VetAssist: An AI Claims Assistant for Veterans, Built in a Garage

A free tool for veterans who deserve better than a spreadsheet

February 2026 · Cherokee AI Federation · ~11 min read

The Ask

It started with a message. A doctor who works with veterans reached out: "Can your AI help veterans navigate the VA claims process?"

The VA disability claims system is notoriously complex. Veterans often file incomplete claims, miss conditions they are entitled to, or do not understand how the combined rating formula works. The difference between a 60% and 70% combined rating can mean hundreds of dollars a month in compensation — compensation that was earned, not granted. And the math that determines these ratings is anything but obvious.

Commercial tools exist. Some of them are quite good. But they charge money — sometimes hundreds of dollars — to people who are already struggling. A veteran dealing with PTSD, chronic pain, and the bureaucratic fog of the claims process should not also need a subscription fee to understand what they are owed.

We decided to build something free.

What VetAssist Does

VetAssist has four core features. Each one addresses a specific point where veterans get stuck in the claims process.

Combined Disability Calculator

Implements the 38 CFR 4.25 combined rating formula exactly as the VA uses it. Veterans enter their individual ratings and see the combined result, the intermediate math at each step, and the final rounded value. No approximations, no "estimates" — the same formula the VA rating board applies to their case.

Evidence Checklist Builder

For each claimed condition, VetAssist generates a checklist of the medical evidence the VA typically looks for when evaluating that condition. This includes diagnostic tests, specialist evaluations, frequency documentation, and severity indicators specific to the condition's rating criteria. A veteran claiming migraines, for example, needs documentation of frequency, duration, and economic impact — not just a diagnosis.

Nexus Letter Guidance

The nexus letter is often the single most important document in a VA claim. It is a medical opinion linking a current condition to military service. VetAssist helps veterans understand what a nexus letter needs to contain, why the "at least as likely as not" standard matters, and what common deficiencies cause letters to be rejected. We do not generate nexus letters — that requires a medical professional. We help veterans have informed conversations with their doctors about what the letter needs to say.

Condition Database

297 VA-recognized conditions drawn from 38 CFR Part 4, each with the rating criteria that determine what percentage the VA assigns. Veterans can look up their conditions, understand the difference between a 30% and 50% rating for their specific diagnosis, and identify what evidence might support a higher evaluation.

The Math

The VA does not add disability ratings together. Instead, it uses the "whole person" theory from 38 CFR 4.25: each rating is applied to the remaining able-bodied percentage, not the original 100%. The logic is that the VA rates based on remaining earning capacity. A person who is already 50% disabled and then sustains a 30% disability has 30% of their remaining capacity reduced, not 30% of their original capacity.

Here is how it works, step by step:

38 CFR 4.25 Combined Rating Formula
════════════════════════════════════════════════════════════

Rule: Sort all ratings highest-first. Start with 100% (whole person).
      For each rating, reduce the remaining capacity by that percentage.

Example: Three service-connected conditions rated 50%, 30%, and 20%.

Step 1 — Start with the whole person:
         Remaining capacity = 100%

Step 2 — Apply highest rating (50%):
         Loss     = 100 × 0.50 = 50.00
         Remaining = 100 - 50.00 = 50.00%
         Combined so far: 50.00% disabled

Step 3 — Apply next rating (30%):
         Loss     = 50.00 × 0.30 = 15.00
         Remaining = 50.00 - 15.00 = 35.00%
         Combined so far: 65.00% disabled

Step 4 — Apply next rating (20%):
         Loss     = 35.00 × 0.20 = 7.00
         Remaining = 35.00 - 7.00 = 28.00%
         Combined so far: 72.00% disabled

Step 5 — Round to nearest 10%:
         72% → 70% combined rating

════════════════════════════════════════════════════════════
50% + 30% + 20% = 70% (not 100%)

If you added those ratings naively, you would get 100%. The VA formula yields 70%. That 30-point gap is the difference between full disability compensation and a partial rating. For a veteran with a family, that gap can represent thousands of dollars per year.

PII Protection

Veterans who use VetAssist enter sensitive information: names, medical conditions, service dates, personal statements. This is not data we take lightly. A claims tool that leaks veteran medical data is worse than no tool at all.

Field-level encryption. Sensitive fields are encrypted at rest individually, not just at the database level. Even if someone gained access to the underlying storage, individual fields containing names, conditions, and service details are encrypted independently. The database sees ciphertext, not cleartext.

PII detection. Free-text fields are scanned automatically for personally identifiable information. Veterans sometimes paste full medical records or include Social Security numbers in text fields where they should not appear. The system flags these before they are stored and prompts the veteran to remove them.

No cloud. All data stays on our hardware. No third-party APIs touch veteran data. The language model runs locally. The database runs locally. There is no point in the pipeline where veteran information leaves machines we physically control.

Session isolation. Each session is independent. We do not build profiles. We do not track veterans across sessions. When a session ends, the working data is gone. The only data that persists is what the veteran explicitly saves, and that data is encrypted as described above.

Crisis Detection

Veterans filing disability claims are a population at elevated risk. The intersection of service-connected trauma, bureaucratic frustration, financial stress, and the emotional weight of documenting your worst experiences creates a context where crisis indicators must be taken seriously.

We built a sentiment analysis layer that monitors user input for crisis indicators in real time. If the system detects language suggesting a veteran may be in distress, it immediately surfaces crisis resources:

• Veterans Crisis Line: 988, then press 1
• Crisis Text Line: Text 838255
• Local VA emergency contacts based on the veteran's indicated region

This feature was a unanimous decision in our internal review — full agreement, no concerns, highest confidence score we had seen at the time. Some features you do not debate. When you are building a tool that asks veterans to describe their service-connected injuries, you have a responsibility to recognize when someone is telling you more than their claim history.

The detection is not a simple keyword match. Context matters. A veteran writing about a past traumatic brain injury for their claim is different from a veteran expressing hopelessness about their situation. The analysis layer considers context, intensity, and temporal framing to minimize false positives while maintaining sensitivity to genuine signals.

The Tech Stack

We are deliberately brief here. The technology is a means, not the point.

No cloud providers. No third-party inference APIs. No vendor lock-in. If a service goes down, we fix it ourselves. If a model needs to be updated, we update it ourselves. The tradeoff is operational complexity; the benefit is that we never have to explain to a veteran why their data was processed on someone else's servers.

What's Next

VetAssist is functional today. It is not finished. Here is what we are building toward:

VSO white-label. Veterans Service Organizations — the DAV, VFW, American Legion, and others — are the frontline of claims assistance. We want to let them customize VetAssist for their veterans: their branding, their workflows, their supplementary resources. A VSO representative helping a veteran in person should be able to pull up VetAssist with their organization's guidance integrated.

Browser extension. The VA.gov claims portal requires manual data entry. A browser extension that auto-fills form fields with calculated data — combined ratings, condition codes, evidence checklists — would eliminate transcription errors and save veterans time on the most tedious part of the process.

More conditions. We currently cover 297 VA-recognized conditions with full rating criteria. The Code of Federal Regulations contains over 800 ratable conditions. Expanding coverage is a matter of careful data entry and validation, not engineering. It is slow work done right.

Buddy statement templates. Lay evidence — statements from fellow service members, family, and friends — can be decisive in claims decisions. But most people have never written a statement for a federal proceeding. Guided templates that walk a buddy through what to include, what language the VA responds to, and how to structure their observations would make lay evidence more effective and more accessible.

What Didn't Work

Honesty about failures is more useful than a highlight reel. Here is what went wrong on the way to what works.

AI for everything, including math. The first version tried to use the language model for the combined rating calculation. This was wrong in a fundamental way. The 38 CFR 4.25 formula is deterministic — given the same inputs, there is exactly one correct output. Language models are probabilistic. They are excellent at natural language tasks like evidence guidance and nexus letter context, where nuance and judgment matter. They are the wrong tool for arithmetic. The calculator is now pure code. The language model handles language. Each tool does what it is good at.

SSL certificate renewal. The reverse proxy handles TLS certificates automatically, renewing them before expiration. In theory. In practice, the certificate authority's validation requests could not reach our server through our firewall rules. The proxy would attempt renewal, fail silently, and the certificate would expire. It took three separate outages before we identified the specific firewall rules that needed adjustment. The fix was two lines in a configuration file. Finding those two lines took weeks.

Memory leak. The backend service would start at reasonable memory usage and slowly climb over the course of days until it consumed everything available. The root cause was the database connection pool: connections were being opened but not properly recycled. Old connections would accumulate, each holding onto memory that was never released. The fix was explicit pool recycling — a configuration change that forces connections to be closed and reopened after a set interval. Simple in retrospect. Invisible until you watch memory graphs for a week.

Encrypting the entire database. Our first approach to PII protection was full-database encryption. Every column, every table, everything encrypted at rest. The performance impact was severe — queries that should take milliseconds took seconds, and the application felt sluggish in a way that would drive veterans away from using it. The right granularity was field-level encryption on sensitive columns only. Names, conditions, service dates, and personal statements are encrypted. Metadata, configuration, and reference data are not. Targeted protection, not blanket overhead.

The best technology is the kind that disappears. Veterans should not need to understand AI to get the benefits they earned.

Cherokee AI Federation · Built on consumer hardware · No cloud · No compromise