasoning & code generation.