supercode168/enterprise-workflow-ai-agent

Enterprise Workflow AI Agent

An AI workflow automation MVP for cross-functional enterprise teams. It demonstrates how one AI platform can support service operations, investment research, and compliance review through structured outputs, retrieval, checklists, drafts, dashboards, and API-ready endpoints.

This project is designed for a Forward-Deployed AI / AI Transformation intern profile: it demonstrates business problem framing, data pipeline design, structured AI outputs, retrieval, dashboarding, and API-ready architecture.

Why This Project

Nan Fung's internship JD emphasizes building production-grade AI tools across business teams: agents, automated workflows, chatbots, data pipelines, dashboards, APIs, and enterprise system integration.

This MVP maps directly to that requirement:

• —Business workflow: service request triage and follow-up
• —AI layer: structured classification, urgency detection, checklist generation
• —Retrieval layer: similar historical tickets and prior resolutions
• —Data layer: SQLite ticket database
• —Interface layer: browser dashboard and JSON API
• —Upgrade path: LLM API, RAG vector store, FastAPI, Streamlit/React, MCP tools

Current Features

• —Load sample maintenance tickets
• —Optional NYC 311 data fetch script
• —Rule-based structured triage
• —Similar case retrieval with lightweight TF-IDF
• —Technician checklist generation
• —Tenant reply drafting
• —Investment document summarization and signal extraction
• —Compliance obligation extraction and risk checklist generation
• —SQLite persistence
• —Browser dashboard
• —JSON API endpoints

Demo Screenshots

Local verification screenshots are generated under screenshots/. They are intentionally not pushed to Hugging Face in the lightweight Space version because the Hub requires Xet/LFS for binary assets.

Architecture

flowchart LR
    A["Ticket / document input"] --> B["Structured triage engine"]
    B --> C["Category and urgency"]
    B --> D["Responsible team"]
    B --> E["Missing information"]
    A --> F["Historical case retrieval"]
    F --> G["Similar resolutions"]
    C --> H["Technician checklist"]
    D --> H
    G --> H
    C --> I["Tenant / stakeholder reply"]
    H --> J["Dashboard and API"]
    I --> J

Project Scenarios

1. 1.Service Operations Copilot
2. 2.Classify tenant or service requests
3. 3.Retrieve similar historical cases
4. 4.Generate technician checklist and tenant reply

1. 1.Investment Research Copilot
2. 2.Summarize annual reports and company documents
3. 3.Extract risks, revenue signals, and deal memo fields

1. 1.Compliance Review Copilot
2. 2.Extract obligations, dates, clauses, and risk flags
3. 3.Generate review checklist

Run

Hugging Face Spaces

The Space runs the Gradio app:

python app_hf.py

Local HTML/API version

cd property_maintenance_ai_agent
/Users/yiplimin/.cache/codex-runtimes/codex-primary-runtime/dependencies/python/bin/python3 scripts/load_sample_data.py
/Users/yiplimin/.cache/codex-runtimes/codex-primary-runtime/dependencies/python/bin/python3 app.py

Then open:

http://127.0.0.1:8765

Pages:

http://127.0.0.1:8765/
http://127.0.0.1:8765/investment
http://127.0.0.1:8765/compliance

Try the JSON API:

http://127.0.0.1:8765/api/triage?text=There%20is%20no%20hot%20water%20in%20the%20unit%20and%20the%20tenant%20has%20a%20baby
http://127.0.0.1:8765/api/investment?text=Revenue%20grew%2018%25%20but%20debt%20increased%20to%20USD%20120%20million
http://127.0.0.1:8765/api/compliance?text=The%20supplier%20shall%20maintain%20confidentiality%20and%20approval%20is%20required

Resume Bullet Draft

Built a multi-scenario enterprise workflow AI agent MVP for service operations, investment research, and compliance review, combining structured ticket triage, urgency detection, historical case retrieval, document signal extraction, risk flagging, checklist generation, stakeholder draft generation, SQLite persistence, browser dashboarding, and JSON API endpoints.

Interview Story

I built this project to simulate how an AI transformation intern could work across different business teams rather than being limited to one function. The system starts with service operations because maintenance tickets are concrete, frequent, and workflow-heavy. I then extended the same architecture to investment research and compliance review to show that the core pattern is reusable: take messy business input, convert it into structured output, retrieve relevant context, generate an action checklist, and expose the result through a usable interface.

The current version runs without paid LLM APIs so it is easy to demo publicly. The next version would replace deterministic rules with LLM structured outputs, add RAG over real internal documents, and expose MCP-compatible tools for enterprise system integration.

Next Upgrade Steps

1. 1.Replace rule-based triage with LLM structured output.
2. 2.Add RAG over NYC 311/HPD historical resolution descriptions.
3. 3.Add real investment reports and compliance policy samples.
4. 4.Add FastAPI once dependencies are available.
5. 5.Add OpenAI/Claude function calling and MCP-style tool wrappers.
6. 6.Add screenshots, evaluation metrics, and a short demo video.