AI-Powered Diligence Automation Roadmap
Prepared for Strike Capital
Executive Summary
This roadmap outlines the development of an internal AI-powered "diligence co-pilot" that automates and standardizes Strike Capital's venture diligence process. The system will evolve through three phases: V0 (Pre-Call DD), V1 (Post-Call DD), and V-Infinity (Advanced DD).
Timeline: MVP (V0) deliverable in 2 weeks upon data availability.
Technical Architecture
Core Technology Stack
| Component |
Technology |
Rationale |
| Foundation Model |
GPT-4o |
Latest OpenAI model with superior reasoning capabilities |
| PDF Processing |
Text extraction + GPT-4o |
Accurate extraction with sentence-level chunking |
| Embeddings |
text-embedding-3-small |
Cost-effective, high-quality embeddings (1536 dimensions) |
| Vector Database |
Pinecone |
Managed, scalable vector storage with metadata filtering |
| User Interface |
Gradio |
Rapid prototyping, Python-native, built-in file handling |
| Initial Hosting |
Hugging Face Spaces |
Fast deployment, easy team access, low cost |
| Production Hosting |
AWS (future) |
Enterprise scalability and security |
Why Gradio for the UI?
- Rapid Prototyping: Functional internal UI in hours rather than weeks
- Python-Native: Seamless integration with data science workflows
- Built-in File Handling: Native drag-and-drop support for PDFs
- Authentication: Built-in authentication options for internal tools
- Hugging Face Integration: One-click deployment to Spaces
- Migration Path: Can be wrapped in FastAPI for AWS deployment
Data Pipeline Architecture
Harmonic PDF Upload
→ Text Extraction + Sentence Chunking
→ Context Window (1 sentence before + after)
→ GPT-4o (structured data extraction)
→ Pinecone Vector DB (1 vector per sentence chunk)
→ GPT-4o (section-by-section generation)
→ Structured V0 DD Document
Vectorization & Chunking Strategy
Sentence-Level Chunking with Context Window
| Content Type |
Chunking Strategy |
Context Window |
| PDF Text |
Sentence-level |
1 sentence before + 1 sentence after |
| Extracted Sections |
Section-level |
N/A (stored as JSON) |
| V0 Document |
Full document |
N/A |
Key Design Decisions
- Sentence-level granularity: Each sentence stored as separate vector for precise retrieval
- Context window: Include 1 sentence before and after for semantic coherence
- Metadata-rich chunks: Every chunk tagged with company_id, chunk_index, version, timestamp
- Batch processing: 20 chunks per API call to reduce costs
Embedding Model: text-embedding-3-small
- 1536 dimensions (good balance of quality vs. storage)
- Cost-effective at $0.00002 / 1K tokens
- Supports up to 8191 tokens per chunk
Phase 1: V0 - Pre-Meeting DD Doc (MVP)
Timeline: 2 weeks (upon data availability)
Objective: Auto-generate a complete V0 diligence document from an uploaded Harmonic PDF report.
PDF Processing Pipeline
- User uploads Harmonic PDF to Gradio app
- Text extracted and split into sentences
- Context window applied (1+1)
- GPT-4o extracts structured data from batched chunks
- Each sentence chunk stored in Pinecone with embedding
- V0 document generated section-by-section
V0 Document Sections
- Deal Overview - Company, round size, stage, intro source
- Company Summary - Product description, users, problem solved
- Investors - Total count, existing investors, angels
- Company Overview - Founded, HQ, Headcount, ARR, Funding, Customers
- 2-3 Key Insights - High-signal facts from Harmonic data
- Team Overview - Founder bullets + team composition
- Market Overview - Target market, TAM/CAGR, tailwinds
- Competitive Landscape - Competitors, differentiation hypotheses
- Smart Talking Points - Credible talking points for founder call
- DDQs - 8-12 evidence-anchored questions
MVP Deliverables
- Gradio web app with PDF upload interface
- Sentence-level Pinecone storage (hundreds of vectors per company)
- GPT-4o prompt chains for each V0 section
- HTML and PDF export
- Deployed on Hugging Face Spaces with shareable link
Phase 2: V1 - Early-Stage DD Doc (Post-Founder Call)
Timeline: 2-3 weeks after V0 completion
Additional Data Sources
- Notion API - Read call notes, write DD documents back
- Box API - Ingest pitch decks and data room materials
- AlphaSense API - Expert calls, industry primers
Key V1 Features
- Version Control: Track V0 → V0.1 → V1
- Change Annotations: Mark updates with "⚠ Update: ..." notation
- Human-in-the-Loop: Edit sections, re-run specific parts
- Diff/Redlining View: Visual comparison between versions
Phase 3: V-Infinity - Advanced DD Doc (Long-term)
Timeline: Ongoing development after V1
V-Infinity Vision
- All V1 sources + expert calls, customer interviews, competitor data
- Long-form synthesis with scenario analysis
- Live updates from news and market developments
- Contradiction detection across sources
- AWS deployment for enterprise scalability
Quality Assurance & Anti-Hallucination Measures
Template Adherence
- JSON schema validation for each section
- Prompt constraints enforcing exact structures
- Output parsing with retry logic
Hallucination Prevention
- Source attribution required for every claim
- Default to "Unknown" when data unavailable (no fabrication)
- Surface contradictions rather than reconciling silently
- Low temperature (0.1-0.3) for consistent extraction
Deployment Timeline
| Phase |
Deliverable |
Hosting |
Timeline |
| MVP (V0) |
Pre-Call DD Generator |
Hugging Face Spaces |
Weeks 0-1 |
| V0 Refinement |
Template tuning, bug fixes |
Hugging Face Spaces |
Weeks 1-2 |
| V1 Development |
Post-Call DD + Version Control |
AWS |
Weeks 3-4 |
| V-Infinity |
Advanced DD System |
AWS |
Ongoing |
Dependencies & Requirements
From Strike Capital
- Sample Harmonic PDF reports (2-3 examples) - For testing
- Sample DD documents (2-3 V0/V1 examples) - For template calibration
- OpenAI API key with GPT-4o access
Infrastructure
- Pinecone account (free tier sufficient for MVP)
- Hugging Face account for initial deployment
- AWS account (for future production deployment)
Questions? Let's discuss any adjustments to scope, timeline, or technical approach.