Skip to content
Back to work

AI & Automation · Healthcare

AI Agents Platform

Seven agents that turn one upload into a recorded, print-ready batch of personalized posters, with a reviewer approving anything that carries commercial risk.

Role
AI Product Owner & Solution Architect
Built on
Google Cloud, Python, React
Industry
Healthcare
Project type
AI & Automation
Primary service
AI & Automation
Scope
Product ownership, solution architecture, agent design, validation and approval workflow, delivery.

Repository not shown as per company policy

  • 0

    agents, each with one clear job and its own screen

  • 0

    tracked codes generated on every poster

  • 0

    upload starts an entire batch, in the background

  • 0

    person approves every incentive before it goes out

01Challenge

Assembly by hand does not scale.

Nothing about the manual process was careless. It simply could not keep up once every recipient needed their own version of the same output.

  • Slow to assemble

    Up to a dozen sections had to fit a fixed layout, to the pixel, for every single recipient.

  • Personal at scale

    A unique tracked code per recipient and location simply cannot be placed by hand across a print run.

  • Hard to trace

    Nothing reliably recorded which output, or which incentive, went to whom.

  • Risky incentives

    One wrong incentive code awards the wrong points to the wrong person. That is a commercial error, not a cosmetic one.

02Approach

Fast where it is safe. Careful where it counts.

  • Automate the mechanical middle

    Layout, code generation, assembly and packaging are repetitive and rule-bound, which makes them the right work to hand to a system.

  • Put the person at the risk point

    The two places a mistake is expensive are an incentive that awards points and an output that prints without a working code. Both are gated.

  • Configuration as data

    Layouts and validation rules live in storage, not in a build, so a rule change reaches users in minutes rather than at the next release.

03Outcome

Seven agents, one personalized output per recipient.

A poster is a puzzle of up to a dozen sections, and each recipient needs their own set of tracked codes. Instead of designers assembling each one by hand, creators pick a layout, upload the content, and the platform does the rest.

  • Set it up

    Layouts and input rules are stored as data, not design files, so the business can change what a valid output looks like without waiting for a release.

  • Build it

    Tracked codes and print-ready PDFs are generated in bulk, in the background, while the creator does something else.

  • Check it

    Incentives need a person’s approval, and every code, poster and layout version is recorded, so nothing goes out untraced.

What changed

  • A batch that previously required a designer per recipient now starts from one upload and returns a print-ready package.
  • Every code, poster and layout version is recorded, so any run can be reconstructed after the fact.
  • Incentives cannot reach the partner without a named reviewer approving them first.

How it works

Follow one batch through the system.

A creator works in the agent dashboard, uploads a recipient list and content sections, validation rules check the input, a job service queues the run, and the agent pipeline generates codes, assembles PDFs and records metadata, with a human approval gate before anything reaches the partner.

batch startsstarts a runone ZIP, ready for printCreatorPicks a layout, uploadscontent, approves perksAgent dashboardEvery agent is a tile,one click awayJob serviceAnswers instantly with ajob number, runs in backgroundCloud StoragePhysician list, postersections, finished PDFsTemplates & rulesLayouts and input rulesedited on a screen, not in codeReviewerApproves or rejectsthe perks batchPerks partnerReceives only codes aperson has approvedAgent pipelineOne shared blueprint — adding an eighth agent is quick1 QR GeneratorA tracked code per physician and location2 Wallboard AssemblerPrint-sharp PDFs, built in small batches3 Metadata CollectorWhat was made, when, and how to find itPerk approval gateA reviewer signs off before anything is sent

One upload in. A finished batch out.

01 / 07

Illustrative walkthrough of a typical run. Sample identifiers are representative, not client data.

Inside the system

What it is made of, and what keeps it safe.

Seven agents, one job each.

Each agent has its own screen and does one thing well. They share one blueprint, which is what makes an eighth agent a week of work rather than a project.

  • 1

    Dashboard

    The front door. Every agent is a tile, one click away.

    Why it mattersA new capability is a new tile, not a new product.

  • 2

    Template Management

    Stores each layout and previews it before anything is built.

    Why it mattersLayouts are data, not design files, so they are easy to reuse.

  • 3

    Validation

    Lets the business define what a valid input file actually looks like.

    Why it mattersRule changes go live in minutes, without a release.

  • 4

    QR Generator

    Makes tracked codes one at a time, or thousands from a spreadsheet.

    Why it mattersRunning the same file twice never creates duplicates.

  • 5

    Assembler

    Builds the output PDFs: one to check a design, or a whole batch in the background.

    Why it mattersPrint-sharp output, and every item is recorded.

  • 6

    Incentive Activation

    Validates incentive codes and sends only approved ones to the partner.

    Why it mattersA person approves before any points are awarded.

  • 7

    Metadata Collector

    The registry: what was made, when, and how to retrieve it again.

    Why it mattersSupport and audit can find anything fast.

Built with

What it runs on.

Cloud
Cloud RunCloud StoragePub/SubFirestoreCloud Logging
Services
PythonFastAPIBackground jobsPDF generationCode tracking service
Interface
ReactNode.jsREST APIs
Security and delivery
Identity-Aware ProxyIAMCI/CD
Services appliedAI & AutomationTechnology SolutionsDevOps & MLOps

Roadmap

Where the platform goes next.

The learning loop

  • Understand each recipient from what they read, open and scan, plus specialty and location
  • Suggest the best content per output, with creators reviewing and rating every suggestion
  • Feed scans and clicks back in, so each round is better targeted than the last

Platform hardening

  • Sturdier large runs that recover on their own and retry only what failed
  • Validation rules enforced behind the scenes as well as on the screen
  • Explicit hand-offs between agents, so they ask each other rather than reading each other’s data

Creator experience

  • A layout designer, so creators build a new layout rather than only previewing one
  • Variant testing, using code tracking to compare versions
  • A reach view tying scans and clicks back to the campaign that produced them

Pieces in. A personalized, trackable output out. People in control of what ships.

More work

Related projects.

View all work
AI & AutomationHealthcare

Autonomous AI Log Monitoring & Observability Platform

For a healthcare media and clinician engagement platform, a five-agent system that reads a production error, writes the fix and opens a reviewed pull request, with an engineer still deciding what ships.

AI & AutomationHealthcare

Drug Competitor Identification

A brand-intelligence tool that asks a language model who a drug competes with, then checks the answer against regulatory reference data before anyone is asked to trust it.

AI & AutomationTechnology & SaaS

Customer Churn Prediction Model

For a subscription software business, a churn model that flags drifting accounts early enough for the customer success team to act, with the reasons behind every score.

Discuss a similar project.

If something here is close to what you need, tell us about your situation and we will walk you through how we would approach it.

Discuss a Similar Projectcontact@genclover.com

Intelligence → Innovation → Automation → Growth