Applied AI & ML
› ▍
open to workEvidence-traceable obligations, risk alerts & summaries from operational documents
AI course generation platform
Coding, Classification & Naming (CRUD) Agent
continuously refining
Hardware
MacBook Pro M4 Pro
iPhone 17 Pro
Samsung Z Fold 7
Software
VS Code
Claude
Codex
Gemini
Ghostty
Agentic workflows
in construction and real estate.
Currently
House of Cards
Show · 2025
Past
Industry
Silicon Valley
Severance
Succession
AI engineer building production LLM systems — RAG pipelines, semantic search, multi-agent workflows. Originally from New Delhi, based in Chicago. Outside the terminal, I'm an avid football fan (the real kind), play guitar, and spend too much time thinking about music and art. I think good taste is a technical skill.
Minor in Architecture
Illinois Institute of Technology
Expected May 2026
After graduation in May 2026, I want to join a team where I can ship AI systems that actually matter — whether that is at a fast-moving startup or a company building infrastructure for the next wave of intelligent software. Longer term, I want to build products of my own at the intersection of AI and systems design, and eventually contribute to research that makes LLM pipelines more reliable and interpretable.
The Syntec Group
Chicago, Illinois
May 2025 – Present
Remote (India)
May 2021 – July 2021
Excelerate (Globalshala)
New Delhi, India
June 2023 – July 2023
Built a staging-deployed SaaS that generates research-backed courses using a multi-stage agent pipeline, orchestrating 5 stages with persisted job state, retry policies, and failure isolation to keep long-running workflows reliable.
FastAPI, Postgres/pgvector, Redis, Inngest, SvelteKit, LiteLLM, PydanticAI, Docker, Sentry, PostHog
Built an agentic Coding, Classification & Naming system using OpenAI function calling that enables users to query, add, update, and delete BIM module codes through natural language with automated sub-code assignment.
Flask, OpenAI Function Calling, SQLite, ChromaDB, React, Docker
Built an evidence-traceable document intelligence platform that extracts obligations, risk signals, and structured summaries from operational documents using LLM pipelines with source-level citation.
FastAPI, LLMs, Vector Search, PDF Ingestion, SvelteKit
Shipped a semantic search and retrieval-augmented chatbot using GPT-4o and ChromaDB embeddings to deliver question answering across PDF documents, CSVs, blog posts, and website content with automated source citations.
Flask, React, Vite, ChromaDB, OpenAI GPT-4o, DeepSeek, Redis, Docker, Nginx, Vercel
RAG system for real estate investment queries with multi-path retrieval (pgvector + SQL fallbacks with MMR), OCR processing, and automated table extraction from PDF brochures.
FastAPI, Postgres/pgvector, OCR, WhatsApp Business API, Airtable
Cotton price forecasting using Facebook Prophet with external regressors (oil, gas, soybeans); evaluated with MAPE.
Python, Prophet, Pandas, Matplotlib
Built an end-to-end soccer video analysis pipeline using YOLOv8 detection and ByteTrack tracking, adding temporal smoothing for stable motion analytics and CPU only local processing with robust cross-platform video input and output.
Python, YOLOv8, ByteTrack, OpenCV
Technical work only matters if others can act on it. My approach: meet the audience where they are — precise specs for engineers, concrete outcomes for stakeholders. A confusing interface is a communication failure. In AI especially, closing the gap between what a model does and what people think it does is one of the highest-value skills on a team.
Audience: Faculty, industry judges & peers
Sole designer of this poster for Illinois Tech's IPRO Innovation Day. The format demands compression — every word earns its place — and the mixed audience of faculty and industry judges forced me to layer technical depth with accessible framing. Fielding live questions sharpened my ability to adapt on the fly.
Audience: General public, patients & families across India
Co-founded a volunteer helpline during India's second COVID wave, writing outreach and coordination copy under real pressure. High-stakes, time-sensitive, directed at people in distress — clarity and empathy couldn't be traded off. Taught me that good communication removes friction for the reader, not credit for the writer.
Audience: Engineering team & non-technical stakeholders at The Syntec Group
Delivered SyntecAgent with docs written for two audiences simultaneously: developers maintaining the code and managers approving its use. Modularizing the document into clearly separated layers — rather than blending jargon with prose — made both versions sharper. Good technical writing is information architecture.
Audience: Potential users, collaborators & early investors
Pitching Clutch meant arguing why the product should exist to people with no reason to care yet. I led with the problem, not the tech — a deliberate choice that respects the audience's skepticism. Compressing a multi-stage AI pipeline into one compelling page made me a sharper communicator across every medium.
Clean, chunk, and index messy docs so retrieval starts from ground truth
Vector search alone hallucinates on structured data — combine with lexical matching and rerank
Function-calling agents with confirmation gates before anything destructive touches production
Every LLM response parsed through Pydantic schemas — if it doesn't conform, it doesn't ship
LiteLLM across models, Redis caching, and fallback LLMs — not every query needs GPT-4o
Inngest job tracing, eval harnesses — catch breaks before users do