<aside> 💡

Paste any extra notes below. Then ask Notion AI: “Rewrite this into a clean portfolio case study, keep it concise, keep the technical clarity.”

</aside>

Overview

What it is: An autonomous RFP + vendor questionnaire copilot that turns messy procurement docs into structured, review-ready responses.

Goal: Cut RFP completion time from days to hours without risking hallucinated compliance/legal claims.

Key idea: Evidence-backed drafting + confidence scoring + human-in-the-loop approval routing.

Architecture (6-layer autonomy stack)

1. Intake + file normalization (universal ingest)
   • Inputs: email attachments (PDF/DOCX/XLSX), shared drive folders, portal downloads, copy/paste text.
   • Normalizes into internal representations:
     • PDF → text + layout blocks + tables
     • DOCX → sections/headings/table rows
     • XLSX → sheet/row/cell mapping
     • Email threads → message-level context
   • Outputs:
     • raw_documents/
     • normalized_chunks.json (page/section/table coordinates)
2. Question extraction + structuring (question miner)
   • Detects paragraph questions, table-row questions, multi-part (a/b/c), yes/no + follow-ups, evidence requests.
   • Writes each question as a structured record:
     • question_id
     • source doc + page/cell reference
     • category (security/privacy/legal/finance/ops)
     • expected answer type (yes/no, short, long, attachment)
     • priority/due date
     • risk level estimate
   • Dedupes with fuzzy matching and groups into canonical “question families”.
3. Knowledge base + evidence index (truth layer)
   • Builds an indexed KB from:
     • Policies (security/privacy/data retention)
     • Product/architecture docs
     • Past RFP submissions (approved “gold answers”)
     • Compliance frameworks (SOC2/ISO mapping)
     • Standard clauses
   • Each chunk is labeled with owner, last updated, allowed usage, and approval tier.
   • Answers must attach evidence chunk IDs + citations.
4. Answer composition engine (guarded drafting)
   • Retrieval-first:
     • Reuse approved canonical answers when possible.
   • Draft generation when needed:
     • Strict constraints (no promises beyond KB, no legal commitments, no security claims without evidence).
     • Suggest required evidence attachments.
   • Confidence scoring:
     • score 0–1 + “why” + reviewer type recommendation.
5. Review workflow + risk routing (human-in-the-loop)
   • Routes high-risk domains:
     • Security (encryption, residency, breach timelines)
     • Legal (indemnity, liability, SLAs)
     • Finance (pricing, payment terms)
   • Produces reviewer queues + diff vs last approved answer.
   • Only approved answers can be exported.
6. Export + submission packager (final output generator)
   • Produces:
     • Filled questionnaire spreadsheet (cell-perfect)
     • PDF summary + evidence list
     • Open items report (blocked/unanswered)
     • Submission checklist
   • Stores final approved answers back into KB + audit trail (who approved what, when).

End-to-end flow

1. Intake: drop files or forward email → create run/job ID.
2. Extraction: question miner builds the question DB.
3. Categorization: auto-label by department.
4. Retrieval + drafting: evidence-backed drafts + citations + confidence.
5. Review routing: low-confidence/high-risk goes to queues.
6. Export: inject approved answers back into original formats + generate packet.

Key components (conceptual)