AI agent builder · Seattle, WA · open to insurance AI roles

Jared Fisher

I direct teams of AI agents to build real things: claims AI, developer tools and games.

ClaimsRAGclaims AI Harnessdev tools JUMP CUTgame
Pick a project to trace the work.ClaimsRAG: cited, refusal-first answers for auto claims adjusters.Agentic App Harness: spec-first framework with six deployed demos.JUMP CUT: a stunt-driving game made with AI agents in the Nov 11-15 jam window.
17years in claims
11as senior appraiser
34golden eval questions
38poisoned docs tested

Featured project

ClaimsRAG, live replay

An assistant for auto claims adjusters. Pick a question and watch what happens behind the answer. This is a scripted replay of the real behavior, on synthetic claims. It does not call a server or a model.

Every document here is synthetic. The assistant advises; the adjuster decides.

Question
  1. PlanBreak the question into searches
  2. Hybrid searchMeaning (vector) + keywords, fused
  3. RerankA second model re-scores the top hits
  4. ComputeTested code does the arithmetic
  5. CheckCite, verify, or hold back

Press Run to replay.

Measured, with dates

Does hybrid search actually help?

34 claims questions, scored on 2026-10-21 by a local judge model. Flip between plain vector search and the full pipeline.

Context precision
The documents it pulled are relevant
0.825
Context recall
It didn't miss a relevant document
0.983
Faithfulness
The answer sticks to what was retrieved
0.839
Reference-fact coverage
The answer covers what the reference requires
0.840

Faithfulness and coverage were scored only for the hybrid pipeline. Faithfulness moves a few points between runs, so read it as a range. Case counts are small and each result is a single run. This is a portfolio proof of concept.

Security

Can a poisoned document hijack it?

The adversarial suite plants hostile and conflicting documents. Slide to turn the defenses on.

87.5%

of injection attempts succeed

Measured on the 16 known attacks. On 12 attacks written after reading the defenses, success fell from 58% to 25%. It did not reach zero, and the README lists what still gets through.

Skills

What I bring, and how I work

I direct, review, debug and ship software with AI coding agents, primarily Claude Code. I have no formal engineering background and I don't claim unaided proficiency in any language below.

Public work

Things you can click and try

Story

From the mailroom to RAG

  1. Started in the Liberty Mutual mailroom and learned claims systems from the bottom.

  2. Claims support and total loss handling, dealing with policyholders, attorneys and shops.

  3. Became an auto appraiser, handling supplements, total losses, fraud indicators, subrogation and shop disputes. I did the job for eleven years.

  4. Start teaching myself to build with AI. First goal: make my own job more accurate. Then: make it work for the whole team.

  5. A 48-feature RAG appraisal assistant demoed to VP and Director level. (A pilot and demo, not deployed team-wide.)

  6. Entered the Epic AI MegaJam solo with JUMP CUT, a stunt-driving game made in Unreal Engine 5.9 with AI agents between Nov 11 and Nov 15.

  7. ClaimsRAG and the Agentic App Harness, both public. Looking for AI product, solutions and implementation roles in insurance.

↑↓ move · Enter go · Esc close