Engineering inside ForgeQuote /
The data model, constraints and recovery decisions behind ForgeQuote.
Turn unstructured requirements into reviewed, precisely priced quotations with traceable catalogue versions.
Interface and interaction architecture
The dashboard and workflow use React; the worksheet runs an independently bundled AG Grid engine. The server retains catalogue, approval and exact-money rules.
Model interpretation does not set prices
Extraction produces structured requirements and questions from supplied text and catalogue context. The quote handler performs a separate SKU lookup in the chosen catalogue; an absent SKU is rejected. Prices are read from that catalogue record. This keeps commercial arithmetic independent of how persuasive or confident the model's language appears.
Exact bounded arithmetic
Prices and quantities use validated integer strings, with BigInt multiplication and addition. Tax is represented in basis points and rounded once using the implemented integer formula. Unsupported magnitude is rejected rather than silently converted to an inexact floating-point value. The supported model uses two-decimal currencies and whole quantities, so fractional unit pricing needs another contract.
A quotation is a reviewed version
Saved lines retain source references, units, prices and calculated totals together with catalogue identity and price-version label. Edits create a new content digest and return the quote to draft. Approval validates both record version and reviewed digest, and unresolved clarification items block it. The PDF and CSV expose the saved state for inspection.
Local persistence and bounded provider use
Workspace records, operation results and change events are persisted behind authenticated, scoped API access. Mutation requests carry CSRF protection and idempotency keys. Optional model requests pass through the application's consent check and save the returned run evidence. A replay returns the recorded run; local account access does not imply configured email, CRM or ERP connectors.
Operating the product
Start with the public interface preview to explore overview and insights using example records. Open /dashboard for account-scoped metrics calculated from your saved records, or /desk for the underlying workflow.
Keep a named, reviewed price catalogue and an accountable owner for each change. Check the version label on the saved quote and export.
When a stale-version or stale-digest response appears, reload the quote and review the current quantities and exception resolutions before approving again.
If extraction is unavailable or incomplete, enter and review requirements manually. Do not turn a missing provider or unknown SKU into an assumed price.
Retain quotation records, source references, catalogue snapshots and exports in the application's backup scope. Test the intended commercial handoff separately before using it with customers.
Independent by construction
ForgeQuote has its own application source, build configuration, local server, database, object store and session cookie. The fifteen applications reuse copies of a common foundation, but their operational data and credentials are independent. A shared dependency cache on this Mac saves disk space; each product includes a lockfile and a command to install its own dependencies.
Mutations validate the workspace and current record version. Approval binds the reviewed content digest; an idempotency key prevents an identical retry from becoming a second action. These controls are implemented in the server, alongside the product-specific rules described above.
A similar problem in your business?
Bring a real workflow, representative inputs and the result that needs to be reliable.
Shape a project brief