VSVenkatesh Shanmugam

Senior AI Engineer · New York

Venkatesh Shanmugam

I build multi-agent systems on top of knowledge bases and knowledge graphs: agents that read filings, CRM history and clinical guidelines, answer with citations a person can open, and pass evals before they ship. As the first AI engineer at Aithon, I own that stack in production, down to the model infrastructure underneath.

  • Multi-agent systems
  • Knowledge graphs
  • Knowledge bases & RAG
  • Evals
  • LLM infrastructure
See my work Résumé LinkedIn GitHub
~10,000 companies

watched by the lead generation agents I built alone, which became Aithon's most adopted product. They read filings, news, job postings and CRM history so sellers don't have to.

42.7M AI calls

in 30 days across 17 services, all through the LLM gateway I built. Every call is traced and priced from a dated price table; the count comes from the gateway's own ledger.

~$8K a month

of AWS Textract spend retired by moving document parsing to a self-hosted open-source model on a GPU endpoint that scales to zero. Financial-table cells captured went from near zero to 82%.

Selected work

Two live agent systems I built and evaluated

Both run in public. Each case study covers the problem, what I built, the decisions that mattered, how I measured it, and what still fails.

The guideline knowledge graph: 10 guidelines, 413 recommendations
Multi-agentKnowledge graphHealthcare

Between-Visit

A graph-backed agent system that turns a hospital discharge into a nurse-approved 30-day care plan, then responds to patient texts, lab results and missed refills.

Guidelines become a queryable graph; patient events change which recommendations apply.

232 of 239 care events caught and 0 unsafe actions across 30 simulated patient months (synthetic patients; a standard care program caught 165)
Read the case study Explore the graph ↗ Live demo ↗
The building code as a graph: 5,961 nodes, 8,421 links
Graph-RAGLangGraphOpen source

VA Building Code Chat

Agents that answer questions about the Virginia Construction Code from a graph of the whole code. Every section or table an answer names is checked against the graph, so a made-up citation can't pass as real.

Citations are verified by code, not by the model.

67% → 100% of answers citing the right section (12-question end-to-end eval, before and after a retrieval rewrite)
Read the case study Ask the code ↗ Source ↗

Experience

Where I've built

Nearly five years shipping production software.

Oct 2025 — Present

Senior AI Engineer

Aithon · New York, NY

First AI engineer at a pre-seed startup building AI for B2B sales teams, promoted to Senior. I own the production AI layer end to end.

What customers use

Production AI systems

  • Lead generation. Agents fuse 10+ sources into one view of each company, with a score and the dated evidence behind it, so sellers know which accounts to work and why now.
  • Company research agent. Finds each company's filings and web presence on its own and builds one profile, searching 917K documents and citing every claim.
  • CRM agent. Reads each customer's own Salesforce or HubSpot as it is, works through thousands of past deals, and keeps a memory for every account.
  • Customers. Onboarded 6 customers and shipped the fixes for why some users weren't adopting. Cut the main screen from about 80 seconds to half a second.
What it runs on

AI infrastructure

  • LLM gateway. The one path every service uses to reach a model: five providers, routing, retries, caching, tracing and per-call pricing.
  • Self-hosted document parsing. Replaced AWS Textract with an open-source model on a SageMaker GPU endpoint that scales to zero.
  • 12+ production services on AWS (ECS, SageMaker) with Terraform, CI/CD and monitoring, each customer's data kept separate.
  • Eval gates. Every prompt, retrieval and model change replays a golden set checked by our CEO and CTO before it ships.
May 2024 — Sep 2025

Machine Learning Engineer

ScriptChain Health · Washington, DC
  • Architected the multi-agent platform (LangGraph) that writes nutrition and exercise plans for metabolic-syndrome patients, with NeMo Guardrails keeping each plan within AHA and U.S. Dietary Guidelines.
  • Predicted 30-day hospital readmissions at about 78% ROC-AUC with a graph attention network plus a time-aware LLM.
  • Cut preprocessing time 99% on 500M+ records by reworking the AWS pipeline, and halved GPU memory use.
Aug 2024 — May 2025Data Consultant, The George Washington University. Statistical consulting for researchers; led 5+ ML workshops for 100+ participants.
Apr 2021 — Aug 2023Software Engineer, Tata Consultancy Services. Automation for 14+ telecom client teams; 12+ RPA bots that cut infrastructure costs 40%.

About

The unglamorous parts that make AI trustworthy

I care about checking every citation, writing the answer key before the system, and knowing what each model call costs. MS in Computer Science from George Washington University (3.88), B.Tech from SRM University, and four papers published as an undergraduate.

Core stack Python · TypeScript · SQL · LangGraph · Claude, OpenAI, Gemini · RAG and knowledge graphs · Neo4j · PostgreSQL · Redis · FastAPI · React / Next.js · AWS (ECS, SageMaker, Bedrock) · Terraform · Databricks

Publications (2018 – 2020)