AI-Assisted RFP & Document Analysis System

LLM & NLP — Proposal Intelligence Workflow (Concept Prototype)

Built a prototype AI-assisted workflow to analyze large RFP documents (50–200 pages), extract structured requirements, and score alignment against an internal capability matrix. The focus is speed + consistency: turning unstructured PDFs into analyst-ready outputs (requirements list, risks, gaps, and an executive fit score).

Use case: 50–200 page RFPs Output: requirements + risks + fit score Pipeline: chunking → retrieval → extraction → scoring

Key Contributions

Document segmentation: designed chunking rules and section-aware splitting to preserve context (scope, SLAs, compliance, pricing assumptions).
Extraction schema: defined a structured requirement model (category, requirement text, priority, evidence span, owner, due date, risk level).
Scoring logic: created a capability-alignment rubric (covered / partial / gap) with risk flags for non-standard terms and high-effort obligations.

Tools: LLM workflow design, retrieval-based chunking, prompt engineering Domain: Insurance proposals / RFP operations

Data

Inputs: PDF RFPs (unstructured), internal capability matrix (spreadsheet/table), optional past proposal library for retrieval.
Output formats: JSON requirements table + risk log; optional CSV export for analysts; summary page for exec review.
Privacy: prototype uses sanitized / synthetic RFP excerpts for portfolio demonstration.

Analyst Experience

Executive summary: overall fit score + top risks + top gaps (bid/no-bid ready).
Requirements table: filterable list by category, owner, priority, and due date; includes citation/evidence span back to the source section.
Gap report: highlights partial coverage and missing capabilities with recommended follow-ups for SMEs.
Exports: CSV/JSON outputs to plug into proposal trackers and response planning templates.

How it Works

1) Ingest: parse PDF → detect headings/sections → normalize text and metadata (page, section, paragraph).
2) Chunk + retrieve: split into section-aware chunks; retrieve top relevant chunks per requirement category.
3) Extract: LLM generates structured requirements using a fixed JSON schema + evidence references.
4) Score: compare extracted requirements to capability matrix → label as covered/partial/gap + risk severity.
5) Review layer: analyst approves/edits high-risk items; exports final tables and summary.

Guardrails: schema validation, confidence thresholds for risk flags, and evidence-linked citations to reduce hallucinations.

Evaluation (Prototype)

Quality checks: % of extracted requirements with evidence spans; schema-valid output rate; duplicate/overlap rate.
Human review: sampled requirements reviewed for correctness and category placement; risk flags reviewed for false positives.
Operational goal: reduce repetitive scanning and improve consistency across reviewers (portfolio prototype).