Lead engineering for an AI-native CAT exam-prep platform (10K+ students): Next.js student app, admin console, NestJS/PostgreSQL API, and a React Native app live on iOS and Android.
Architected and built Opto, the platform’s personal AI coach, end to end: agent design, student context and memory, safety and evals, and production infrastructure.
Built Opto’s real-time voice pipeline (streaming STT → reasoning model → sentence-level TTS over WebSockets, with barge-in), 2.5 s median reply; removed a hard 20-concurrent-user ceiling with per-utterance STT leases and a Postgres limiter shared across autoscaled Cloud Run instances.
Cut worst-case AI coach cold start from 60+ s to under 20 s by replacing an unbounded dependency wait with a timed race and instant status feedback.
Led production debugging and performance work: traced a crawler traffic surge to database connection-pool limits, and cut the founder dashboard from ~190 to ~50 queries per load.
Migrated Firestore to PostgreSQL (Prisma); built the GCP infrastructure in Terraform, with CI/CD, Langfuse LLM tracing, a per-call cost ledger, alerting and logging.
Zarle Infotech - Software Engineer
June - August 2025 · Noida
Built Next.js/TypeScript interfaces for Optima Learn as an agency engineer, with accessibility in mind; later joined full-time as its Founding Engineer.
Integrated Firebase Auth, Firestore and Cloud Functions; shipped production releases on Vercel.
GEMS of JCI - Software Development, Research & Data Analysis Intern
September 2023 - May 2024 · Delhi
Built social-platform features with React/Redux, Node.js/Express and MongoDB in a team of three.
Cut data-processing time 20% with aggregation pipelines and indexes.
Projects
Askcode
January 2025 · Next.js, tRPC, PostgreSQL/pgvector, Prisma, LangChain, Gemini
Built a RAG assistant for GitHub repositories that embeds file summaries into pgvector and retrieves context for natural-language questions.
Implemented cosine-similarity retrieval over 768-dimensional embeddings, token-streamed answers through the Vercel AI SDK, and prompt-level grounding.