Mayank Goel

AI training simulator for hospitality front-line staff — practise hard guest conversations against a realtime voice AI, with transcript-grounded coaching

• Built a realtime voice pipeline on pipecat and WebRTC with a containerized Python worker, diagnosing and fixing a production connectivity failure caused by TURN relay selection
• Designed multi-tenant Postgres schema with row-level security, verified by adversarial cross-tenant testing against production
• Maintained a 946-test suite behind six CI workflows including two hard gates — full-history secret scanning, 173 tenant-isolation tests, and a weekly backup-restore round trip that performs the restore
• Built a gold-set evaluation harness for LLM grading quality, catching a token-truncation bug that silently corrupted scored reports

WhatsApp-native AI product photography for Indian micro-sellers — send a product photo on WhatsApp, get a brand-ready ad back in minutes

• Profiled an image quality-gate pipeline and found decode and resize accounted for ~99% of runtime against under 1% for the pixel arithmetic; eliminated five redundant decodes of the same image, cutting per-image latency 57% (40.6ms to 17.4ms) with byte-identical output
• Built and benchmarked a Go reimplementation of the same pipeline before adopting it, measuring it 2x slower because the workload was decode-bound rather than compute-bound, and rejected the rewrite on the evidence
• Built an AI image-generation pipeline with cross-provider failover (Gemini 3 Pro Image to OpenAI gpt-image-2), gating every result on deterministic quality checks so a degraded image escalates a tier or is refunded rather than delivered — and cut a planned Gemini Flash middle tier after testing showed it refused the same content as the primary, since both share a safety backend
• Architected a pnpm monorepo of 3 deployable apps and 10 shared packages — a Fastify HTTP API, a queue worker and a Next.js storefront over PostgreSQL and Prisma