LT SmartTender

Evidence-backed AI procurement decisions for global tenders, launching with US and EU sources and expanding to the UK, Canada, and Singapore.

Comment

posted an update —

Milestone 3 — Gemini safety boundary and 91-test regression

LT Tender Trace now has a tested, credential-safe two-model Gemini server boundary.

Completed

  • Added the official @google/genai 2.12.0 SDK as an exact dependency.
  • Fixed triage routing to gemini-3.1-flash-lite and decision routing to gemini-3.5-flash.
  • Added provider JSON response schemas plus independent Zod validation.
  • Made every customer decision impossible at the triage stage.
  • Bound decision citations to the supplied evidence identifier, label, HTTPS URL, and access date.
  • Rejected malformed JSON, invented citations, invented requirement identifiers, model-created deterministic exclusions, incomplete evidence, contradictory evidence, and unreviewed Bid outputs.
  • Added an accessible agent-run panel for queued, triaging, deciding, completed, and needs-review states.
  • Verified lint, TypeScript checking, 91 automated tests across 18 files, and the Next.js production build.

Current evidence

  • TED v3 remains live-validated.
  • Supabase immutable notice versions and durable amendment events are live.
  • Supabase security advisor remains clear with zero outstanding findings.

Still pending

A live deployed Gemini call requires the server API credential. SAM.gov live validation separately requires its API key. Google and Microsoft sign-in still require provider-console configuration and clean-browser verification. These items are not represented as complete.

Next milestone

Run a live TED notice through both Gemini roles, persist the immutable agent run, validate every citation, present the advisory decision for human review, and begin the expert-labeled US/EU golden-set evaluation.

Log in or sign up for Devpost to join the conversation.