FDE AI Systems Portfolio

Two runnable enterprise AI systems for demonstrating secure RAG, governed agents, evals, traces, audit logs, and approval gates.

Most AI app demos stop at chat. Real enterprise deployments need permission boundaries, evidence, human approval, debugging surfaces, and regression tests. This repo implements those patterns in two local-first systems that run without paid APIs, while leaving clean upgrade paths to OpenAI Responses API, Agents SDK, PostgreSQL/pgvector, OpenTelemetry, and enterprise connectors.

Projects

Project	What It Demonstrates	Local URL
Secure Enterprise Knowledge Copilot	Permission-aware RAG, citations, abstention, prompt-injection handling, traces, audit logs, evals	http://127.0.0.1:8765
Regulated Customer Operations Agent	Tool calling, business workflow automation, approval queue, side-effect blocking, supervisor approval, unsafe-action evals	http://127.0.0.1:8770

Why This Exists

FDE / AI application interviews increasingly test whether you can build systems around models, not just call a model. These projects focus on the parts that usually separate production AI systems from demos:

• permissions before model generation
• citations and abstention instead of unsupported answers
• retrieved-content prompt-injection handling
• tool permissions and approval gates
• trace and audit surfaces
• eval gates and smoke tests
• clear production upgrade paths

Quickstart

Verify everything from a clean checkout:

python -B scripts/dev.py verify

Start both demos:

python -B scripts/dev.py start

Or start them separately:

cd secure-enterprise-knowledge-copilot
python -B app.py --reset --port 8765

cd regulated-customer-operations-agent
python -B app.py --reset --port 8770

Useful commands:

python -B scripts/dev.py health
python -B scripts/dev.py evals
python -B scripts/dev.py smoke
python -B scripts/dev.py report
python -B scripts/dev.py quality
python -B scripts/post_publish_check.py

Current verified status:

health check: both services ok
smoke tests: 9/9 passed
Project 1 eval: 11/11 passed, unsafe_leak_failures = 0
Project 2 eval: 8/8 passed, unsafe_direct_side_effect_failures = 0

Evidence Matrix

Production Concern	Where To Look	Verification
Permission-aware RAG	secure-enterprise-knowledge-copilot/src/copilot/retrieval.py	Alice finance query abstains; Morgan finance query answers
Prompt-injection handling	secure-enterprise-knowledge-copilot/src/copilot/security.py, secure-enterprise-knowledge-copilot/src/copilot/answering.py	eval-005, eval-008 through eval-011
Governed tool use	regulated-customer-operations-agent/src/ops_agent/tools.py	direct send_notice is blocked for investigator
Human approval	Project 2 approval queue and supervisor endpoint	supervisor approval sends the notice once
Regression gates	scripts/dev.py, project eval runners	python -B scripts/dev.py verify

See Portfolio Evidence Matrix for the full claim-to-evidence map.

Screenshots

Secure Enterprise Knowledge Copilot	Regulated Customer Operations Agent

Project 1: Secure Enterprise Knowledge Copilot

Open:

http://127.0.0.1:8765

Show:

1. Alice asks: How many days per week can employees work remotely?
2. The system answers with Remote Work Policy 2026 citation.
3. Alice asks: What is the finance retention plan?
4. The system abstains because Alice cannot access confidential finance evidence.
5. Morgan asks the same finance question.
6. The system answers with Finance Retention Plan 2026 citation.
7. Run evals.

Core claim:

The model never receives evidence the user is not allowed to access.

Project 2: Regulated Customer Operations Agent

Open:

http://127.0.0.1:8770

Show:

1. Ivy investigates Market Blue / RX-900 recalled product.
2. The agent searches policy and listings.
3. It creates a violation, drafts seller notice, schedules follow-up.
4. It creates an approval request before sending.
5. Direct send_notice is blocked for investigator.
6. Supervisor approval sends the notice.
7. Run evals.

Core claim:

The model may propose actions, but side effects are enforced by application logic.

Architecture

Portfolio
  +-- Secure Enterprise Knowledge Copilot
  |   +-- role-aware retrieval
  |   +-- citation answer shape
  |   +-- abstention logic
  |   +-- prompt-injection detection
  |   +-- trace/audit/evals
  |
  +-- Regulated Customer Operations Agent
      +-- intent routing
      +-- business tools
      +-- approval queue
      +-- side-effect blocking
      +-- trace/audit/evals

Both projects are dependency-free by default so they run reliably anywhere with Python 3.12. Optional OpenAI gateways are included but disabled by default.

Optional OpenAI Mode

$env:OPENAI_API_KEY="..."
$env:OPENAI_MODEL="gpt-5.2"
$env:OPENAI_REASONING_EFFORT="medium"
$env:OPENAI_TEXT_VERBOSITY="low"
$env:COPILOT_MODEL_PROVIDER="openai"
$env:OPS_AGENT_MODEL_ROUTER="openai"

Security boundaries remain outside the model:

• Project 1 filters permissions and unsafe retrieved content before generation.
• Project 2 enforces approval gates in deterministic application code.

Docker

Docker config is included:

docker compose up --build

Docker config is included for Docker-enabled machines. The local Python runtime is the verified path.

Repository Structure

fde_portfolio/
  secure-enterprise-knowledge-copilot/
  regulated-customer-operations-agent/
  scripts/
  docs/
  .github/

Key Docs

• Project Content Index
• Changelog
• Final Demo Runbook
• Demo Report
• Resume And Interview Package
• Production Upgrade Notes
• Model Runtime Configuration
• Final Completion Audit
• GitHub Launch Plan
• Published Repository Status
• GitHub Repository Settings
• Community Backlog
• Public Release Audit
• Differentiation Strategy
• Hard Interview Playbook
• System Design Deep Dive
• Portfolio Evidence Matrix
• ADR 0001: Local-First Portfolio Runtime
• ADR 0002: The Model Is Not The Security Boundary
• ADR 0003: Eval State Isolated From Demo State
• Secure RAG Case Study
• Governed Agent Case Study
• Demo Video Script
• Demo Recording Checklist
• Star Growth Plan
• Launch Copy Pack
• GitHub Initial Issues
• Reviewer Perspective Checklist
• GitHub Release Commands
• Post-Publish Checklist
• Roadmap

Interview Narrative

I focus on deploying AI into real business workflows. The first project handles enterprise knowledge access with permissions, citations, abstention, traces, audit logs, and evals. The second project connects an agent to operational tools while preventing unsafe side effects through approval queues and governance checks.

Together they show I can build AI systems that are useful, inspectable, and safe enough for enterprise deployment.

License

This project is released under the MIT License. See LICENSE.