Tech Stack
Overview
TravelClaim AI is a hackathon prototype exploring agentic AI workflows for military travel reimbursement. The idea was to let service members describe travel in plain English, ask policy questions against the Joint Travel Regulations, auto-fill DD Form 1351-2, and validate uploaded receipts before the package reaches finance review.
This was built as a concept prototype rather than a deployed production system. The focus was reducing administrative friction in a workflow that is both high-frequency and regulation-heavy.
Problem
Military travel reimbursement is slow and painful because service members have to navigate JTR rules, DTS, per diem rules, receipts, and DD Form 1351-2 all at once. Even a relatively standard TDY trip can become a long administrative task with multiple failure points before reimbursement is approved.
The opportunity behind TravelClaim AI was not just answering policy questions, but turning that knowledge into concrete workflow actions: structured claim data, form fill, receipt checks, and a cleaner handoff to reviewers.
Validation
The team validated the problem through Reddit analysis across military communities. The pitch deck surfaced:
- 822 posts across military subreddits discussing reimbursement and voucher pain.
- 9,754 total upvotes, showing that the frustration is widely recognized and shared.
- Repeated complaints around delayed reimbursement, DTS friction, confusing policy interpretation, and waiting weeks or months to get paid back.
This mattered because it grounded the prototype in a real administrative pain point rather than a generic “AI assistant” idea.
System
- RAG assistant over Joint Travel Regulations — soldiers or reviewers can ask policy questions and get regulation-grounded answers with citations.
- Natural-language travel input — conversational travel descriptions get converted into structured claim fields.
- DD Form 1351-2 PDF autofill — the system generates a filled voucher preview and PDF rather than leaving users with a blank form.
- Receipt upload / vision extraction concept — receipt images can be parsed into structured expense data for review and claim packaging.
- Finance NCO review queue concept — the prototype direction included a cleaner downstream review flow for flagged claims.
My Role
I contributed to the agentic workflow design, product framing, research validation, and technical architecture for the prototype.
What I Learned
- Agentic systems are strongest when they take concrete actions, not just answer questions.
- RAG needs citations and uncertainty handling for regulatory workflows.
- Form automation is valuable when it reduces real administrative pain.
- User research can come from messy public data when handled carefully.