Manifesto

Building agents
that work when
no one is watching.

Pranav
Garg

AI Product · Strategy · Engineering

New York, NY  ·  Open to Relocate

Building agentic systems that run in production, not just in demos.

View ProjectsRead Blog
LinkedInGitHubXpranavgarg71@gmail.com
About

AI engineer at EXL Services in New York, building multi-agent workflows for enterprise data migration and finance operations on the Data.AI platform. I work on LLM orchestration with LangGraph, RAG pipelines, and evaluation frameworks that let agents learn from their own failures — from architecture through client onboarding.

MS Business Analytics from UT Austin McCombs (GPA 3.8, May 2025). Previously at Kimberly-Clark as Solutions Engineer, where I shipped a GenAI procurement chatbot that cut query resolution time 70% across 100K+ supplier interactions.

Now

3 Things I Am Building

Writing

Latest Posts

01NemoClaw & DeepAgents: A Technical Deep DiveInteractive CourseJul 202602What It Actually Takes to Ship an AI Agent to ProductionJul 202603Predicting ICU Mortality Using Machine Learning: A Global PerspectiveMediumNov 2024View All Posts →
Work
Projects
01
Claude Code Chat Viewer
Zero-build single-file HTML viewer for Claude Code conversation history — drag-and-drop .jsonl files to explore messages, tool calls, thinking blocks, token usage, and session cost estimates with full-text search and a collapsible TOC sidebar.
HTMLJavaScriptmarked.jsPrism.js
02
Intelligent Customer Support Router
GenAI document-analysis and classification system with LangGraph StateGraph, conditional routing, verification nodes, and a RAG layer for real-time knowledge retrieval — 90% processing accuracy.
LangGraphRAGVector DBLangChainPython
03
Investment Analytics Pipeline
Production-grade portfolio tracking system — Airflow orchestration, dbt transformations, Sharpe Ratio and Alpha calculations, Streamlit dashboards, Great Expectations data quality.
Apache AirflowdbtPostgreSQLStreamlitDocker
04
Connect 4 AI Bot
CNN and Transformer models trained via MCTS self-play on 1.8M+ game states. CNN hit 76% accuracy on spatial tactics; deployed on AWS Lightsail via Docker.
TensorFlowPyTorchRLDockerAWS
05
ICU Mortality Prediction
ML models across MIMIC-III and eICU datasets using SMOTENC for class imbalance and PCA for dimensionality reduction. MLP achieved 0.89 recall, outperforming clinical APACHE scoring.
Scikit-learnTensorFlowSMOTEPCAPython
06
NLP Fragrance Recommender
BERT-based topic modeling over 44,000+ fragrance reviews scraped from Base Notes forum. Sentiment-weighted recommendations served via Flask API with SQLite storage.
BERTFlaskSQLiteBeautifulSoupNLP
View All Projects →
Education