VC Pitch Screening Agent

The Problem

Early-stage VCs review 2,500–5,000 pitch decks to make just 5–10 investments. That's a brutal funnel — and most of the filtering happens in minutes, by tired analysts working through a pile.

The cost of getting it wrong goes both ways:

False negatives — great startups buried in the pile. Missed unicorns that define fund returns.
False positives — "vibe-heavy" decisions that drain time and capital on weak fundamentals.

Solo GPs and emerging fund managers feel this most acutely. They're running lean teams with deal flow that rivals much larger funds, and every hour spent screening a bad deck is an hour not spent with a founder who deserves attention.

What We Built

A multi-agent AI pipeline that automates the initial pitch deck screening workflow — from raw PDF upload to a structured investment memo with a PASS / REVIEW / ARCHIVE decision.

The system handles the full workflow that currently takes analysts 1–3 days, and produces a consistent, thesis-aligned output every time.

Agent Architecture

The pipeline uses four agents running in a coordinated sequence:

Agent	Role
Agent 1 — Parser	Extracts content from pitch decks (text + vision fallback)
Agent 2 — Analyst	Evaluates thesis fit, team, market, and deal-breaker flags
Agent 3 — Fact-Checker	Cross-references startup claims against trusted web sources
Agent 4 — Memo Writer	Synthesises findings into a structured investment memo

Agents 2 and 3 run in parallel to minimise processing time. A post-processing decision engine then applies Agent 2's deal-breaker vetoes to Agent 3's output before the final memo is generated.

Technical Challenges Solved

Image-heavy PDFs — Most pitch decks are design-tool exports, not text-native PDFs. We built a two-pass parser: first attempts pdfplumber extraction. If a slide returns fewer than 50 characters, it's flagged as image-based and routed to Claude's Vision API to read the slide visually.

Fact-check noise — Open web search returns opinion-based, low-quality results that can incorrectly contradict legitimate startup claims. We domain-whitelisted Tavily to trusted sources only and introduced a 4-level verification status: verified / contradicted / unverified / not_found.

LLM non-determinism — Claude uses probabilistic sampling by default, meaning the same deck uploaded twice could produce different scores. We set temperature=0 across all agents and added file-hash deduplication — if the same PDF is uploaded again, the pipeline is skipped and the cached result returned instantly.

Parallel agent coordination — Agent 3's REVIEW/PASS/ARCHIVE decision needed to be overridden by Agent 2's deal-breaker flags. We resolved this with a post-processing engine that applies the veto logic after both agents complete.

Target Market

We designed the go-to-market around the highest-pain, fastest-adoption segments first:

Emerging fund managers & solo GPs — Small teams, high deal flow, most acutely feel the bandwidth problem. Primary target.
Mid-sized funds ($50M–$500M AUM) — Use as a triage supplement before human review.
Accelerators & incubators — Y Combinator, Antler, 500 Global review hundreds of applications per cohort. Same bottleneck, different context.
Corporate venture arms — Process-driven and open to tooling (e.g. Singtel Innov8).

Competitive Positioning

Two categories of competition — neither solves the full workflow:

Direct AI-native competitors (Hebbia, Harmonic, DeckMatch) — Powerful analysis but not purpose-built for end-to-end VC screening.

Data giants adding AI features (PitchBook, Affinity) — Rich data and distribution, but AI features are shallow additions to prevent platform leakage, not native to the decision workflow.

Our gap: the only tool built exclusively for AI-native screening end-to-end — from raw deck to investment memo.

Impact

Stage	Before	After
Initial Screening	Manual, inconsistent, high volume	Automated, thesis-aligned, consistent
First Meetings	Low context, noise-heavy	Higher quality, better prepared
Internal Diligence	Analysts spread thin	Time focused on the right deals
Partnership Meeting	Variable conviction	Higher conviction, better use of partner time

What's Next

Deeper thesis learning — Model learns from past memos and investment decisions to capture implicit preferences
Multi-user collaboration — Role-based access (Partner / Associate / Analyst), deal annotations, @mention teammates
VDR integration for DD — Expand from pitch deck to full data room documents (financials, cap table, GTM contracts)
CRM-style deal pipeline — Track deals across stages with personalised founder outreach
Proactive deal sourcing — Flag thesis-matching companies before they formally raise, via LinkedIn hiring signals and Crunchbase tracking