04 · Governed Procurement Intelligence · Governed AI / Public Sector

Public procurement: a prototype that helps people check evidence

Challenge: An anonymized public-sector procurement scenario needed connected evidence across supplier, tender, compliance, contract, and delivery records. Solution: I designed a FastAPI/LangGraph prototype with search, document extraction, role-based workflows, explanations linked to source records, and human approval. Outcome: a controlled non-production prototype; client identity, live data, scale, and automated decisions are intentionally omitted. Next: Message me to map this to a controlled workflow.

AI Product Architect2026Prototype — non-production

Case study

What was built and why it mattered.

Outcome, ownership, architecture, and evidence in one view.

This anonymized prototype explores how a procurement team could bring planning, supplier, tender, compliance, and contract records into one controlled workspace. It is intentionally shown without client identities, live data, or sensitive workflows.

The main rule is simple: software may summarise, search, and highlight evidence, but an authorised person remains responsible for every approval and award. The design combines FastAPI, LangGraph/LangChain, vector search, document extraction, a Copilot-style interface, and a role-based dashboard.

It shows the judgement needed for high-stakes software: set clear limits, link answers to sources, keep an audit trail, and stop automation from becoming the decision-maker.

The prototype connects search, document reading, explanations, and approval while keeping important decisions with people.Solid = primary flow · dashed = support / control path

Download editable Excalidraw

SOURCE_LINKS

Search, extraction, and AI-assisted analysis point people back to records instead of presenting unchecked conclusions.

HUMAN_APPROVAL

The AI layer can sort and explain information; it cannot issue approvals, awards, or compliance decisions on its own.

CONNECTED_RECORDS

Supplier, tender, evaluation, compliance, contract, and delivery records stay connected from start to finish.

WORKING_FOUNDATION

FastAPI, LangGraph/LangChain, vector search, document parsing, and a Copilot-style interface form the base for the prototype.

Confidentiality note

This is an anonymized non-production prototype drawn from a confidential public-sector procurement scenario. It intentionally omits client names, live data, and sensitive operational detail. The purpose of the case study is to show the product and AI-governance approach—not to disclose the environment it was designed for.

The problem

High-governance procurement depends on interconnected requirements, supplier information, tenders, evaluations, compliance checks, contracts, and delivery records. When those records are spread across spreadsheets, document folders, and point tools, leadership loses a coherent view of work, risk, and accountability.

The design challenge was to use AI to surface evidence and speed up review without letting the system make decisions that belong to authorized people.

My role

I designed the prototype’s product architecture, decision-support boundaries, information flow, and AI foundation. I translated a broad operating model into a controlled system that combines work management, retrieval, document analysis, and a Copilot-style interface.

Architecture

Design principles

AI assists; people authorize

The AI layer can retrieve relevant material, summarize records, highlight gaps, and help users understand a case. It is explicitly not allowed to approve, award, or make a compliance decision. The design keeps responsibility with the authorized reviewer.

Evidence has to remain visible

Generative output is useful only if a reviewer can inspect the underlying material. The prototype therefore centers the relationship between the answer, the source record, and the workflow stage that gives it meaning.

Governance is product functionality

Supplier, tender, evaluation, compliance, contract, and delivery work are treated as one connected lifecycle. This makes status, risk, and accountability legible without pretending that a model can replace formal controls.

Technical foundation

FastAPI with LangGraph/LangChain patterns for the AI and API layers.
Vector search, document extraction, and retrieval for source-aware assistance.
A Copilot-style interaction layer alongside an operator dashboard.
Role-aware workflow concepts, auditability, and human authorization as core design constraints.

Delivery status

Prototype, non-production. No claims are made here about deployment scale, live users, or automated decision-making.

What this case study proves

This project shows how I design AI systems for consequential workflows: define the boundary, make evidence inspectable, preserve human authority, and connect the model to an operating system rather than a standalone chat box.