INSEAD MBA · Generative AI for Business · Group Project
VC Pitch Screening Agent
The Problem
Early-stage VCs screen 2,500–5,000 pitch decks to make just 5–10 investments per year — a brutal funnel where most decisions happen in minutes, by tired analysts working through a pile.
The cost of getting it wrong goes both ways:
- False negatives (missed winners) — Great startups get overlooked due to analyst fatigue, noise, or a weak deck design. Missing the next unicorn can define a fund's returns.
- False positives (wasted time / bad bets) — "Vibe-heavy" decisions that skip weak fundamentals and red flags drain time and capital.
Three structural forces make this worse:
- Due diligence demand — More founders than ever are pitching, and solo GPs are operating with even leaner teams.
- Asymmetric risk — One missed unicorn or one bad investment can define a fund's returns.
- Inconsistent judgement — Human analysts apply different standards to deck #1 vs deck #47 on a Friday afternoon. An agent doesn't.
The VC process has 6 stages: Initial Screening → First Meeting/Call → Internal Due Diligence → Partnership Meeting → Term Sheet Negotiation → Final Diligence & Legal Closing. We focused exclusively on Stage 1 — where most decks are rejected, and where consistent, thesis-aligned filtering has the highest leverage.
What We Built
Insiders Den — 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.
A VC analyst uploads a pitch deck PDF. The fund's investment thesis is pre-configured (target sectors, geographies, stage, ARR targets, and deal-breaker rules). The pipeline does the rest in minutes.
Tech Stack
| Layer | Tool | Hosting |
|---|---|---|
| Frontend | Next.js | Vercel |
| Backend | FastAPI (Python) | Render |
| Database | PostgreSQL | Supabase |
| Agents 1–3 | Claude Sonnet 4.6 | Anthropic API |
| Agent 4 | Claude Opus 4.6 | Anthropic API |
| Web Search | Tavily API | — |
| Resend API | — |
The 4-Agent Pipeline
Agent 1 — Claim Extractor (Claude Sonnet 4.6)
Reads every slide of the pitch deck and extracts structured data — TAM, ARR, team background, funding stage, geography, product description, and customer traction. Flags fields where information is missing or ambiguous.
Output: ClaimsJSON — a structured object passed to Agents 2 and 3.
Agent 2 — Fact Checker (Claude Sonnet 4.6, runs in parallel)
Takes ClaimsJSON and searches trusted external sources (Crunchbase, Statista, industry databases) to validate the startup's claims. Assigns each claim one of four verification statuses:
verified— corroborated by a trusted sourcecontradicted— directly contradicted by evidenceunverified— searched but no relevant result foundnot_found— insufficient data to assess
Also detects the three hard deal-breakers configured into the fund thesis:
- No pre-product / pre-revenue
- Hardware business model
- D2C / consumer products
Output: FactCheckResult — claim statuses + deal-breaker flags. Agent 2 holds veto power over Agent 3.
Agent 3 — Thesis Scorer (Claude Sonnet 4.6, runs in parallel)
Compares ClaimsJSON against the fund's pre-configured investment thesis across four scored dimensions:
| Dimension | What it measures |
|---|---|
| Sector Fit | Does the business model match target sectors (B2B SaaS, Enterprise, Deep Tech)? |
| Geography Fit | Is HQ and primary traction in target markets (Southeast Asia)? |
| Stage Fit | Is the round Pre-seed, Seed, or Series A/B as expected? |
| Financial Traction (ARR) | Does ARR evidence support a credible path to $1M ARR in 12 months? |
Each dimension scored 0–100%. Confidence levels adjust based on missing data points (0–1 missing: High confidence; 2–3: Medium; 4+: Low).
Bonus points awarded for: prior institutional/renowned VC backers, clear path to profitability within the round, and repeat/serial founders.
Produces a preliminary REVIEW / PASS / ARCHIVE decision — subject to Agent 2's veto.
Output: ThesisFitResult — dimension scores, overall fit %, and preliminary decision.
Decision Engine
After Agents 2 and 3 complete in parallel, a post-processing engine applies Agent 2's deal-breaker veto to Agent 3's output:
| Thesis Fit | Deal Breaker Triggered? | Final Decision |
|---|---|---|
| ≥ 50% | No | REVIEW |
| < 50% | No | PASS |
| ≥ 75% | Yes | ARCHIVE |
| < 75% | Yes | PASS |
Agent 4 — Memo Drafter (Claude Opus 4.6, triggered only for REVIEW)
Only fires when the Decision Engine returns REVIEW — avoiding unnecessary API cost on decks that won't proceed. Synthesises all prior outputs (ClaimsJSON + FactCheckResult + ThesisFitResult) and writes a full structured investment memo with sections covering Business Summary, Market, Team, Traction, Thesis Fit, Fact-Check Findings, and Recommendation.
Output: InvestmentMemo — displayed in the web app and emailed to the analyst via Resend API. All deals stored in PostgreSQL.
Technical Challenges
1. Image-Heavy PDFs
Most pitch decks are Keynote or Figma exports — not text-native PDFs. Standard text extraction returns empty strings for these slides, causing Agent 1 to miss critical data.
Mitigation: Built a two-pass parser. First attempts pdfplumber text extraction. If a slide returns fewer than 50 characters, it's flagged as image-based and routed to Claude's Vision API, which reads the slide visually.
2. Fact-Check Noise
Open web search returns opinion-based, low-quality results that can incorrectly contradict legitimate startup claims — creating false negatives on good deals.
Mitigation: Tavily search is domain-whitelisted to trusted sources only (Crunchbase, Statista, sector databases). Introduced the 4-level verification status system (verified / contradicted / unverified / not_found) to avoid treating "no data found" as a contradiction.
3. Parallel Agent Coordination
Agents 2 and 3 run in parallel to reduce processing time, but Agent 3's REVIEW/PASS/ARCHIVE decision needed to be overridden by Agent 2's deal-breaker flags — which wouldn't exist yet during parallel execution.
Mitigation: Agent 3 runs with an empty placeholder for deal-breakers during parallel execution. After both agents complete, the post-processing Decision Engine applies Agent 2's veto logic to produce the final decision.
4. LLM Non-Determinism
Claude uses probabilistic sampling by default — the same deck uploaded twice could produce different thesis fit scores, making the tool unreliable for repeated use.
Mitigation: Set temperature=0 across all four agents. Added file-hash deduplication — if the same PDF is uploaded again, the pipeline is skipped entirely and the cached result returned instantly.
Target Market
Primary focus: Southeast Asia — ecosystem is maturing, deal flow is growing fast, most funds are still small teams (bandwidth problem is acute), and the region is more open to new tooling.
Segment priority:
- Emerging fund managers & solo GPs — Highest pain, fastest adoption. Small teams running the same deal flow volume as larger funds.
- Mid-sized funds ($50M–$500M AUM) — Use as an analyst triage supplement before human review.
- Corporate venture arms — Process-driven and open to tooling (e.g. Singtel Innov8).
- Accelerators & incubators — Y Combinator, Antler, 500 Global review hundreds of applications per cohort. Same bottleneck, different context.
- Family offices moving into venture — Less experienced teams that need structured frameworks most.
- Angel investors — Very lean or no team. Maximum leverage per decision.
Competitive Landscape
Direct agentic competitors (Hebbia, Harmonic, DeckMatch) — powerful analysis capabilities, but not purpose-built for VC screening end-to-end. Strong at analysis; not optimised for the full workflow.
Data giants adding AI features (PitchBook, Crunchbase, AlphaSense, Affinity, DealCloud) — rich data and distribution, but AI features are shallow additions built to prevent platform leakage, not native to the decision workflow.
Our gap: The only tool built exclusively for AI-native deal screening end-to-end — from raw deck to investment memo, with thesis alignment built in from the start.
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 Due Diligence | Analysts spread thin | Time focused on the right deals |
| Partnership Meeting | Variable conviction | Higher conviction, better use of partner time |
| Term Sheet | Slower cycles | Stronger terms, faster negotiation |
| Final Closing | Surprises and delays | Smoother diligence, fewer surprises |
What's Next
- Deeper thesis learning — Model learns from past memos and investment decisions to capture implicit preferences the firm has developed over time — not just the explicit thesis document.
- Multi-user & team collaboration — Role-based access (Partner / Associate / Analyst), deal annotations, @mention teammates for second opinions.
- VDR check for DD assistant — Expand from pitch deck to full data room documents: business plans, financial models, GTM contracts, cap tables.
- CRM-style deal pipeline — Pipeline dashboard by stage, personalised founder outreach, and a feedback loop from users on memo quality.
- Proactive deal sourcing — Flag thesis-matching companies before they formally raise, by tracking LinkedIn hiring signals, Crunchbase activity, and PitchBook data.