My role
I designed the module and was the main implementer across its three codebases: the Node backend, the Python AI worker and the React frontend.
- Reply matching: Introduced a reference token in outgoing RFQ emails and built the backend logic that matches each reply to its request.
- Email ingestion: Built the workflow that routes replies from a shared mailbox, stores attachments in blob storage and posts the email to the backend.
- AI worker: Developed a stateless Python service that sends the email and its attachments natively to an LLM with strict structured outputs.
- Validation: Implemented backend re-validation of every AI suggestion against the system catalogs.
- Review UI: Built the prefilled review form and the confirmation flow in React.
Key project features
- Automatic linking of vendor replies to the request they answer; unmatched replies go to a manual triage inbox.
- Classification of each reply and per-item quote proposals with evidence and confidence.
- Versioned prompt with a prompt-injection guard: the email is always treated as untrusted data.
- Output restricted to the system catalogs and re-validated by the backend.
- Nothing is saved without explicit human confirmation.
- Field-level usage logging that measures how many AI suggestions are accepted versus edited.
Impact achieved
In real test cases, the extraction matched the source documents correctly. The module is currently in testing, with a pilot planned, and aims to remove the manual, item-by-item typing of vendor replies.
Technologies
| Technology | Use / Implementation |
|---|---|
| TypeScript | Typed code for the backend and frontend. |
| Node.js / Express | Backend for data, reply matching and validation. |
| Python / Flask | Stateless AI worker that calls the LLM. |
| React | Review interface with a prefilled form and confirmation. |
| OpenAI | Reply classification and quote extraction with structured outputs. |
| PostgreSQL | Storage of requests, replies, suggestions and usage logs. |
| Azure Blob Storage | Storage of reply attachments. |
| Power Automate | Workflow that ingests replies from the shared mailbox. |
Project flow
