Skip to content
Back to Work
002

Optum

Automating RFP Response at UnitedHealth Group

Role

Forward Deployed AI/ML Engineer

Duration

Feb 2026 – Present

Team

Enterprise AI/ML Team

Status

Current

Private Repo

Overview

The business receives tens of thousands of RFPs. Each one is a dense document that has to be parsed, routed to the right teams, answered, reviewed, and returned on a deadline. That work used to live in email threads, spreadsheets, and tribal knowledge. I build the platform that replaces it. AI parses incoming RFP documents into structured, answerable questions, and every team involved works in one place to communicate, draft, and complete the response. The hardest technical problem is parsing accuracy. If the system misreads a requirement, everything downstream is wrong, so improving extraction accuracy is the constant focus.

Problem

An RFP is not one document for one team. A single response pulls in multiple teams, each answering their own sections against a shared deadline. Before the platform, that coordination happened over email and spreadsheets. Questions got missed, answers got rebuilt from scratch, and nobody had one view of what was done. The source documents themselves are hostile inputs: long PDFs, Word files, and embedded tables in formats that change with every issuer. Parsing them wrong is worse than not parsing them at all, because a misread requirement produces a confident, wrong answer.

Approach

  • 01Parse incoming RFP documents into structured, answerable questions with AI document extraction. Accuracy here gates everything downstream
  • 02Centralize every team that touches an RFP into one platform. Assignment, drafting, review, and communication happen in the same place
  • 03Treat parsing accuracy as ongoing work, not a one-time build. Keep improving extraction against the real documents the business receives
  • 04Put questions in front of the right team inside the platform instead of an inbox
  • 05Work forward deployed: build directly against the business's real RFP workload and iterate with the teams completing responses in the tool
  • 06Handle sensitive healthcare business data under enterprise security and compliance requirements from day one

Design Decisions

Technology Stack

Languages

PythonTypeScriptSQL

AI/ML

LLMsDocument ParsingOCREvaluation Pipelines

Infrastructure

DockerKubernetes

Compliance

Enterprise SecurityAudit Logging

Impact

Volume

Tens of thousands

RFPs the business receives. Every one flows through parsing, so accuracy improvements compound across the whole pipeline

Teams

One platform

All teams involved in an RFP communicate and complete responses in the same system instead of email and spreadsheets

Core metric

Parsing accuracy

The number that gates everything downstream. Improving it is the central ongoing technical work

Next Case Study

MedVanta Platform