Agents

Agentic AI for finished reports.

AI agents inspect the data, write the analysis code and assemble the report.

The PEP Health reporting agent takes a plain-English request and turns it into a self-contained report. It reads the request, inspects the data, writes and runs analysis code, checks the result, and keeps iterating until the report is ready to download, email or present.

Talk about a use case

Agent loop

What the report agent does.

Starts from a plain-English request

The user asks for a report in a chat window. There is no dashboard to configure, and no spreadsheet to prepare. The agent works out which data and analysis the request needs.

Built for ad hoc questions over a large data warehouse, from people who do not write code.

Writes and runs its own analysis

The agent inspects the data, writes Python analysis code and runs it in a controlled sandbox. It checks what came back, then repeats while more analysis is needed.

The loop is deliberate: decide what to analyse, run the code, check the result, then continue or stop.

Assembles a finished report

The output is a polished report in PEP Health's visual identity, with embedded charts and no extra software needed to open it or share it.

The report is the product, not an intermediate chart for someone else to interpret.

State of play

What is built, and what is one test away.

The earlier version is built and proven end to end on one customer's data. The newer version takes the same design across PEP Health's full customer base, with reasoning handled by Llama 3.3, self-hosted on PEP Health's own AWS infrastructure.

That generalised version is built and one live test away. It has been checked against a scripted test response, but has not yet run end to end against a live model server.

The 60 per cent figure remains a design target for reducing manual report-writing effort, not a measured result. It stays framed that way here.

Built, one live test away 60.6 million comments

The volume the reporting agent is built for: 4.6 million facilities and practitioners, with reasoning on Llama 3.3 inside PEP Health's own AWS.

The agent loop. It receives the request, decides whether more analysis is needed, writes and runs code against real data, checks what came back and repeats. It only stops once it is confident the analysis is complete.

Method and stack

Infrastructure, guardrails and status.

Self-hosted reasoning

The reasoning engine is Llama 3.3 on PEP Health's own AWS infrastructure. Patient data, and the questions asked about it, do not leave PEP Health's environment.

Contained execution

The agent writes and runs Python analysis code in a controlled sandbox. That is what lets it answer a new request, rather than fill a fixed dashboard template.

Conservative status

The generalised version is described as built and one live test away: checked against a scripted response, but not yet verified against the live model server.

Llama 3.3, self-hosted
vLLM
AWS
Contained code execution
Data stays inside the client's infrastructure
Plain-English report requests
Finished report output

Llama 3.3 is self-hosted on the client's own AWS via vLLM, so the reasoning runs inside their infrastructure. Neither the data nor the questions about it leave. Agent-written code executes in a contained environment, then the agent assembles the report in the client's visual identity.

Get in touch

Reach out and tell us what you're working on. Every message gets a personal reply within two working days.

Prefer email? Write straight to hello@naviins.com.