AI bid intelligence & proposal OS
RFPs in. Decision-ready proposals out.
AI-assisted bid operations that connect requirements, evidence, drafting, review, QA and export in one workflow — and refuse to claim what it cannot cite.
Synthetic company Northstar Cloud Systems answering the synthetic Orion Enterprise Cloud Modernization Program: 83 requirements, 13 deliberate traps with no evidence, one hidden prompt-injection line. No real customer data.
Problem
Proposal teams lose weeks to re-finding what they already know
Requirements get missed, answers are copied from stale decks, reviewers cannot tell what is verified, and the same questionnaire is answered from scratch every quarter.
Stance
Honest by construction
Every factual sentence must cite approved evidence. Where nothing supports a claim the system says Needs evidence, generates the question for the right expert, and routes it — it never invents a certification.
Workflow
One pipeline, four human gates
- Intake & parsePDF, DOCX, XLSX, CSV, Markdown with cell-accurate locators.
- ExtractSpan-anchored requirements; injected instructions are ignored, not obeyed.
- Gate 1 · ConfirmHumans confirm the matrix.
- QualifyRadar scorecard as evidence for the bid / no-bid call (Gate 2).
- Draft & verifyHybrid retrieval → evidence gate → grounded draft → sentence-level claim verifier → label.
- SME loopGaps become routed tasks; statements become citable evidence; answers re-draft automatically.
- Gate 3 · ReviewReviewers approve; approved answers enter the library.
- Assemble, QA, Gate 4, Export16+ deterministic checks block export while blockers exist; DOCX, PDF, MD, JSON, CSV and questionnaire round-trip.
Product screens
Built for review, not chat
Requirement workspace
Split view: matrix on the left, source span, evidence rail, verified draft and review controls on the right.
Bid command center
Stage tracker, gate actions, run log, qualification radar and the primary next action.
Launch checklist QA
Ready score, blockers, warnings and info with deep links to the exact issue.
Evidence-backed AI
Labels come from facts, not confidence scores
Evidence gate
A chunk qualifies only if it covers the requirement's salient terms and mentions its hard terms (numbers, acronyms, proper nouns). The union of qualified chunks must cover every hard term, or the answer stops at Insufficient evidence.
Claim verifier
Every factual sentence needs a citation marker; every number and hard term must appear in the cited chunk; lexical overlap must clear a threshold. Unsupported sentences are highlighted, counted and blocked by QA.
Provider router
Local model → configured provider → deterministic demo floor. Sensitivity gate, cache, token budgets, schema validation with one repair, logged fallbacks. The demo needs no keys.
Prompt-injection resistance
Buyer documents are untrusted data. Instruction lines are filtered at extraction and the synthetic RFP contains one to prove it.
QA
Export is a gate, not a button
Missing mandatory answers, weak or absent evidence, missing citations, placeholders, stale or expired sources, contradictions, unsupported sentences, unresolved SME tasks, missing attachments and sections — each finding links to its entity and blocks export until resolved or dismissed with a reason.
ROI model
Estimated operational savings, formula visible and editable
hours = drafted × baseline hours × draft saving + reused × baseline hours × reuse saving + bids × admin hours × 50%. Every input is a setting on the Analytics page; every output is labelled an estimate.
Architecture
One Go binary, no external services required
Security principles
Server-side authorization
Sessions, RBAC per action, organization scoping on every query, CSRF tokens, rate limits, upload validation.
Keys never reach the browser
Provider credentials live in the environment; every result is labelled with its provider mode.
Sensitivity gate
Internal or confidential data only goes to providers explicitly allowed for it; otherwise the local model or demo floor answers.
See it run on synthetic data
3-minute script: start the pipeline → extraction with source spans → a strong answer → a trap question → answer as the SME → upload the addendum → run QA → export the round-trip questionnaire → Quality page.