Building production AI infrastructure from Hyderabad → Germany
Multi-agent AI pipelines, high-throughput backend systems, and production-grade APIs that handle real traffic under real constraints.
At RefractOne I cut LLM latency ~60% and API costs ~40% via async parallel agent architecture on GCP. Outside work I ship independently — two apps live on Google Play, stress-tested at 1.9M+ API requests with P95 < 200ms at 500 concurrent users.
Nexora AI — Executive & Corporate Intelligence Platform
Turn a company domain or a person's name into a boardroom-ready intelligence dossier
- Company track — 12-section AI report (financials, tech stack, competitive landscape, SWOT, analyst verdict); sections 1–10 run in parallel via thread pool, SWOT + verdict synthesize sequentially
- Persona track — 12-node sequential LangGraph pipeline: identity → professional background → skills → personality → online presence → content leadership → social intelligence → network influence → achievements → red flags → engagement playbook → analyst verdict
- RAG chat — Pinecone-indexed reports; ask any question about a saved persona or company
- Stack — FastAPI · LangGraph · Gemini · Node.js · TypeScript · PostgreSQL · Pinecone · React
VendorIQ — AI Decision Intelligence
Multi-agent boardroom simulation for evidence-based vendor selection
Built for Google DeepMind × AI & Big Data Expo North America Hackathon 2026
- 13 specialized agents (CFO, CTO, Legal, Devil's Advocate, Governance Auditor) debate across 3 structured rounds
- Live bias detection per round — biased agents auto-downweighted in final verdict
- Full replay system — debates saved to PostgreSQL, replayable without re-running LLM
- Stack — LangGraph · Gemini 2.5 Flash · FastAPI · PostgreSQL · React · TypeScript
Convoxa — Community Discussion Platform
Production modular monolith — 322 installs, 65 MAU, 1.9M+ API requests
- 3-mode feed ranking (New/Hot/Top), threaded comments, nested voting
- Read-optimised PostgreSQL schema — denormalised aggregate counters, eliminated COUNT() on hot read paths
- k6 stress-tested: P95 < 200ms at 500 concurrent users · homefeed 199 RPS · thread 449 RPS
- Stack — Node.js · TypeScript · PostgreSQL · Redis · BullMQ · Socket.io · GCP Cloud Run
Lepus — Campus Social Platform
Student-only platform with real-time messaging, live bus tracking, QR attendance — 6 Indian languages
- Stack — Node.js · MongoDB · Redis · Socket.io · React Native · Expo · GCP · Firebase
Backend Node.js TypeScript FastAPI Python Express REST WebSockets BullMQ
AI/ML LangGraph LLM Orchestration RAG Pinecone Prompt Engineering Gemini
Databases PostgreSQL MongoDB Redis SQLite Prisma Mongoose
Cloud GCP (Cloud Run · Cloud Build · GCS · IAM) Docker GitHub Actions AWS
| LeetCode | Knight — 1900+ rating (top 3% globally) |
| Load tested | 1.9M+ API requests · P95 < 200ms @ 500 concurrent users |
| Play Store | 2 apps live · 322 installs · 65 MAU |
| LLM latency | ~60% reduction via async parallel agent architecture |
| API costs | ~40% reduction via prompt orchestration + RAG |
- Building AI agent systems at RefractOne, Gurugram
- Actively targeting backend / AI engineering roles in Germany — EU Blue Card eligible, Anabin H+ degree, German A2 → B1
- Open to relocation: Berlin · Munich · Hamburg