Loading...
Client: type2.wiki — a free patient resource for people living with Type 2 Diabetes
Engagement: End-to-end AI/ML capability — design, build, and production operation on AWS
People with Type 2 Diabetes are routinely denied coverage for the medications and devices that keep them well. They have the right to appeal — but the process differs across all 51 US jurisdictions, winning requires citing the right clinical evidence in the right language, and the paperwork lands exactly when a patient is least equipped to handle it. Most people simply give up.
The type2.wiki community knew what a winning appeal looked like. The hard part was delivering one to any patient, on demand, safely — without hallucinating clinical claims, without holding onto sensitive health data, and without an unbounded AI bill ending a free tool overnight.
Jacobian Engineering designed and operates a generative-AI platform that turns a denial into a polished, state-correct, evidence-backed appeal letter in minutes — free, no login.
The design treats the model as a writer wrapped in deterministic guardrails. Two tightly-scoped Claude Haiku 4.5 calls (extract, then generate) sit on top of a hand-curated medical-evidence corpus selected by auditable rules — retrieval-augmented generation reduced to its most defensible form. The letter streams to the browser token-by-token over Server-Sent Events, so patients watch it being written.
The decision Jacobian is proudest of is privacy: patient PII is tokenized in the browser and never reaches the model or any server in the clear. There is no database to breach — the model works on redacted placeholders, and the browser re-inserts the real values locally. Privacy is enforced by the data flow, not by policy.
It is 100% serverless, defined as AWS CDK infrastructure-as-code: CloudFront + WAF at the edge, a streaming Lambda for the pipeline, Secrets Manager for every credential, and a security plane (GuardDuty, Inspector, Security Hub) feeding one Slack alerting surface. Cost is engineered in from day one — every letter’s token spend is measured and dashboarded, a composite alarm guards a hard daily ceiling, and a kill switch degrades the tool gracefully to a fill-in-the-blank fallback rather than breaking.
A vulnerable community gets an expert-grade appeal letter on demand, with a privacy posture they can actually trust and economics that keep the tool free and alive. Behind it: spec-driven delivery, an AI eval harness gating quality, and 485+ automated tests.
It’s how Jacobian builds production AI — measurable quality, measurable cost, security by construction — for workloads where “it mostly works” isn’t good enough.
Read the full engineering story → The complete type2.wiki case study on TrustEdge
Challenge: Turn insurance denials into polished, state-correct, evidence-backed appeal letters in minutes — free, privacy-preserving, with hard cost ceilings. Solution: A 100% serverless AWS CDK platform with browser-side PII tokenization, Claude Haiku 4.5 with deterministic guardrails, an AI eval harness, and 485+ automated tests.