← All Projects Requirements ← Overview Tech Stack Data Model API AI Pipeline Presentation →

Project 13 of ~34

📊 Due Diligence Lite — Architecture

Technical specification for Due Diligence Lite — investor research service. SvelteKit dashboard + Deno backend + Supabase + Ollama (Mac Mini) for research agents.

SvelteKit Deno Mac-native AI Research Pipeline

🏗️ Architecture Overview

Due Diligence Lite is primarily a research pipeline with a web dashboard. The core intelligence is the research agent that aggregates data from multiple sources, synthesizes findings, and compiles a PDF report. The Mac Mini serves as the永远-on research server for nightly data refreshes and async report generation.

┌──────────────────────────────────────────────────────┐ │ USER INTERFACE │ │ ┌──────────────┐ ┌─────────────┐ ┌──────────┐ │ │ │ Web Dashboard │ │ Order Form │ │ Report Viewer│ │ │ │ (SvelteKit) │ │ (pitch deck │ │ (PDF embed) │ │ │ └──────┬───────┘ └──────┬──────┘ └────┬─────┘ │ └──────────┼──────────────────┼─────────────┼──────────┘ │ │ │ ▼ ▼ ▼ ┌──────────────────────────────────────────────────────┐ │ APPLICATION LAYER (Deno) │ │ ┌────────────┐ ┌────────────┐ ┌──────────────┐ │ │ │ Order Mgmt │ │ Research │ │ Report Gen │ │ │ │ Service │ │ Orchestrator│ │ (PDF output) │ │ │ └────────────┘ └────────────┘ └──────────────┘ │ │ ┌────────────┐ ┌────────────┐ ┌──────────────┐ │ │ │ Customer │ │ Financial │ │ Notification │ │ │ │ Outreach │ │ Sanity │ │ Service │ │ │ │ Agent │ │ Checker │ │ (Resend) │ │ │ └────────────┘ └────────────┘ └──────────────┘ │ └──────────────────────────────────────────────────────┘ │ │ ▼ ▼ ┌─────────────────────┐ ┌──────────────────────────┐ │ DATA SOURCES │ │ DATA STORES │ │ ┌───────────────┐ │ │ ┌──────────────────────┐ │ │ │ LinkedIn API │ │ │ │ Supabase (Postgres) │ │ │ │ Crunchbase │ │◄────│ │ Orders, Reports, │ │ │ │ IBISWorld │ │ │ │ Research Cache, │ │ │ │ News API │ │ │ │ Customer Outreach │ │ │ │ PitchDeck (PDF)│ │ │ └──────────────────────┘ │ │ └───────────────┘ │ │ ┌──────────────────────┐ │ └─────────────────────┘ │ │ Cloudflare R2 │ │ │ │ Pitch decks, PDFs │ │ │ └──────────────────────┘ │ └──────────────────────────┘ │ ▼ ┌─────────────────────┐ │ AI RUNTIME │ │ ┌────────────────┐ │ │ │ Ollama (Mac Mini)│ │ │ │ Mistral 7B │ │ │ │ Research agent │ │ │ │ synthesis │ │ │ └────────────────┘ │ │ ┌────────────────┐ │ │ │ MiniMax API │ │ │ │ (report polish)│ │ │ └────────────────┘ │

🛠️ Tech Stack

Frontend

Component	Technology	Why
Dashboard + Order	SvelteKit	Web app: order form (upload pitch deck), report viewer (embedded PDF), order status tracking
PDF Generation	Puppeteer (server-side)	Generate investor report as HTML → PDF using a headless Chromium render. Deployed on Mac Mini or Deno Deploy.
PDF Viewer	pdf.js (Mozilla)	Embed PDF reports directly in browser. No download required.

Backend

Component	Technology	Why
Runtime	Deno 2	Runs on Mac Mini. TypeScript-first, clean permission model for running agents.
PDF Generation	Puppeteer + Chromium	Headless browser renders report HTML at high quality. LaTeX-style report formatting possible.
Queue	Upstash QStash	Report generation jobs queued async. Customer calls scheduled. Email triggers.
Payments	Stripe	Checkout session per order. Deep Dive tier includes customer outreach.
File Storage	Cloudflare R2	Store pitch decks (uploaded by investors) and generated PDF reports.

Data Sources

Source	What It Provides	Access
LinkedIn API	Founder profiles, employment history, connections	Restricted. Use PhantomBuster or Airtable social scraping as fallback.
Crunchbase	Funding history, investor list, company milestones	Basic free tier. Airtable LinkedIn data enrichment.
IBISWorld	Industry reports, market size, growth rates	Subscription. Use Statista free tier as alternative.
Google News	Company and founder news mentions	Google News API or RSS.
G2 / Capterra	Competitor reviews, pricing, market share signals	Public scraping.
Hunter.io	Find verified email addresses for customer outreach	API access.
Apollo.io	Contact data + enrichment for customer outreach	Subscription.

AI Stack

🤖 AI Philosophy: Research Agent on Mac Mini, Polish in Cloud

Research synthesis: Ollama Mistral 7B on Mac Mini — good enough for pulling data together and drafting findings
Customer call analysis: Ollama Whisper (if available) for transcription, Mistral for synthesis
Report writing quality: MiniMax M2.1 API — final polished prose for investor-facing document
Financial sanity check: Local Python scripts + Ollama for scenario modeling

🗄️ Data Model

deals
id	uuid	Primary key
company_name	varchar(255)
founder_names	varchar[]
industry	varchar(100)
stage	enum('pre-seed','seed','Series A')
pitch_deck_url	varchar(500)	R2 URL to uploaded deck
requested_by	uuid (FK, users)	Investor who ordered
tier	enum('standard','deep-dive')
status	enum('ordered','in_progress','customer_outreach','report_pending','delivered')
ordered_at	timestamp
delivered_at	timestamp
stripe_payment_id	varchar(255)

research_reports
id	uuid	Primary key
deal_id	uuid (FK)
founder_report_md	text	Founder background check
market_report_md	text	Market size analysis
competitive_report_md	text	Competitive landscape
financial_review_md	text	Projection sanity check
customer_calls_md	text	Customer interview synthesis (deep-dive only)
executive_summary_md	text	1-page overview
red_flags_md	text	Key concerns
pdf_url	varchar(500)	R2 URL to final PDF
confidence_score	integer (0–100)	Research completeness confidence

customer_outreaches
id	uuid	Primary key
deal_id	uuid (FK)
customer_name	varchar(255)
customer_title	varchar(255)
customer_company	varchar(255)
outreach_email_sent	boolean
agreed_to_call	boolean
call_notes	text
call_recording_url	varchar(500)
synthesis	text	AI-synthesized pattern analysis

users (investors)
id	uuid
email	varchar(255)
name	varchar(255)
stripe_customer_id	varchar(255)
monitoring_active	boolean	For portfolio monitor subscribers

🔌 API Design

Orders

POST /api/orders

Create new deal research order. Body: {company_name, founder_names[], industry, stage, tier, pitch_deck_base64?}. Creates Stripe checkout session. Returns order ID.

GET /api/orders/:id

Get order status and details.

GET /api/orders/:id/report

Get generated report (PDF URL + section previews).

Research (Internal)

POST /internal/research/deal/:id

Trigger full research pipeline for a deal. Async via QStash. Returns job ID.

GET /internal/research/deal/:id/status

Check research pipeline status per section.

POST /internal/outreach/start/:dealId

Start customer outreach for deep-dive orders. Triggers email sequence to 5 customers.

Portfolio Monitoring

GET /api/portfolio

List monitored companies for subscriber.

POST /api/portfolio/company

Add company to portfolio monitoring.

GET /api/portfolio/:companyId/alerts

Get recent news alerts and signals for a monitored company.

🤖 Research Agent Pipeline

Step 1: Founder Background Check

Agent takes founder names → searches LinkedIn (via PhantomBuster/Airtable), Crunchbase, Twitter/X, Google News. Compiles: employment timeline, education, funding history, news mentions. Identifies red flags: timeline gaps, credential discrepancies, negative news. Drafts 1-page founder bio. Runs on Ollama Mistral 7B on Mac Mini.

Step 2: Market Size Analysis

Agent queries IBISWorld/Statista for industry data. Builds TAM/SAM/SOM model. Calculates 3-year projections using CAGR from public comps. MiniMax API for final narrative polish.

Step 3: Competitive Landscape

Agent searches G2, Capterra, LinkedIn for direct/indirect competitors. Maps positioning matrix. Extracts pricing from public sources. MiniMax API for competitive narrative.

Step 4: Customer Interview Synthesis (Deep-Dive)

For each of 5 customers: agent sends outreach email via Hunter.io → schedules call → takes notes (manual or transcribed via Whisper) → Ollama synthesizes patterns across all 5 calls. Highlights: what users love, what they wish was different, why they chose this over alternatives.

Step 5: Financial Projection Sanity Check

Agent parses pitch deck numbers (or uploaded financial model). Benchmarks against public SaaS comparables at similar stage (from MicroCapCampbell, OpenView reports). Generates 3 scenarios: bear/base/bull. Identifies top 3 most aggressive assumptions.

Step 6: Report Assembly + PDF

All sections compiled into report HTML template. Puppeteer renders HTML → PDF. Uploaded to R2. Link shared with investor via email. MiniMax API writes executive summary and red flags sections.

🚀 Deployment

Component	Where	Notes
Dashboard (SvelteKit)	Cloudflare Pages (free)	Auto-deploys from Git
Backend API (Deno)	Mac Mini (local) or Deno Deploy	Runs research pipeline. Mac Mini is always-on server.
PDF Generation	Mac Mini + Puppeteer	Headless Chromium. Can run on same Mac Mini.
Database	Supabase (free tier: 500MB)	Sufficient for <100 active reports.
File Storage	Cloudflare R2 (free tier)	Pitch decks + PDF reports.
AI (local)	Mac Mini Ollama (Mistral 7B)	Research synthesis. Free.
AI (cloud)	MiniMax API	Report polish. Pay per report ~$0.10.

← Requirements | All Projects | Presentation →