u
utkarshg7

Utkarsh G

@utkarshg7

AI Engineer building LLM Agents RAG Chatbots and Voice AI

India
Hindi, Engels
Sommige informatie wordt in het Engels weergegeven.
Over mij
I build AI agents that survive real users, plus the backend that keeps them running. 20 months shipping production AI at two startups, not demos. What I build: - LLM chatbots and agents with tool calling against your live data - RAG pipelines with guardrails and evaluation harnesses - Voice AI: real-time agents, speech-to-text, latency tuning, Hindi/English - Backends in Python and Go: APIs, WebSockets, workflow engines Results: cut an AI copilot's token cost 17%, lifted answer accuracy 40% on a multilingual pipeline, scaled a call platform to 4,000 concurrent calls.... Lees meer

Skills

u
utkarshg7
Utkarsh G
offline • 

Bekijk mijn diensten

AI-integraties
I will build a production ai chatbot with rag over your docs
VanafUS$ 95 / projectMeer informatie
Full stack webapplicaties
I will build a full stack web app with ai features built in
VanafUS$ 95 / projectMeer informatie

Werkervaring

Opptimise

Full Stack Developer

Opptimise • Fulltime

Jan 2026 - Present • 9 mos

Build and own production AI features end to end on a behavioural team-analytics platform. - LLM coaching chatbots in Python and FastAPI, with per-organisation competency frameworks injected at request time so each client is coached against its own rubric - SigV4 request-signing middleware with SSRF protection for service-to-service calls on AWS (ECS, Lambda, IAM), secrets moved to SSM - Replaced platform-wide permissions with a per-job authorisation model across production subdomains - Built the CI gate on GitHub Actions (lint, tests-required check, integration suites, migration safety) and cut build spend 40-50%

Finn_— AI Voice Agents for Enterprise

Founding Engineer

Finn — AI Voice Agents for Enterprise • Fulltime

Jan 2025 - Jan 2026 • 1 yr

Founding engineer on a production voice AI platform for financial services, building the backend that ran live customer calls at scale. - Built the sequencing execution engine from scratch: a durable state machine with atomic step-claim preventing double-advance, driving multi-step voice, SMS, WhatsApp and email cadences - Fixed three production races in the call pipeline: Redis ZSET leases for per-campaign and global concurrency caps, a number-allocation leak stranding numbers on 7 of 10 campaigns, and a 2x count inflation marking campaigns complete at half their true progress - Migrated realtime WebSocket and call-status microservices from Node.js to Go, scaling the pub/sub path to roughly 4,000 concurrent calls at production peak - Built the agentic layer: an MCP server exposing the voice agent to Claude, in-call tool calling against live application status, multi-agent RAG with guardrails, and an eval harness; cut AI copilot completion tokens 17% - Fixed the multilingual post-call analysis pipeline for Indian languages, lifting answer correctness 40%