EthosPowerAIReference design · Intelligence
Confidential RFP Scoping on Local Models
- Scoper
- Ollama
- gemma3:27b
- FastAPI
- SQLite
- Git
The problem
A tender for a substation battery replacement arrives with drawings, site constraints and a deadline. Pasting it into a public chatbot to scope it faster sends the customer's documents and the firm's own rates to someone else's server.
What I would build
Scoper's Standard chain runs on a local model behind Ollama: ARIA parses the tender and extracts the requirements and the ambiguous scope, NOVA drafts phases, hours per role and risks, CHRONOS builds the programme and flags timelines that assume an unbooked outage, and PRISM writes acceptance criteria. The design replaces the rates Scoper hardcodes in a prompt with a version-controlled rate file the firm owns, pins the provider to Ollama and checks the pin on every start, and adds a statement-of-work template filled from the JSON estimate, which an engineer prices, edits and signs.
The logic flow
Click any node for what it does, what it hands on, and where a person is required.
What it returns
This is a reference design on someone else's open-source project, not a system I run, so there is no measured saving to quote. Scoper's README claims an estimate in 'two minutes' and its author's draft says 'within 20-30%' on his own testing; neither has been checked independently. The number I would watch is each delivered job's actual hours against the signed estimate, kept beside the rate file.
Why it fits
A utility's procurement team can now hand a tender for a substation battery replacement to its own assistant and get back a list of requirements, a rough programme and a guess at the hours before the bidders have opened the email. The first read of a tender is no longer scarce. A firm that answers by pasting the tender into a public chatbot to bid faster has kept the old habit and added a new exposure, because the drawings, the site constraints and its own rates now sit on somebody else's server.
What the customer's assistant cannot know is what this firm charges, how long its crews take to change out a string of cells in a live substation, and which clauses have cost it money before. So I would keep the whole chain on one host. A local model behind Ollama reads the tender, pulls out the requirements, drafts a phased estimate with hours for each role and flags the scope marked 'to be confirmed' and any timeline that assumes an outage nobody has booked. The rates come from a version-controlled file the firm owns, not from a prompt. The provider stays pinned to Ollama, and a person prices the work and signs the statement of work before anything leaves the building.
One five-year strategy it could serve, offered as a sketch rather than a plan: bid less on hours and more on the evidence behind them. Every tender the firm scopes and then delivers adds an actual against an estimate, and after a few dozen the rate file and the risk flags rest on its own record instead of a model's guess, which is a record the customer's assistant has no way to read. The firm can then tell a customer how far out its last twenty estimates were. Few competitors can.
See it run
Want one of these?
Every part of it is open source and self-hostable. What I bring is the build, the checks around the project's gaps, and the judgement about where your gates should sit.