Vyankatesh
Sabu I build multi-tenant backends that stay up under load — and AI pipelines that don't get to hallucinate.

Multi-tenant fleet management platform. I own backend features and schema design end-to-end, with isolated tenant stacks running on separate deployments.

Zero-downtime blue-green releases · Redis leaky-bucket rate limiting · Flyway expand/contract migrations

LLM extraction pipeline inside Fleetcore. Per-form-type JSONB schemas drive the extraction, a Strategy layer swaps providers, and a model registry lives in the database rather than in config.

Two-layer dedup (Redis + MinIO) · pgvector entity resolution · full parse-log audit trail

Event-driven pipeline that streams live mailbox events through the Gmail Watch API into phishing classification and an automated quarantine workflow.

OAuth 2.0 with AES-GCM encrypted refresh tokens — no plaintext credentials at rest

Built for a nephrology hospital, where a wrong number matters. The LLM only picks ingredients and quantities; macros are computed deterministically from IFCT and USDA datasets.

±7% validation gate rejects hallucinated output before a dietitian ever sees it

Urban services marketplace with three isolated roles — customer, partner, admin — covering booking, payments and fulfilment, plus a reactive Android client.

28+ production-ready features · JWT multi-role access control · Redis on hot read paths

Four-way finance reconciliation across orders, gateway reports, bank statements and the GL ledger. Deterministic matching runs first; the LLM only adjudicates what's left.

LLM as Tier-3 adjudicator, not as the source of truth · natural-language Q&A over results