← All writing

If It Can Hold Psychiatric Records, It Can Hold Your Memoir

Clinical patient files and family histories demand the same architectural rigor. The system has to be a vault or it's a liability.

Clinical patient files and family histories demand the same architectural rigor. The system has to be a vault or it’s a liability.

The Blindfold Comes First

Before a single token of user data reaches an LLM inside PureTome, it gets physically redacted.

I call it the Blindfold Protocol. I built it because I assume every environment is compromised on day one. After architecting a fully HIPAA-compliant psychiatric platform this quarter, I’m more convinced than ever that most AI products on the market today are one breach away from extinction.

Right now, thousands of companies are bolting generic AI wrappers onto their products and funneling sensitive user data into public models. They ship fast. They treat privacy as a compliance checklist instead of a foundational engineering primitive. When the breach comes (and it always comes), the post-mortem reads the same: we didn’t think anyone would target us yet.

That assumption is the bug.

I’m officially launching PureTome Labs, the consulting and C2C engineering arm operating under the PureTome parent. To mark the launch, I’m walking through the architecture of our first deployment: QuietHorizons Telepsychiatry, a HIPAA-compliant clinical platform for psychiatric patient management.

Why Psychiatric Data Is the Hardest Test

Psychiatric records are the highest-stakes patient data in healthcare. A leaked diagnosis ends careers, custody cases, security clearances, and lives. HIPAA fines run up to $1.9M per violation category per year, and a single LLM that hallucinates a patient detail into the wrong session can put a clinic’s license at risk overnight.

You cannot bolt this together with off-the-shelf SaaS and a prayer.

I engineered QuietHorizons from the ground up to make breach almost mechanically impossible and to make every clinical AI interaction auditable, stateless, and verifiable.

The Production Stack, and What Each Piece Defends Against

  • [Foundation] Google Workspace under BAA. The entire ecosystem operates inside a signed Business Associate Agreement, so compliance is enforced at the infrastructure root rather than patched on at the application layer.
  • [Backend] NestJS microservices on Cloud Run with Drizzle ORM and Cloud SQL Postgres. Cloud Run executes stateless. No persistent process memory means no session artifacts for an attacker to scrape between requests. Drizzle gives me typed, auditable query construction so a developer can’t accidentally write a leaky raw SQL string under deadline pressure.
  • [Frontend] Angular with Spartan UI. Strict typing end-to-end and a component model that forces explicit data flow. Clinicians get a precise, modular interface; I get a codebase where unauthorized data access has nowhere to hide.
  • [Ingestion] Google Speech-to-Text over secure WebSockets. Real-time encrypted clinical dictation streams directly into the patient’s isolated dossier. The encrypted channel is the only path in.
  • [The Ambient Scribe] Vertex AI / Gemini, stateless inference. This is the key architectural decision. The AI scribe writes nothing to a draft-state database. Inference happens in memory, the result is committed to the immutable medical record after provider review, and the working state evaporates. No draft tables, no shadow copies, no orphaned PHI sitting in a cache waiting to be subpoenaed or stolen.

I just opened several repositories on my GitHub publicly documenting some of these integration patterns, DevOps pipelines, and vector infrastructure choices. The build is the proof.

The Thread Connecting Telepsych and Memoir

The architectural thread between a psychiatric clinical portal and an AI memoir engine is identical: the user has to trust the vault.

A patient handing over their psychiatric history and a 72-year-old grandmother handing over the story of her marriage are doing the exact same thing. They’re trusting that their most private narrative won’t escape, won’t be sold, won’t be used to train someone else’s product. The technical countermeasures are nearly identical. The standard of care should be too.

Most AI companies haven’t figured that out yet. PureTome Labs exists for the ones that need to.

What PureTome Labs Does

I solve complex, high-stakes engineering bottlenecks for early-stage teams and established operators. 0-to-1 builds. Monolithic technical debt untangled. Advanced RAG implementations, full-stack development with robust CI/CD, cost-containment algorithms for LLM context windows, HIPAA-grade architecture for regulated industries.

I do not farm out tickets. I execute the build.

If your team is staring down a critical architectural decision, or you need a lead engineer who builds like a founder to execute a high-velocity bridge contract, my calendar is open.