Overview
Investment analysts spend 3–5 hours reviewing each inbound pitch deck — time that could be spent on higher-value work. This concept demonstrates a system that performs initial triage on pitch documents using semantic analysis, scoring each submission against a configurable investment thesis, and surfacing the top-ranked submissions for human review. The analyst still makes every decision — the AI handles the reading.
The Problem
The challenge is not reading comprehension — modern LLMs handle that well. The challenge is evaluation consistency. Two analysts reviewing the same deck will sc…
The Approach
The system uses a two-stage pipeline: (1) Extraction — parse the PDF, extract structured data (team, market, traction, technology, competition) using a combinat…
The Outcome
The primary performance metric was accuracy — does the system's assessment align with the analyst's assessment on the same document? In a retrospective test on …
The Challenge
The challenge is not reading comprehension — modern LLMs handle that well. The challenge is evaluation consistency. Two analysts reviewing the same deck will score it differently based on their current mood, recent portfolio decisions, and individual cognitive biases. The system must encode the firm's actual investment thesis in a way that produces consistent, explainable scoring — not a black box number.
Research
Research involved mapping the firm's current evaluation criteria: market size, team quality, technology differentiation, traction, and competitive moat. Each criterion was decomposed into specific signals that can be extracted from a pitch document. The research also identified what pitch decks systematically omit (unit economics, churn, competitive analysis), which became mandatory data points to flag as missing.
Strategy
The system uses a two-stage pipeline: (1) Extraction — parse the PDF, extract structured data (team, market, traction, technology, competition) using a combination of GPT-4o structured outputs and traditional NLP for numerical data, (2) Scoring — apply the firm's thesis criteria against the extracted data, producing a score per dimension with explicit citations to the source document. Analysts see both the score and the exact text that generated it.
Design
The interface is deliberately minimal. The goal is to give an analyst enough information to decide "worth reviewing" or "pass" in under 90 seconds per deck. The submission view shows: overall thesis alignment score, per-dimension breakdown, extracted key facts (team, market size, traction), and highlighted text from the source document for each score dimension. One click opens the full PDF in a side panel.
Development
Built in Next.js with a Python FastAPI backend for the AI pipeline. PDFs are processed asynchronously via a queue (Redis + BullMQ). The AI pipeline uses LangChain for document parsing, OpenAI embeddings for semantic search within documents, and GPT-4o structured outputs for criterion extraction. All extracted data is stored in PostgreSQL with full provenance — which page and paragraph generated which data point.
Performance
The primary performance metric was accuracy — does the system's assessment align with the analyst's assessment on the same document? In a retrospective test on historical deal flow, alignment on "pass" decisions was 94%. Alignment on "review" decisions was 78% — lower, because the system uses conservative thresholds to avoid missing potentially good deals.
Architecture Overview
PDF submission → S3 storage → BullMQ queue → FastAPI worker (LangChain + GPT-4o) → structured extraction → PostgreSQL (scores + provenance) → Next.js dashboard (real-time via WebSocket for queue status). Analyst makes every final decision — AI provides ranked context.
SEO Strategy
The application is private (not indexed). The marketing site explaining the service targets keywords around "deal flow automation", "VC AI tools", and "pitch deck analysis software". Technical content (how LLM structured outputs work in document processing) builds topical authority through a blog.
Accessibility Notes
The analyst dashboard is a complex application with data tables and document viewers. All table headers are associated with data cells. Sort controls announce state changes via aria-live. The PDF viewer uses an accessible iframe wrapper. Keyboard shortcuts are documented and announced to screen readers.
Performance Optimisations
PDF processing is asynchronous and does not block the UI. Queue status is streamed via WebSocket — no polling. The analyst dashboard uses React Suspense boundaries to load the scored deck list immediately while individual deck details load on demand. Aggressive browser caching on all extracted data (immutable once processed).
Lessons Learned
The highest-value investment in the system was not the AI pipeline — it was the investment thesis encoding. Making the firm articulate exactly what they look for (and in what order of priority) was a 4-week process that surfaced significant disagreement within the team. The system forced a conversation that produced alignment on investment criteria that the team had never formally documented.
Future Improvements
Benchmarking extracted data against public comparables (Crunchbase, PitchBook API) to validate claimed metrics. Portfolio fit analysis that scores how a deal would complement (or cannibilise) existing portfolio companies. Automated founder research across LinkedIn, GitHub, and news sources to augment the pitch deck data.