Skip to main content

AP: Accounts Payable AI Workflow with AILANG Parse and a sprinkling of AI Protocols

· 6 min read
Mark Edmondson
Mark Edmondson
Founder

I entered a Google competition on DevPost but it turned out as a Google Developer Expert I wasn’t eligible, but its ok as now I have this cool demo app which is the next step in my journey of AI engineering.

AP Showcase running a four-agent pipeline: extracted invoice fields on the left, an MCP App vendor knowledge graph and validator citations on the right

It combines a lot of what I think are good investments in an AI engineering journey:

  • AILANG created the document parsing library ailang-parse for 100% deterministic e.g. can be trusted more than AI/OCR PDF parsing
  • The parsed document is then fed into the AI agent framework from Google - ADK
  • ADK has been configured to use the AI protocols I think will help future proof AI applications: A2A, MCP, AG-UI on the back-end and A2UI and MCP Apps on the front-end - these are what I will talk about at WebSummerCamp next month in Croatia
  • I could boot everything up fast I used the rapidly evolving open-source template I’m maintaining and using in all current client engagements, the AI Protocol Platform (GitHub)
  • It was all deployed on Google Cloud Platform’s Cloud Run, Agent Engine, Firebase etc.

I also had to create lots of marketing assets for the contest that I always mean to do when I make demos but never get around to (part of why I made this Substack to encourage it more)

So you, dear reader, get to be the first to see my practical demo on AI document processing, AP Showcase:

AP Showcase landing page reading 'Invoices in. Audit-grade postings out.' with the intake, extract, validate and post pipeline

The dev URL is now running here, with the app and documentation. I may swap it out for a more official domain if it develops further.

Its hard to get people to log-in and try stuff though, so a 2min video is much easier to grok what its trying to do:

2 mins seems very short when trying to show all your application features, but very long when trying to do an error free voice-over…

The AP Showcase Pitch​

I think the competition format is a good structure to present it so repeat it below:

Problem to solve​

Accounts-Payable teams still process invoices manually because the available tools force a trade-off: rule engines are deterministic but brittle; LLM extractors are flexible but hallucinate vendor data and skip audit trails. Finance can’t accept either failure mode. The result is a £50k–£500k/yr cost in any mid-market AP function — a handful of FTEs typing data into an ERP and reconciling exceptions.

Our solution​

A four-agent AP pipeline on Google ADK that turns an incoming invoice into a trustworthy posting decision with a complete audit trail.

  • The orchestrator (Gemini Pro) delegates to three specialists (Gemini Flash):

    • invoice-extractor uses ailang-parse for 100% fidelity deterministic field extraction (no LLM tokens) and then creates structured output
    • ap-validator grounds every claim against the vendor master via Vertex AI Search
    • ap-poster writes the journal entry or escalates with citations.

Each step is observable live:

  • users see an AG-UI pipeline visualiser
  • an A2UI invoice review card in the workspace pane
  • an embedded MCP App vendor knowledge graph showing exactly which records were cited. The whole system is four declarative SKILL.md files plus ~80 lines of wiring — no orchestration loop, no prompt-chaining glue.

Technologies used​

An architecture diagram shows how all the below fits together

Architecture diagram: a Next.js AG-UI frontend and FastAPI ADK runner feeding an ap-pipeline SequentialAgent that delegates to invoice-extractor, ap-validator and ap-poster

Also available on the website at https://gde-ap-agent-blqtqfexwa-ew.a.run.app/tech

  • Google ADK — agent orchestration, SequentialAgent pipeline, sub-agent delegation, sessions
  • Gemini Pro / Flash via Vertex AI — orchestrator + specialist models
  • Vertex AI Agent Engine (Reasoning Engine) — managed ADK session service + Memory Bank for cross-session recall; chat history and user memory persist across deploys
  • Vertex AI Search — grounded validation against vendor master, open POs, approval policy
  • Google Cloud Run — backend, frontend, and MCP Apps sandbox services (europe-west1)
  • Firebase Auth + Firestore — auth, skill registry, session mirror
  • Cloud Build — branch-based CI/CD (dev → test → prod promotion)
  • AG-UI / A2UI / MCP / MCP Apps / A2A — the protocol stack (CopilotKit-backed AG-UI client, A2UI v0.9 surfaces, MCP tool registry, MCP Apps sandboxed iframes, A2A discovery card)
  • ailang-parse — deterministic Office/.docx/.eml/.tex extraction (<1s, no LLM cost) - handoff to AI for PDF/image parsing
  • AILANG - A new programming language made exclusively for AI Coders - used by ailang-parse
  • Open-source template: sunholo-data/ai-protocol-platform (we built and maintain it)

And outside the competition, since A2A was all ready hooked up, I could import it into Gemini Enterprise as a custom made agent, to seamlessly integrate with Google Drive, NotebookLM etc.

The AP Showcase agent imported into Gemini Enterprise, processing an Acme GmbH invoice and returning a needs_review verdict

That process alone was good to iron out some kinks with document transfer etc.

I would do it again​

All in all entering the competition was a great experience despite not being able to claim any prize and I will look at others in the future. It helps hone your skills and is a good audience for cutting edge techniques. I also now have a good demo to show in the workshops and for potential clients looking for AI engineering or AI platforms in the future.