Selected work

Featured project / Evidence-aware AI process discovery

SME Process Discovery Agent

An AI system that interviews users about business processes, investigates connected systems and internal documentation, builds a structured AS-IS model, and produces a verified automation proposal.

The problem

Automation needs a reliable picture of the work.

Business processes rarely live in one place. The documented procedure says one thing, employees may actually do another, and critical details are buried across systems, spreadsheets, emails, and institutional knowledge.

Automating the first description of a process can simply automate the wrong process.

The agent investigates first.

What I built

Discovery, structured state, then a verified proposal.

The application conducts a process-discovery interview while independently consulting available business systems and internal documentation. Facts are captured into application-owned structured state rather than left only in model conversation history.

When discovery has enough substance, an explicit orchestration pipeline analyzes the AS-IS process, proposes automation opportunities, and verifies those recommendations against the evidence collected.

Architecture

The application owns the system boundaries.

MCP handles operational capabilities. RAG handles semantic company knowledge. The application owns state and orchestration.

Why it is not just a chatbot

The model participates. The application stays in control.

01

Application-owned state

Model output proposes typed patches; Python validates and owns the durable process model.

02

Evidence provenance

User-reported practice, documented policy, MCP system facts, and unresolved unknowns stay distinguishable.

03

Explicit orchestration

Analyst → Designer → Verifier is application-controlled rather than autonomous agent-to-agent conversation.

04

Observable by design

The execution trace is generated from real backend events—no fake model ‘thinking’.

Discovery example

Preserve the conflict. Do not smooth it away.

Demo boundary: Northstar Industrial Services is a fictional demo company using synthetic data.

Reported practice

“No management approval unless the discount is above 15%.”

Documented policy

“Sales Manager approval is required above 10%.”

Unresolved conflict

The system preserves both claims and creates an unresolved policy/practice conflict rather than silently choosing one. Above 15%, it can then map the manual path: Sales emails Marta → Marta prints → signs → scans → emails back → Sales waits.

Application-controlled orchestration

Analyze. Design. Challenge.

  1. 01

    Process Analyst

    Identifies bottlenecks, handoffs, gaps, duplicate work, and unresolved questions.

  2. 02

    Automation Designer

    Proposes TO-BE improvements while preserving required human decisions and documenting assumptions and dependencies.

  3. 03

    Evidence Verifier

    Challenges each recommendation against the evidence and marks it supported, partially supported, unsupported, or blocked by unknown information. It is not designed to rubber-stamp the proposal.

Visual evidence

Screenshot slots prepared.

The portfolio does not yet contain approved screenshots for this project. These slots document the required real assets without presenting fabricated UI.

Real screenshot pendingDiscovery interfacesme-process-agent-discovery.png
Conversation and real Agent Activity events
Real screenshot pendingStructured process modelsme-process-agent-process-model.png
Policy/practice conflict and branch-aware process flow
Real screenshot pendingVerified proposalsme-process-agent-orchestration.png
Process Analyst, Automation Designer, and Evidence Verifier

Behavioral evaluation

Tests for code. Evals for intended behavior.

Traditional tests can prove that software behaves as programmed. They cannot fully prove that a nondeterministic AI system behaves as intended.

Hybrid strategy

  • Deterministic checks where Python can inspect behavior objectively
  • Semantic LLM judging only where meaning requires interpretation
  • Scenario-based regression testing
  • Visible failures retained as regression signals
  • Capability routing 11/11
  • RAG grounding 4/4
  • Conflict preservation 3/3
  • Verifier calibration 2/2

Public-demo hardening

Designed for a public AI surface.

  • Anonymous cookie-backed session isolation
  • Per-session ProcessState and conversation context
  • Rate and lifetime request limits
  • Global model-operation emergency fuse
  • Restricted MCP subprocess environment
  • Safe Markdown and DOM rendering
  • Non-root Docker runtime
  • Secure production cookies and health checks
  • Explicit synthetic RAG release artifact

Deployment boundary

The application is containerized with Docker, deployed on Railway, and served through a custom domain. It deliberately runs one application worker because session state is currently held in memory. Horizontal scaling would require moving that state to shared storage such as Redis or a database.

Technical stack & contribution

Engineering ownership, with AI-native implementation.

My role: system architecture, product design, AI-system behavior, trust boundaries, state model, evaluation strategy, technical direction, and deployment—using AI coding agents heavily for implementation.

Backend
Python, FastAPI, Pydantic
AI
OpenAI Responses API, GPT-6 Luna, text-embedding-3-small
Agent architecture
Tool calling, MCP, RAG, Structured Outputs, application-controlled orchestration
Frontend
HTML, CSS, JavaScript, SSE
Quality
unittest, deterministic integration tests, AI eval harness, browser smoke tests
Deployment
Docker, Railway, Cloudflare DNS/custom domain

Engineering judgment

  • Durable state belongs outside model history
  • MCP and RAG address different integration needs
  • Structured outputs remain proposals until application validation accepts them
  • AI evals and deterministic tests measure different failure modes
  • Public AI demos need spend controls and session isolation
  • Visible AI activity should correspond to actual system events

Explore the system

See the discovery flow or inspect the implementation.

Back to selected work