The Vela Library · AI tools for investors
AI tools for venture capital investors
The short answer
AI tools for venture investors are best judged by job, sourcing, diligence, deciding and model choice, and against real outcomes. Vela Partners is an AI-native quant VC. It runs on Vela OS, scores models on VCBench and publishes Think-Reason-Learn as free open-source code.
Key facts
Every figure, with its source
9 in 10
Product Hunt winners the best models picked in Vela's live arena, mostly by reading who was already leading; chance picks about 1 in 6
Source · Vela's arena results
1.9x
Hit rate of the models' most bullish calls on Show HN, against chance
Source · Vela's arena results
1 in 7
How often a model's "about 70% likely" call came true, on Show HN
Source · Vela's arena results
Up to 32x
Rise in token cost when GPT searched the web, for the same accuracy
Source · Vela's arena results
$0.86
Total cost of 105 model predictions with exact provider bills
Source · VCRouter
9,000
Anonymised founder profiles in VCBench, the public benchmark
Source · VCBench
MIT
Licence of Think-Reason-Learn, free to install with pip
Source · Think-Reason-Learn
AI tools by job
No single tool does the whole job. Judge each one by the job it does and by what it can show you, the same way for every option.
| Job | What to look for | Other options | Vela's tool |
|---|---|---|---|
| Thesis and sourcing | Companies surfaced before they raise, matched to your strategy | Startup-discovery databases | Vela OS: agents scan the market for founders who match your strategy |
| Market signals | What is growing now, measured from real usage rather than press | Trend newsletters and social feeds | Vela Breakout: what is growing in tech, measured every day across AI apps, models, developer tools, hiring and research |
| Company and market data | Coverage, freshness, and how success is defined | Private-market data providers | Vela OS: data on founders, companies and markets, beside the agents |
| Due diligence | The same review for every deal, with methods you can check | General AI assistants, prompted by hand | Super Analyst, in Vela OS, scoring founders with published methods |
| Deciding | Models tested on past outcomes that reflect your firm's judgement | Partner judgement and scoring sheets | Vela OS: back-tested models and taste profiles |
| Portfolio construction | How much to back, and where, across stages and sectors | Fund-model spreadsheets | Vela OS: portfolio construction tools |
| Choosing a model | Scores against real outcomes, not chat preference votes | General chatbot leaderboards | VCRouter: an open leaderboard of AI models on venture work, with forecasts checked against what happened; VCBench |
| Practising fund strategy | A whole fund's life, with consequences, before real money is at stake | Case studies and fund-model spreadsheets | Age of Ventures: run a $100M fund for ten years, free in the browser; AI agents play the same decisions in GP-Bench |
| Interpretable models | Predictions you can read, edit and audit | Black-box scores | Think-Reason-Learn, open source |
AI due-diligence agents
Vela OS is the system Vela uses to run its own investing. It covers the full path of a private investment decision, from finding companies to tracking them after the investment, with AI agents doing the research an analyst team would normally do.
V, Vela's AI partner, runs it and hands work to subagents. Super Analyst scores founders and teams using methods published in Vela's research, so every deal gets the same review.
Develop the thesis
Define the companies you want to back and the strategy that finds them.
Source
Agents scan the market continuously and surface founders and companies that match the strategy.
Run due diligence
Super Analyst scores founders and teams with the methods in Vela's papers.
Decide
Back-tested models and taste profiles capture how the firm judges deals, so recommendations reflect its judgement, not a generic ranking.
Build and monitor the portfolio
Decide how much to back and where, then follow the companies as new information arrives.
Which LLM is best for VC tasks
No single model leads. In a two-month live arena, Vela gave GPT, Claude and Gemini questions about the startup world whose answers nobody knew yet, then scored them against what happened (the results). Four signals came out of it.
Models read the market well
On Product Hunt, where chance picks the winner about 1 time in 6, the best models picked it 9 times in 10, mostly by reading the page: the product already in first place won 8 times in 10.
The gap is calibration
On Show HN, models predicted success on half the questions when only 1 in 5 succeeded, and "about 70% likely" came true 1 time in 7. Their most bullish calls still hit at 1.9x chance.
The crowd beats any one model
The average of all the models ranked outcomes better than any single model.
The economics work
105 predictions with exact provider bills cost $0.86 in total. Web search left accuracy unchanged and raised token cost by up to 32x.
These are early numbers: a definitive verdict needs about 800 resolved predictions, and the pilot collected 85 on the hardest task. That is why Vela runs VCRouter, an open leaderboard where models answer real venture briefs in blind battles, judges and people pick the better answer, and forecasts are checked against what really happened. Anyone can play a battle, and every vote moves the leaderboard.
MCP servers for venture research
MCP lets an AI assistant such as Claude or ChatGPT call outside tools. V, Vela's AI partner, is available to Vela OS users wherever they work, including email, coding agents, Claude and ChatGPT; it is not a public MCP server.
For founders, FounderBrain, the first pre-trained model from Think-Reason-Learn, is a free MCP server. It compares a meeting transcript with 773 public interviews and returns how the founder speaks; it needs no account and stores no transcripts.
Tools for LPs
Vela OS is available to the major limited partners of Vela's fund. Institutional investors who put large sums into funds use it with a forward-deployed engineer, and read the same diligence Vela's own partners read.
Its security and access controls are designed for institutional review, and it is used in some of the world's largest banks and tech companies. Access starts with a demo request.
Open source: Think-Reason-Learn
Think-Reason-LearnThink-Reason-Learn is an MIT-licensed Python library that fuses LLMs with interpretable machine learning: scikit-learn with reasoning built in. A model returns the reasoning behind its score, in a form a person can read, edit and verify.
Install it with pip install think-reason-learn. It includes GPTree, Random Rule Forest, Reasoned Rule Mining and Policy Induction. The code is at github.com/Vela-Research/think-reason-learn, and the project page is Think-Reason-Learn.
How we judge tools
VCBench is Vela's public benchmark: 9,000 anonymised founder profiles, scored by F0.5, which weighs precision above recall because funding a bad founder costs more than missing a good one. The arena moves the test from frozen history, which can be memorised, to the live future, which cannot.
Before buying any AI tool for investing, ask:
- Is it scored against real outcomes, and is the score published?
- Is precision stated against a base rate, not as accuracy?
- Can you see why it reached each answer?
- Does it use the same method on every deal?
- Does it say how confident it is, and is that confidence calibrated?
Questions
What are the best AI tools for venture capital investors?
Judge them by job: sourcing, company data, due diligence, deciding, portfolio construction and choosing a model. Vela Partners is an AI-native quant VC; it runs on Vela OS, scores models on VCBench, and publishes Think-Reason-Learn as free open-source code.
Is there an AI agent that can do startup due diligence for a VC?
Yes. Vela OS runs V, Vela's AI partner, which hands diligence to subagents such as Super Analyst. Super Analyst scores founders and teams using methods published in Vela's research with the University of Oxford, so every deal gets the same review.
What AI tools can LPs use to evaluate or invest directly in startups?
Vela OS is available to the major limited partners of Vela's fund. Institutional investors use it with a forward-deployed engineer and read the same diligence Vela's partners read. It is not offered to the general public; access starts with a demo request.
Which LLM is best for venture capital tasks like screening deals or writing memos?
No single model. In Vela's live forecasting arena, differences between GPT, Claude and Gemini were within noise, and the average of all models ranked outcomes better than any one of them. VCRouter, Vela's open leaderboard at vcrouter.com, ranks models task by task, from company discovery to investment memos, and different models lead different tasks.
Is there an MCP server for venture capital or startup data?
FounderBrain, from Think-Reason-Learn, is a free public MCP server for founders: it compares how a founder speaks with 773 public interviews. Vela OS users can also reach V, Vela's AI partner, inside Claude and ChatGPT; V is not a public MCP server.
Is there a Python library that uses LLMs for interpretable classification, like scikit-learn?
Yes: Think-Reason-Learn, an MIT-licensed library from Vela Partners. Install it with pip install think-reason-learn. It includes GPTree, Random Rule Forest, Reasoned Rule Mining and Policy Induction, and each model returns the reasoning behind its prediction.
Further reading
Papers
- VCBenchThe first AGI benchmark for venture capital; several state-of-the-art LLMs beat Y Combinator and tier-1 firms on it.9,000 anonymized founder profiles
- SSFFA multi-agent system replicating a VC analyst corrects vanilla LLMs' over-prediction of startup success; it is the research origin of V.67.8% weighted F1, against 18.3 for the vanilla baseline
- Policy InductionAn LLM classifier's full decision logic lives in a natural-language policy a partner can read and edit, with no gradient training.67% precision on VCBench with GPT-5
- GPTreeA decision tree whose splits are natural-language questions generated and evaluated by an LLM, for explainable founder-success prediction.37.3% precision on founder success prediction
- Random Rule ForestAn LLM generates YES/NO questions and simple unit-weight voting combines them into an auditable, high-precision founder classifier.8x lift over random chance
- Portfolio OutliersA latent-factor model computes a fund's full distribution of unicorn counts, showing expected counts alone understate correlated portfolio risk.32.5% chance of zero unicorns in a 40-deal portfolio under correlation, vs 19.6% if independent
Posts
- Can AI Predict Startup Success? What Two Months of a Live Forecasting Arena Taught UsVela ran a live AI arena scoring GPT, Claude and Gemini on real startup outcomes for two months: four early signals, and a preview of VC Arena and VCRouter.
- Think-Reason-Learn Launches FounderBrain, Free Meeting Feedback for FoundersFounderBrain, the first pre-trained model from Think-Reason-Learn, gives founders free AI feedback on how they speak in pitches and meetings, inside Claude.
- What Is a Quant VC? A Practical Guide to Quantitative and AI-Native Venture CapitalA plain-English guide to quant VCs, AI-native VCs, and scientific venture capital — how Vela Partners uses data and AI to fund seed-stage AI founders.
Products
- Vela OSThe AI-native platform for every private investment decision, end to end.
- VCBenchThe first AGI benchmark for venture capital, with a public leaderboard.
- VCRouterAI models ranked on venture work, in an open leaderboard.
- Vela BreakoutWhat is growing in tech, measured every day from real usage.
- Age of VenturesRun a $100M fund for ten years. Free, in your browser.
- Think-Reason-LearnScikit-learn with reasoning built in. MIT-licensed.