A self-hosted system connecting tender intake, document extraction, supplier research, pricing, approvals and finance. Models run on the company’s hardware; people review records and approve outgoing actions. Delivered in phases, this illustrates a broader implementation rather than the scope of a starter pilot.
Client work under NDA. We show the problem, the system and the results, but not the client's name or screenshots. More detail is available on a call.
One integrated platform, delivered in phases as deployment packages. Agents read the mail, extract the documents, source suppliers, draft the outreach and chase the replies, while a pricing engine reproduces the company's own sheet to the cent. Three hard rules run through all of it: the model handles language and code handles numbers; nothing becomes a record without a person; nothing is sent without a person.
Every connected mailbox is read. A router agent classifies each message (new request, quotation, purchase order, shipment update, invoice, supplier reply) and files it against the right record. Anything ambiguous goes to a short triage list instead of being guessed.
PDFs, spreadsheets, Word files and scans, with OCR only where it's needed. Extracted fields are shown beside the original with low-confidence values flagged, and nothing is saved until a person confirms it.
Per-line supplier research with token and time budgets, a knowledge graph of suppliers, products and routes, and drafted enquiries and follow-ups that a person reviews and sends.
Duty, freight, landed cost, margin and selling price are fixed arithmetic, never a model output. One screen shows the whole bid with all evidence, and approval is stored against the exact document version.
A multi-currency ledger, instalments from each purchase order's terms, working-capital exposure per deal, shipment tracking with document checklists, and a daily digest whose every figure is checked by a verifier agent.
Self-hosted Ollama and llama.cpp with open-source models, GPU if present and CPU if not. Built entirely on dummy data with no developer access to production, and handed over as Docker deployment packages.