Hacktoberfest 2026 · Side Hustle · Open-weight AI
ScopeLedger
Keep good work profitable.
A client asks for one more thing. Is it included, extra work, or too vague to estimate? ScopeLedger compares the ask with the agreement, shows the source lines, and helps you prepare a change-order draft. AI proposes. You decide.
Prototype, not an automatic quoting service. The website hosts a hand-authored sample walkthrough, not live model inference. For real documents, download the app and run Gemma on your own computer. No automatic invoices or client messages are sent.
Evidence before an extra invoice.
01 / Compare
Paste the agreed deliverables and new requests, one item per line.
02 / Inspect
Review each suggestion beside the exact request and referenced scope lines.
03 / Decide
Override the suggestion, enter your effort and rate, and select changes to quote.
04 / Agree
Export the owner-reviewed draft. Discuss it with the client before starting work.
Measured, not assumed
The small model was not enough.
Gemma 3 1B matched 1 of 8 development cases. Gemma 3 4B matched 4 of 8 with the initial prompt, then 6 of 8 after adding explicit rules for vague, conflicting and mixed requests. Both remaining failures are published.
These are eight synthetic development cases reused during prompt refinement, not a held-out accuracy benchmark. The model still mishandled an included revision and a mixed request. Every decision needs owner review.
Inspect the evaluation report →What is verified?
- 23 application and HTTP tests passed during the project build.
- The DOM workflow covers estimates, overrides, export and stale-decision clearing.
- A real local-model API-to-export run passed with explicit owner decisions.
- Source IDs and prices are validated by code; semantic judgments are not guaranteed.
No paying-user, recovered-revenue or time-saving claim has been validated.
Your client documents. Your machine.
Install Python 3.10+ and Ollama, download Gemma 3 4B, then start the application. After setup, the live app does not depend on a hosted AI API. Allow several GB of memory and time for CPU inference. Local SQLite records are not encrypted.
Gemma is open-weight under Google’s terms. The application code is MIT-licensed; no model weights are bundled. This is a single-user local prototype, not legal advice.
Quick start
ollama pull gemma3:4b
python3 server.py
# Open http://127.0.0.1:8765