Build a regulated AI inquiry workflow

Project 1 of the Agentic AI Buildcamp for Engineers & PMs. Four 10-minute sessions, each with a lab you run and a test that proves the property.

Session	What you build	Labs
1. Design a Regulated AI Workflow	The outcome, the metric and the must nevers, then a routing table with safety first	Route inquiries
2. Build Grounded Agent Replies	Pinned agents that select approved block ids, and a check on every sentence	Ground replies, Pin the model
3. Secure Agent Tool Calls with Cedar	A Cedar check on every tool call, and an audit trail nobody can edit	Guard tool calls, Hash chain
4. Prove Every Release with an Eval Gate	Twenty eval cases, a gate that fails on one missed adverse event, and mutants	Gate releases

The job: answer doctors' product questions for a pharmaceutical company in minutes instead of days, without ever sending a sentence that Compliance cannot trace to current, approved content, and without ever missing an adverse event.

Quillon Biopharma is synthetic, and so is every doctor, product and content block. The architecture is the real one: agents on the Claude Agent SDK, all on claude-haiku-5-5, and AWS services running locally on Floci. The same boto3 code runs against AWS.

Before	After, in this repository
Each question routed by hand between Medical and Legal, taking days	A fixed graph routes each inquiry in one pass; safe, cleared topics are answered at once
Nobody can prove a reply was approved content	Every sentence carries its block, version and MLR approval, checked again at the moment of sending
An adverse event can sit in a general inbox	Any possible adverse event goes to pharmacovigilance before any other rule; the eval gate fails the release if one is missed
Permissions live in application code	Every permission is a Cedar policy in Verified Permissions, including the agents' own tool calls
Audit is a log someone could edit	A hash-chained ledger, archived under S3 Object Lock, where any edit is found

Run it

Needs Docker, uv and Node 22.

make up       # Floci (local AWS) and every resource: DynamoDB, S3, SQS, EventBridge,
              # Step Functions, Cognito, Verified Permissions, SSM, CloudWatch Logs
make serve    # the console on http://localhost:8765
make check    # the gate: tests, labs, evals, mutants, prose, web build, diagrams

Model calls replay from recordings/ by default, so all of the above is free and needs no API key. make record calls Haiku 5.5 for every inquiry, about two cents in total, under the spend cap in policies/autonomy.toml.

Read it

The four sessions teach the build. These pages are the reading around it, and a PM starts with the case study and the delivery plan.

• Case study: the platform, agentic AI in a regulated setting, the delivery plan, and why this slice.
• Architecture walkthrough: each part, and the three flows Compliance cares about, traced through the code.
• Product walkthrough: the console screen by screen, with the moments it refuses to answer or act.
• Delivery plan: from first conversations to a first release in Germany.
• Architecture decisions: each one names the failure it prevents and the test that enforces it.
• Labs: one property per lab, a test that proves it, and a change that breaks it.

Layout

Path	What it holds
src/hcpagents/	The graph, agents, grounding, guards, Cedar client, ledger, API and AWS provisioning
policies/	Cedar policies and the autonomy settings Compliance controls
world/	The synthetic company: products, doctors, people, content blocks, inquiries
evals/	The release gate thresholds and the last report
recordings/	Recorded Haiku 5.5 calls, keyed on agent, prompt version, model and input
web/	The Astro console
.claude/	Rules, a compliance reviewer subagent, skills and a model hook for working on this repository with Claude Code