Vaibhav Kumar.
I build production AI systems — multi-agent orchestration, RAG services, and LLM pipelines — on top of large-scale data engineering. Currently at Plivo, where my agents automate 100% of US messaging compliance vetting and classify 2M+ messages a day. Previously moved 50M+ daily transactions at Razorpay.
Experience
Four years shipping systems that stay up: agents, pipelines, and the data infrastructure underneath them.
AI & Data Engineering @ Plivo
- Built a multi-agent compliance vetting system — 4 specialized agents (brand validation, content extraction, campaign generation, decisioning) with tool-calling and structured outputs across Claude, OpenAI, and Azure OpenAI. Took accuracy from 62% → 84% and turnaround from days → 6 hours, automating 100% of US messaging compliance vetting.
- Shipped an AI support agent: a 12-step orchestration with 11 tools — SQL retrieval, vision-based document parsing, RAG search, Zendesk and Slack integrations — that auto-resolves customer tickets across 7 categories.
- Built a RAG knowledge service from scratch: FastAPI + LanceDB with 14,492 vectors over docs, runbooks, and past tickets; hybrid filtered vector search with quality-based reranking.
- Engineered a fraud classification pipeline processing 2M+ daily messages with parallel batch scanning of 10M+ historical records — cut manual audits by 70% and response time by 85%.
Technical Consultant @ Razorpay
- Designed payment analytics infrastructure: star-schema warehouse ETL moving 50M+ daily transactions into Redshift, with reconciliation across MySQL, PostgreSQL, and MongoDB at 99.97% accuracy on 200GB+/day.
- Improved dashboard P95 latency 850ms → 320ms through indexing, caching, and query rewrites.
- Owned integrations for 20+ enterprise clients — REST APIs and webhooks at 99.9% delivery across 2M+ daily API calls.
Live on the Internet
Not slides, not repos — things you can click and try right now.
fcstudio.co.in e-commerce, in production
A real store with real customers — storefront, payments, and shipping integrations built and operated end-to-end since 2024. Next.js · MongoDB · Razorpay · Shiprocket.
Visit the store LIVEThe Great Train Robbery WebGL, zero libraries
A single-file wild-west heist animation written in raw WebGL — no frameworks, no build step, runs entirely offline. Built purely for the joy of it.
Watch the heistThe Glow-Up 📈
Every dev has an origin story. Mine is kept live — for humility.
- 2022Weather App still live, adorably ↗ The first thing I ever put on the internet: one search box, one API call, and unreasonable pride. Preserved in amber.
- 2022Sign Language Translator College capstone gone right — real-time deep learning, and a Project of the Year award.
- 2024fcstudio.co.in live ↗ First thing with actual customers and actual money moving through it.
- 2024+Production AI at Plivo Multi-agent systems vetting compliance for real US carriers — the pipelines grew up.
- 2026The Great Train Robbery live ↗ Came full circle: back to building toys for fun, except now the toy has shaders.
Things I've Built
Production systems and deep-learning projects — agents, RAG, and vision pipelines.
Multi-Agent Compliance Vetting
Four LLM agents with tool-calling and JSON-schema outputs vet messaging campaigns end-to-end. Event-driven concurrency control in plain SQL, webhook callbacks, and A/B prompt evaluation baked in.
RAG Knowledge Service
Embedded vector search microservice: LanceDB, Voyage embeddings, per-source chunking strategies, and reranking driven by human quality ratings. Serves three search indexes to production agents.
AI Support Agent
A 12-step agent that reads support tickets, routes intent, pulls context from databases and RAG, parses uploaded documents with vision models, and drafts resolutions — across 7 ticket categories.
Sign Language Translator
Real-time sign-to-text: MediaPipe landmark extraction over 150K+ augmented video samples feeding a bidirectional LSTM — 85% accuracy at 30 FPS with sub-50ms latency.
AI Drawing Analysis (DDSL)
Vision models read architectural fit-out sheets; a deterministic calc engine turns the recognized schedules into auditable material quantities. CAD files parsed directly for ground-truth counts.
Toolbox
What I reach for, grouped by how I actually use it. Don't know one? Click it — you'll get the ELI5.
AI / LLM
Data Engineering
Languages
Cloud & Tools
Get In Touch
I'm interested in AI engineering roles and hard problems at the intersection of agents, data, and production reliability. My inbox is open — say hi.