Inspiration
Small businesses lose public-sector opportunities before proposal writing begins. Tender notices are fragmented across portals, dense with eligibility requirements, and repeatedly amended. Small teams spend scarce time assessing opportunities they cannot pursue, while strong-fit tenders are missed because the supporting evidence is difficult to assemble quickly.
LT SmartTender turns fragmented discovery into a defensible procurement decision workflow.
What it does
LT SmartTender ingests opportunities from authoritative public sources, preserves immutable notice versions, and evaluates each tender against a subscriber's approved capabilities, locations, certifications, delivery capacity, and explicit exclusions.
Every evaluated opportunity receives one source-cited advisory outcome:
- Bid — approved evidence supports a direct pursuit.
- Partner — the opportunity is viable with a qualified delivery partner.
- Monitor — material facts remain incomplete or may change.
- Decline — a documented eligibility, capability, timing, or risk constraint makes pursuit unsuitable.
The product also exposes evidence gaps, risks, next actions, human review, amendment history, and a compliance matrix. High-impact actions stay under human control. LT SmartTender is a qualification and decision system; it is not an autonomous bid-submission tool.
The v1.0 launch sources are US SAM.gov and EU TED. UK Find a Tender, CanadaBuys, and Singapore GeBIZ are scheduled for v1.1 only after each adapter passes source-access, permitted-use, normalization, citation, eligibility, amendment, reliability, privacy, pricing, support, and pilot-readiness gates.
Subscriber workflow
- Create an organization workspace and verify an account with email/password, Google, or Microsoft Entra work/school login.
- Complete organization basics, operating markets, capability facts, approved evidence, saved searches, and notification settings.
- Receive deduplicated opportunities from live sources with explicit fit, deadline, and risk signals.
- Open an opportunity to see its recommendation, confidence, official citations, matched evidence, gaps, risks, and next actions.
- Accept the advisory decision, record a reasoned override, save, monitor, decline, or create a pursuit.
- Build an amendment-aware compliance matrix and record outcomes without deleting the audit trail.
The first onboarding step now persists in the live Supabase project. A verified server action invokes a service-role-only, idempotent database function. One transaction creates the organization, active owner membership, default subscription, default entitlements, five progress rows, a private retry record, and one durable audit event. Repeating the same request returns the original organization instead of creating duplicates.
How we built it
The application uses Next.js, TypeScript, Supabase Postgres/Auth/Storage, the Google GenAI SDK, and Google Cloud. The responsive web application is the primary workspace. Android Google Play distribution remains the v1.0 mobile target. iOS App Store distribution and Apple sign-in remain disabled for v1.0.
The Gemini pipeline has two fixed server-side roles:
gemini-3.1-flash-liteperforms bounded triage and routing.gemini-3.5-flashperforms evidence synthesis and advisory procurement decisions.
The implemented @google/genai boundary uses provider JSON schemas plus independent Zod validation. The triage model cannot issue a customer decision. Decision validation rejects malformed JSON, missing citations, unknown evidence identifiers, changed citation labels or URLs, incomplete evidence, contradictory evidence, model-created deterministic exclusions, and unreviewed Bid outputs. The Gemini credential remains server-only. A live deployed Gemini call is still credential-gated and is not represented as complete.
Supabase provides forced row-level security on exposed application tables, private organization-scoped evidence storage, immutable opportunity versions, durable amendment events, tenant-visible agent-run summaries, private redacted agent steps, immutable human reviews, and append-only audit events. The organization-bootstrap function cannot be called by anonymous or authenticated browser roles. Live rolled-back checks confirm retry idempotency, member reads, viewer write denial, outsider isolation, and one audit event per bootstrap. The Supabase security advisor reports zero findings.
Google Cloud staging includes an immutable Cloud Run image at verified source revision 022e32d, a public health route, a scheduled smoke job, workload-specific service identities, empty Secret Manager containers, structured Cloud Logging, an uptime check, and alert routing. The product display-name build has not yet replaced that verified staging revision.
Challenges
The hardest problems are provenance, amendment drift, multi-jurisdiction eligibility, tenant isolation, abstention calibration, and presenting enough reasoning for a founder to trust a recommendation. The design treats retrieval, structured extraction, deterministic checks, immutable citations, and human approval as core product behavior.
A second challenge is truthful progress reporting. Static and controlled UI surfaces are labeled as such until their server persistence is complete. Planned country sources remain visibly planned. Missing credentials remain explicit gates. A passing unit test is not described as a production customer outcome.
Accomplishments so far
- Implemented normalized SAM.gov and TED adapters with official-URL validation, bounded pagination, deterministic hashing, and sanitized failures.
- Validated the TED v3 adapter against the live public endpoint. Live SAM.gov validation remains API-key gated.
- Implemented immutable opportunity versions, replay idempotency, durable amendment events, fixed Gemini routes, strict evidence and citation validation, persisted agent runs, and immutable decision reviews.
- Implemented secure email/password, Google, and Microsoft Entra work/school authentication flows. Provider-console configuration and clean-browser verification remain.
- Implemented a protected responsive subscriber workspace with dashboard, opportunity feed, organization onboarding, capability profile, compliance matrix, evidence upload states, cited decision detail, saved-search preferences, pursuit actions, billing usage states, and explicit online/offline behavior.
- Applied the resumable onboarding schema and hardened its privileged bootstrap behind the trusted server boundary.
- Reached commit
406b131. GitHub Actions CI #26 passes lint, typecheck, 249 tests across 47 files, production build, Cloud Run container build, dependency audit, and credential scanning.
The canonical repository is gilberthlimocon/LT-SmartTender. Pre-launch privacy, terms, and deletion drafts use support.lt-smarttender@limocontechnologies.com.
What we learned
Procurement teams do not need another tender feed. They need an evidence-backed system that explains why an opportunity fits, what could disqualify them, what changed, and what action is safe to take next. Trust, traceability, and disciplined abstention are product value.
What's next
The next vertical slice will complete the remaining onboarding forms and authoritative minimum-profile unlock, then run a live TED notice through both Gemini roles, validate every citation against the immutable source version, and display the advisory decision for human review. Remaining release gates include Gemini and SAM.gov credentials, Google and Microsoft provider configuration, billing-provider integration, Storage API deletion evidence, Android release evidence, clean-browser PWA validation, a public judge environment, the demo video, and production redeployment.
Built With
- codex
- gemini
- google-cloud
- gpt
- next.js
- postgresql
- supabase
- typescript
- vitest
Log in or sign up for Devpost to join the conversation.