Rohan Bansal is a backend engineer exploring how to train small models for useful technical tasks. His work includes QORL, a query-planning experiment that learns from actual PostgreSQL execution times, and an account of Traba’s migration from Firestore to Postgres.[3][1]
Start with Training a 4B model to produce 81% faster query plans than Postgres →
QORL repository 
X Career
Bansal’s site describes being Crosby’s first engineer, building precedent search and evaluation tooling. Earlier he worked at Traba on warehouse-shift matching and engineering workflows, and at Google on conversational AI for support. Those roles precede his current independent experiments with small models, interfaces and developer tooling.[3]
Contributions
Reward plans that run faster on the actual database
QORL trains a four-billion-parameter model to propose PostgreSQL plan hints, execute candidate queries and compare their runtimes with the database’s default plan. The model can make several attempts and select a candidate or keep the default. Reinforcement-learning rewards come from measured relative runtimes, with controls for noisy execution. Bansal’s reported speedups apply to the tested workloads and measurement setup, not arbitrary production databases.[1]
Change the data model without replacing everything at once
Bansal’s Traba migration report explains why increasingly relational staffing data became difficult to query in Firestore. It traces preparation and an incremental move to Postgres, including how the team kept services working while changing storage. The practical concern is the system around the database: connected records, application behavior and operational checks must migrate together.[4]
Tools, writing & talks
Writing, research & talks
Start here
An implementation and evaluation report on reinforcement learning from measured PostgreSQL query performance.
Start with the optimizer and hint examples before the training details. SQL, query-plan and basic machine-learning familiarity help; compare the reported workload and baseline with your own before generalizing a speedup.
Author and experiment creator[1]Original sourceAn engineering account of moving a growing staffing application from a document database to relational storage.