Healthcare/Engineer

Score reimbursement likelihood before the device ships.

A DME startup was writing off revenue because nobody could predict which placements would get reimbursed. We built an AI engine that checks every case against Medicare and Medicaid policy before anything goes out the door.

Client
Healthcare Startup (DME)
Engagement
Ongoing
Focus
AI Agents + RAG
Compliance
HIPAA, Medicare, Medicaid
AI reimbursement engine for durable medical equipment

01

The work

A healthcare startup in the durable medical equipment space, running patient programs that provide medical devices under Medicare and Medicaid coverage. Industry-wide, an estimated 20 to 50 percent of DME placed in the field fails reimbursement and gets written off.

What was getting in the way

  • A meaningful share of devices shipped were never reimbursed. Equipment and effort turned into bad debt.
  • Eligibility logic was scattered across staff, binders, and tribal knowledge. No two people applied it the same way.
  • Checking reimbursement meant manually cross-referencing Medicare and Medicaid rules against patient records. It didn't scale.
  • There was no system to score how likely a case was to be paid, or flag what documentation was missing.
  • When payers denied claims, the team couldn't quickly build a structured argument for medical necessity.

02

What changed

Train AI on real policy. Feed it real patient data.

We took the specific Medicare and Medicaid policies relevant to the client's DME portfolio and structured them into a knowledge layer the AI could reason over. Organized by device type, program, conditions, thresholds.

03

What we built

Then we connected it to live patient data from their EMR via HL7 and FHIR. Diagnoses, clinical history, program enrollment, device data. All normalized and governed for PHI compliance.

The engine matches each case against the relevant rules, produces a reimbursement score, and explains which criteria are met, which are borderline, and what's missing. Staff see the answer before the device ships. If a claim gets questioned later, the documentation is already there.

Under the hood, the engine runs on Anthropic's Claude models: Opus 3 for the heavy policy reasoning and Sonnet for faster, high-volume scoring. Our team built it with Claude Code, which let a small crew move through the policy structuring, integration work, and testing much faster.

AI AgentsRAG PipelinesSystem IntegrationHL7/FHIR IntegrationHIPAA Audited

04

How it works

  • AI engine trained on the Medicare and Medicaid rules that actually govern this client's DME categories
  • Live patient and clinical data flowing from the EMR via HL7 and FHIR into the scoring engine
  • Per-case reimbursement scores with plain explanations of what's satisfied and what's not
  • High-risk cases and missing documentation flagged before a device goes out
  • Justification material for audits and appeals, generated from the same scoring run
  • A platform built to extend as new device types, programs, and payer rules come in

05

Business impact

industry write-off rate
20-50%
industry write-off rate
Of DME placed industry-wide fails reimbursement. That's the baseline we're fixing.
scoring
Real-time
scoring
Per-patient reimbursement likelihood, checked before the device ships
documentation
Audit-ready
documentation
Justification tied to specific Medicare and Medicaid rules, generated automatically
“We used to guess. Now we know before we ship whether a placement will get paid.”
Founder, DME Healthcare Startup

06

What’s next

Stop writing off revenue you should be collecting

If reimbursement decisions at your organization depend on tribal knowledge and manual policy review, we should talk.

Talk to an AI engineerMore work

Team

  • Justin TannenbaumSolutions Architect
  • Lukasz ChmielewskiLead Engineer
  • Mikolaj KaminskiEngineering
  • Daniel BukalaProduct Manager

Stack

.NET · C# · Azure · Semantic Kernel · Claude Opus 3 · Claude Sonnet · Claude Code · HL7 · FHIR

More work

Keep reading.