Healthcare & life sciences — patient access to pre-approval treatmentsMulti-agent systems & regulatory knowledge graphs

Agents that watch global regulators and keep a graph of treatment access pathways current

A team helping patients reach pre-approval treatments tracked regulatory change across jurisdictions by hand. We built agents that detect, validate and write each change into a knowledge graph of access pathways — delivered in five and a half weeks.

Three kinds of regulatory source — US, UK and EU registries, government websites, and regulatory publications — feed a change detection agent that classifies each update by significance and region, followed by an LLM-as-judge agent that checks the extraction against defined accuracy criteria. Clear updates are written to the graph with provenance and a timestamp; high-stakes or ambiguous updates are reviewed and approved by the regulatory affairs team first. Both flow into one regulatory knowledge graph of jurisdictions, regulatory bodies, access pathways, requirements and history, which answers questions with the source cited and sends alerts by subscription. Headline figures: under 24 hours from source change to graph, above 90% accuracy, under 5 seconds to a cited answer, and above 99% availability.
Agents read the regulators. People approve what matters. Everyone asks the same graph.

Project snapshot

Client
A company connecting patients and physicians with treatments that are not yet approved
Industry
Healthcare & life sciences — patient access to pre-approval treatments
Business function
Regulatory affairs and compliance monitoring
Challenge
Whether a patient can reach an unapproved treatment depends on access pathway rules — expanded access, compassionate use, orphan drug and clinical trial programmes — that differ by jurisdiction and change without notice. The regulatory team tracked registries, government websites and publications by hand, so updates arrived late and answers depended on who had read what.
Solution
A multi-agent system orchestrated with Agno. Ingestion pipelines pull US, UK and EU regulatory sources into a knowledge graph of jurisdictions, regulatory bodies, access pathways and requirements, with provenance. A change detection agent classifies updates by significance and region, an LLM-as-judge agent validates them, and high-stakes or ambiguous changes go to regulatory experts for approval. A GraphRAG agent answers pathway questions with the source regulation cited, through a role-based API.
Result
Regulatory changes reach the graph within 24 hours, with above 90% accuracy, cited answers in under 5 seconds, 99%+ availability and 100+ concurrent queries — delivered as seven epics in five and a half weeks.

Key outcomes

< 24h
From a regulatory change to the knowledge graph
90%+
Accuracy on regulatory pathway data
< 5s
To a cited answer on any access pathway
100+
Concurrent queries through a role-based API

The client

Getting patients to treatments that are not yet approved

A company that connects patients and physicians with treatments still in development. When a patient has run out of approved options, it helps their physician find a lawful route to an investigational treatment — a clinical trial, an expanded access or compassionate use programme, or an orphan drug scheme.

Which route is open depends on where the patient is, and on rules that regulators revise continuously.

The challenge

Compliance tracking that depended on someone reading

Every regulatory update was found, read and recorded by hand. Registries, government websites and regulatory publications each change on their own schedule, in their own format.

The surface area

3
jurisdictions — US, UK and EU
4
access pathway types to track

Knowledge sat with whoever had last read the source. An answer to “is this pathway still open here?” depended on who was asked, and how recently they had checked.

Constraints

What we had to design around

  • StakesA wrong answer about an access pathway can send a physician down a route that is closed. Speed mattered, but never more than being right — and being able to show where the answer came from.
  • SourcesSome regulators publish structured registries with APIs. Others publish web pages and guidance documents. The pipeline had to read both and normalise them into one model across jurisdictions.
  • ConsistencyThe same question asked twice had to get the same answer. A system that paraphrases a rule differently each time cannot be relied on for compliance.
  • Human authorityRegulatory affairs owns the final word. Changes in high-stakes jurisdictions, and anything ambiguous, had to reach an expert before reaching the graph.
  • Existing estateThe client already ran a graph database, cloud infrastructure and monitoring. The system had to plug into them, meet ISO 27001 and SOC 2 expectations, and leave room for further jurisdictions.

Our approach

Model the rules first, then let agents keep them current

A search engine over regulatory documents finds text. It does not tell you which pathway is open. That needs the relationships: which body governs which pathway, in which jurisdiction, with which requirements, as of which date.

So we built the ontology before any agent. A knowledge graph holds the structure. Agents keep it current — one to spot changes, one to judge them, one to answer questions — and people approve the changes that carry real risk.

The solution

Five layers, from a regulator’s page to a cited answer

  1. Ingestion — Cloud-native ETL pulls from US, UK and EU registries, government websites and regulatory publications. Extraction modules handle APIs and semi-structured pages; transformation modules normalise formats across jurisdictions, with completeness and accuracy checks on the way in.
  2. Regulatory knowledge graph — An ontology of regulatory bodies, jurisdictions, access pathways — clinical trials, expanded access, compassionate use, orphan drug programmes — and their requirements, plus a historical timeline. Every fact keeps its provenance.
  3. Agents — Orchestrated with Agno. A change detection agent watches the sources and classifies updates by significance and region. An LLM-as-judge agent validates each one. A GraphRAG agent retrieves the relevant subgraph and answers with the source cited. Contextual memory recalls earlier decisions and how a framework has evolved.
  4. Human review — Flagged and high-stakes changes land in a review workflow. Regulatory experts approve or correct them before they are written to the graph.
  5. API and alerts — A role-based API with query and suggest-update endpoints serves the regulatory affairs team, site managers and client applications. Alerts go to stakeholders by subscription, and every detected change sits in an audit trail.
The regulatory monitoring system. At the top, data sources: US, UK and EU registries, an APAC registry shown dashed as an extension, government websites and regulatory publications, all monitored by the regulatory change detection agent. Inside the Agno agent framework, a GraphRAG query agent and the change detection agent sit above an LLM judge validation agent; the change detection agent passes changes to the judge for validation. On the left, the knowledge graph contains jurisdictions, which contain regulatory bodies, which govern access pathways; a historical timeline tracks changes to access pathways, and each pathway has requirements. The GraphRAG agent queries every part of the graph and the judge validates updates to it. On the right, a user interface and API layer with an alert system, a query endpoint and a suggest-update endpoint serves site managers, external clients and the regulatory affairs team, who approve suggested updates. Success metrics: latency under 5 seconds, availability above 99%, accuracy above 90% and data freshness under 24 hours.
The graph on the left is the model of the rules. The agents in the middle keep it current; the API on the right is how everyone else reads it.
Seven epics over five and a half weeks. Epic one, half a week: ontology and knowledge graph schema for regulatory bodies, jurisdictions, access pathways and requirements, validated with domain experts. Epic two, one week: source integration with ETL from registry APIs and websites, normalised across jurisdictions with quality checks. Epic three, one and a half weeks: GraphRAG queries with subgraph retrieval, cited sources, consistent answers and responses under five seconds. Epic four, one week: change detection classified by significance and region, with subscribed alerts and an audit trail. Epic five, half a week: LLM-as-judge validation with an accuracy dashboard, drift detection and human review of flagged items. Epic six, half a week: a secure API with query and suggest-update endpoints, role-based access for regulatory affairs and site managers, handling 100+ concurrent queries. Epic seven, half a week: containerised deployment with infrastructure as code, existing monitoring, and performance and security testing.
The ontology lands first because every agent after it reads and writes the same graph.
5.5 weeksfrom ontology to a deployed, monitored system, across seven epics

Responsible by design

Four rules for agents near patient access decisions

  • Cite

    No answer without its source.

    Every response points to the regulation it came from, so a reviewer can check it in one click rather than trusting the model.

  • Judge

    Every extracted change is checked before it is stored.

    An LLM judge tests each update against defined accuracy criteria. An accuracy dashboard and drift detection show when that standard slips.

  • Escalate

    High-stakes and ambiguous changes go to a person.

    Regulatory experts approve anything the agents are unsure of, and anything in a jurisdiction where an error would be costly.

  • Trace

    Every change has a history.

    Provenance on each fact, versioning on each document and an audit trail on each detected change. The graph can say what a rule was last month, not just what it is today.

The third rule is the one the others serve. The agents do the reading and the checking. The regulatory team keeps the authority, and spends it on the changes that need it.

Results

What changed for the regulatory team

  • Regulatory changes in the graph within 24 hours — from a registry update or a government notice to a validated, queryable fact, without anyone reading the source by hand.
  • Above 90% accuracy on regulatory pathway data, with an LLM judge on every change and an expert on the ones that matter most.
  • Cited answers in under 5 seconds, deterministic, so the same question gets the same answer next week.
  • 99%+ availability and 100+ concurrent queries through a role-based API that the regulatory affairs team, site managers and client applications all use.
  • Manual compliance tracking replaced — the regulatory team reviews flagged changes instead of searching for them.

The team stopped searching for changes and started deciding on them.

Beyond the numbers

What else changes

The rules become something you can ask, not something you have to remember. “Which pathways are open for this treatment in this jurisdiction, and what changed since last quarter?” is one query against the graph, answered with the sources attached.

And because the schema was built to extend, a new jurisdiction is a new set of sources feeding the same graph and the same agents, not a new system.

If your compliance team spends its week reading regulators

The first question is which sources change most and which changes matter. That is answerable from a month of your update history, before anything gets built.