Our Research

Research toward HIPAA-native intelligence.

Theo is an AI research lab. We build the orchestration, safety, and compliance layers that let healthcare and insurance teams put models to work on real patient data, from the routing engine that runs on every call to a PHI-native model platform of our own. It stays auditable by design and compliant from the very first call.

300+: Models orchestrated
1: API call, full pipeline
5: Stage orchestration
100%: HIPAA-native

What we do

We build the layer that makes models safe to use on regulated data.

Healthcare and insurance teams don't need another model. They need everything around it: an engine that routes each request to the right capability, a model layer that keeps protected data out of prompts and logs, and a compliance posture that travels with every product. That's the research we do.

Where we focus

Six problems we keep returning to.

Orchestration & routing

Classifying intent and routing each prompt across 300+ models in a single call, with automatic failover and a five-stage pipeline.

A PHI-safe model layer

Redaction, access controls, and tamper-evident audit trails designed directly into the model layer, so PHI never leaks into prompts or logs.

Compliance as a framework

Baking SOC 2, HIPAA, and HITRUST into the framework itself, so every product inherits our posture instead of earning it from scratch.

Agentic clinical workflows

A think-act-observe agent loop with human-in-the-loop approvals and an evidence ledger that cites back to the source record.

Memory & context

A memory protocol that resolves intent and assembles the right context before the model sees a request, with provenance on every fragment.

Production reliability

Isolated environments, blue-green deploys, circuit breakers, and end-to-end observability so a research idea survives contact with real traffic.

How we do it

One call runs a five-stage pipeline.

Every request through the Theo engine runs the same five-stage orchestration pipeline (classify, load skills, route, run an agent loop, and respond) across hundreds of models behind a single key.

The pipeline

Classify

Load Skills

Route

Agent Loop

Respond

Classify

Detect the intent and task type behind the prompt.

Load Skills

Inject domain expertise, tools, and prompt extensions.

Route

Select the optimal engine, with automatic failover.

Agent Loop

Think, act, and observe with tools until the task is done.

Respond

Return content with citations, artifacts, and usage.

How routing decides

Intent over instruction

The mode a caller asks for is a preference, not a lock. Every prompt is classified for intent so an image request returns an image, whatever mode it arrived on.

Confidence-gated promotion

A prompt is only re-routed when the classifier clears a confidence floor. Below it, we honor the caller's mode and still surface the signal for observability.

Deterministic locks

Sensitive agentic flows, such as quoting, claims, and extraction, are pinned. They never accept a silent, intent-based override; routing into them is caller-driven only.

Automatic failover

Circuit breakers watch the engine stack and reroute on provider failure, so a single upstream outage never becomes a failed completion.

The production environment

Research only counts once it survives real traffic.

Behind the engine is the operational work that turns an idea into a dependable service: isolation, reversibility, durability, and an audit trail you can stand behind.

Isolated environments

Production, QA, and dev run fully apart, with separate data, caches, and credentials, so nothing under test can ever reach a patient record.

Blue-green deploys

Releases roll out behind a pointer with an instant rollback, so shipping is reversible and a bad build never has to become an outage.

Durable background workers

Long-running work, such as video, research, documents, and memory synthesis, runs on durable job queues that survive restarts and scale on their own.

Idempotent, guarded APIs

One middleware pipeline gives every endpoint auth, rate limits, idempotency keys, timeouts, and a fire-and-forget audit trail.

Observability & uptime

Structured logs, error tracking, and a heartbeat feed one source of truth, so platform status in the dashboard can never drift from reality.

Immutable audit

High-value actions are written to a tamper-evident trail, turning compliance from a quarterly scramble into a property of the system.

Published research

Our model platform, in full.

The flagship of our research: a standalone, HIPAA-grade model platform. The complete paper covers architecture, build pipeline, PHI guardrails, runtime, and benchmarks.

Technical Report18 min read

Theo Arca 5.1

Theo Arca is a deployable, self-contained AI stack for regulated industries. It serves the Arca 5.1 model family (Velox, Magnus, Edge, and Code) behind an OpenAI-compatible API, wrapped in a native privacy guardrail, an agentic tool runtime, an immutable audit ledger, and a fleet control plane that manages nodes on customer infrastructure, including air-gapped environments. Our central claim is that trustworthy AI in healthcare and insurance is a systems property, not a model property, so the guarantees are built around the model and verified continuously by a promotion-gating benchmark harness of seven suites. No model version reaches production without a passing scorecard.

Model platformBenchmark-gatedPHI-nativeAgentic

Read the paper

From the lab

Ideas and notes shaping how we build.

EngineeringVol. 1

Routing across 300+ models in a single call.

ResearchVol. 1

Designing PHI-safe audit trails into the model layer.

NotesVol. 1

Why we baked HITRUST into the framework itself.

FieldVol. 1

The agent loop: think, act, observe, until it's done.

Build on Theo

Put our research to work.

Start directly on hitheo.ai: sign up, get an API key, and call the model in minutes.

Start on hitheo.ai Read the docs