Here's a description you can paste straight into the form: Downshift is a self-improving harness for data questions on MongoDB Atlas. The first time a new kind of question arrives, a frontier model (Claude Sonnet 4) writes the aggregation pipeline, answers it, then generalizes it into a skill: a template with typed parameters and its own test questions. A promotion gate only makes the skill live if a cheap 8B model (Llama 3.1) can use it correctly. If it fails, the frontier reflects on the passing and failing traces and rewrites the skill. From then on the cheap model answers that kind of question, protected by code guardrails: typed and grounded parameters, date checks and read-only pipelines. When the schema changes, an Atlas change stream flags every affected skill. The frontier rewrites each one, and it only ships if every stored answer is identical. Results on held-out questions: 100% accuracy at $0.003 per question including learning, vs 37% for the cheap model alone. After a skill exists, answers cost about $0.00002 and take under a second. A field rename across 5,000 documents was repaired automatically, with 21/21 answers identical. Skills, ledger, events, schema versions and repairs all live in MongoDB.
Team
Placement
Participant
