May 2026 — Aug 2026 · Remote
Posit PBC
Software Engineering Intern
Built a feedback-memory MCP server and an agent harness that raised great_tables generation accuracy from 15.8% to 87.9%.
MCPAgent SDKEvaluationPython
Highlights
- 01
Built Whetstone, a Python FastMCP feedback-memory server achieving a 97.4% stick rate across 38 preferences and consistent application within 2 runs by reinforcing corrections and promoting recurring feedback after 4 occurrences.
- 02
Created GTskill, a Claude Agent SDK harness benchmarking 4 skills against ground-truth tables, raising great_tables generation accuracy from 15.8% to 87.9% for approximately $0.08 per invocation with deterministic checks and a grounded LLM judge.