This repository contains three sanitized, runnable demonstrations developed from internship prototypes for enterprise operations. The public version preserves the architecture, state logic, tests, and local setup while removing employer-specific infrastructure, customer material, internal documentation, deployment details, identifiers, and credentials.
| Project | Problem | Main components |
|---|---|---|
legal-review-bot |
Classify contracts and produce checklist-based risk reviews | Flask, OpenAI-compatible LLM, embeddings, SQLite, Lark/Feishu output |
financial-reporting-bot |
Normalize expense records and generate management-report summaries | Flask, CSV/XLSX parsing, LLM classification, SQLite, reporting API |
customer-service-bot |
Answer product questions after a human-response window and escalate high-risk cases | State machine, retrieval, LLM response synthesis, cache, CSAT, tests |
- Human escalation before automation. Legal, billing, security, refund, and churn-risk messages are routed to staff instead of being answered automatically.
- Structured outputs. The model is constrained to classification or review schemas rather than free-form generation alone.
- Retrieval before response. Answers and reviews are grounded in an approved document collection or synthetic templates.
- Operational observability. Health endpoints, SQLite event records, latency tests, and explicit failure states make the prototypes inspectable.
- Secrets stay outside Git. All credentials are read from environment variables; example configuration files contain placeholders only.
intern/
├── legal-review-bot/
├── financial-reporting-bot/
├── customer-service-bot/
├── .gitignore
└── README.md
Each project has its own setup instructions, architecture notes, limitations, and test commands.
- Removed real server addresses, ports, App IDs, chat IDs, employee names, customer records, internal repository URLs, and SSH deployment scripts.
- Replaced organization-specific model gateways with generic OpenAI-compatible environment variables.
- Removed historical customer-service QA and replaced real product branding with
ExampleProduct. - Retained only synthetic contract and expense examples.
- These are portfolio demonstrations, not production services.
- The legal-review project is not legal advice and requires qualified human review.
- The financial project is not an accounting or audit system.
- The customer-service project requires an approved documentation corpus before use.
- Reported demo accuracy is not presented as a general benchmark.
Code is released under the MIT License. Third-party trademarks and platform APIs remain the property of their respective owners. No employer or client data is included.