.png&w=3840&q=75)
Zebra AI
The Challenge
Managing transactional consistency, preventing credit double-spending race conditions, and handling brittle un-structured LLM parsing errors under concurrent workloads.
What I Built
A resume-tailoring backend engineered with Next.js 15 and Drizzle ORM over a Neon serverless Postgres pool.
Technical Work & Hardening
- Eliminated N+1 Query Fan-out:Migrated the main dashboard read path from a sequential point-lookup loop into a single relational join database roundtrip.
- Atomic State Integrity:Replaced a 'check-then-spend' application-level logic flow with atomic SQL transactions inside Drizzle, enforcing token balance verification at the exact database decrement boundary.
- Deterministic Payload Structuring:Dropped fragile substring regex methods ({}) for parsing raw LLM responses; implemented strict JSON schema-enforced structured outputs natively via the Gemini SDK.
