Resume / CV
Download PDFPriyank Agrawal
Professional Summary
AI engineer with 3+ years building AI systems that hold up in production. Built and own KarbonIQ, a multi-tool ESG agent that answers questions, generates dashboards and produces Excel/PDF exports, with every figure computed from approved data, so it cannot invent a number. Largest contributor to a 240,000-line platform with 900+ APIs — 41% of its commits, the most of 8 contributors — where I own the database query design and cut its slowest dashboard queries from over three seconds to under one. Mentor a team of 5.
Work Experience
Software Engineer
- Built KarbonIQ, the platform's ESG agent on AWS Bedrock and LangGraph, as a deliberately single-agent system with 21 tool-calling functions rather than a multi-agent one — a router is itself a model call with less context, and every handoff loses conversational continuity quietly. Tools bind per user role and resolve server-side, and every figure is computed from approved data, so the model explains numbers but cannot invent one — cutting manual analysis time by 70%.
- Rebuilt how emission summaries are queried, moving unit conversion and date bucketing to write time: 12× fewer documents scanned per query (38K to 3.2K), with 22 join and expansion stages removed from the read path.
- Cut a settings page from over 3 seconds to under 1 second by replacing a per-item database check with a single grouped query and the missing index it needed.
- Largest contributor to a 240,000-line platform with 900+ APIs, where I own the database query design. Mentor a team of 5.
- Engineered an AI data-ingestion agent (AWS Bedrock vision + OCR) extracting ESG metrics from 8+ document formats (PDF, Excel, scanned images, email), auto-calculating Scope 1/2/3 emissions at 95% extraction confidence; deployed on AWS ECS Fargate.
- Built an ESG Benchmarking Agent over BRSR/XBRL filings for 987+ listed companies (SASB-weighted peer scoring, percentile rankings, gap-to-leader analysis) and a CCTS carbon-credit compliance module.
- Architected a Supplier Due-Diligence Agent that scrapes public data, auto-builds assessments for gaps, and runs an autonomous follow-up sub-agent to collect responses — targeting ~60% faster supplier onboarding.
Associate Software Developer
- Engineered a Resource Management System for 400+ employees with role-based access, automated allocation algorithms, and real-time analytics on Node.js and PostgreSQL
- Built a social platform with interest-based matching, Socket.IO chat, and Agora audio/video calling, reaching 10,000+ downloads in 4 months
- Architected Speedbox, a white-labelled logistics product on serverless microservices (AWS Lambda, Cognito, SNS), sustaining 99.9% uptime and auto-scaling to 10,000+ concurrent requests
- Implemented CI/CD pipelines with GitHub Actions + Jest testing (85% coverage), reducing bugs by 40%
Featured Projects
NextRole
7-agent AI system for job search automation with intelligent CV parsing, job matching, cover letter generation, and mock interview prep. Built with FastAPI, LangGraph, LangChain, OpenAI GPT-4o, and Next.js. Deployed on AWS Lambda + Amplify.
YojanaKhoj
AI-powered government benefits discovery platform for India. A conversational quiz matches user profiles against 101 central and state schemes across 14 states using three-valued rule logic (true / false / cannot-evaluate), then the LLM explains each verdict — it never decides eligibility. UI in 10 languages, scheme content in Hindi and English. WhatsApp bot built, pending BSP approval.
AI Job Finder Agent
Intelligent job search automation that sweeps 35 configured career boards (Greenhouse, Lever and Ashby pages) plus the major aggregators daily, scores each listing against your resume using a Claude → Gemini → OpenAI fallback chain, and delivers the best matches to your email and WhatsApp. Features smart deduplication via MongoDB and LinkedIn HR email extraction for cold outreach.
ESG Agent
An ESG agent that queries company sustainability data, builds dashboards on the fly, produces Excel and PDF exports, and covers framework reporting. Deliberately single-agent rather than multi-agent: 21 typed tools bound per request against the caller's role, with every figure derived server-side so the model routes and explains but never computes a regulated number. Cuts manual ESG analysis time by roughly 70%.
Making a slow ESG dashboard load instantly
A dashboard query took over three seconds. It now returns in under one. I rewrote the read path to denormalise at write time, moving unit conversion and date bucketing to write time — removing 22 join and expansion stages (14 $lookup, 8 $unwind) and cutting documents scanned per query 12×, from ~38K to ~3.2K — added the compound indexes the queries actually needed, and made the backfill resumable so it could be re-run against production safely.
Alumni Connect
Real-time social platform connecting university alumni and students. Features live chat, event management, job board, and role-based access control with college admin approval flow. Built with Node.js and WebSocket for real-time communication.