# Vela Partners

> Vela Partners is an AI-native quant VC.
> We fund AI founders in days, not months, with $100K to $500K checks
> and soft commits in 2 to 7 days.

## Core pages
- [Home](https://vela.partners): Firm overview, thesis, and investment process.
- [About](https://vela.partners/about): Our vision, the four pillars of the firm (quant, AI-native, product-led, scientific), and the research program.
- [Research](https://vela.partners/research): The flagship research portfolio. Four programs on LLMs with verifiable methods for high-stakes decisions, plus probabilistic portfolio construction.
- [Contributors](https://vela.partners/contributors): The 90+ people who have built Vela since 2017. The core team, committees, advisors, fellows, and the University of Oxford research community. Vela is a quant VC team sport, not a partner-brand vehicle.
- [Portfolio](https://vela.partners/portfolio): Vela is the world's first AI-only venture capital fund, before any other VC, with top-decile performance. AI portfolio across three funds since 2017. 50+ companies grouped into four thematic buckets: agentic AI and autonomous systems, foundation models and scientific AI, AI for industry, and AI-native infrastructure and data. 10+ exits including Limitless (Meta), Promoted (Dropbox), Bear Robotics (LG), Golden (ComplyAdvantage), WaveOne (Apple), Vartana (Capchase), OwnId (Unico), LeakSignal (F5), and Jobox (Talus Pay).
- [Products](https://vela.partners/products): Vela productizes venture capital. Products span four audiences. For Vela and its LPs: Vela OS (the operating system Vela built to run its own firm, also used by major LPs to invest directly). For aspiring founders: Founder Readiness (psychometric self-assessment based on Paul Graham's Founder Mode at enterfoundermode.com), AI Theses (live map of investable AI startup theses, derived from labor-market data, at thesisradar.com), Entrepreneur OS (Vela investment decision in minutes). For founders building: Vela Coach (open-source coaching from meeting transcripts at coach.vela.partners), Cairn (open-source end-to-end test agent for AI-generated code at cairn.vela.partners). For breakthrough research: Scientific Research Agents (the engine behind Vela's Oxford research). Every product for founders is free or open-source.
- [GPTree](https://vela.partners/research/gptree): LLM-powered decision trees for explainable founder-success prediction. The foundational Think-Reason-Learn paper. arXiv:2411.08257.
- [Random Rule Forest](https://vela.partners/research/random-rule-forest): Interpretable LLM-generated YES/NO question ensembles. 13.1% precision (6.9x over random), 15.3% with expert-informed questions. ICAIF 2026.
- [Reasoned Rule Mining](https://vela.partners/research/reasoned-rule-mining): Calibrated LLM classification for quant VC. 24.5% precision at 2% base rate (12.3x uplift). IEEE CSCloud / SecureFinAI 2025.
- [Policy Induction](https://vela.partners/research/policy-induction): Editable natural-language policies for quant VC. On VCBench: 67% precision (7.4x base rate, 3x tier-1 VC firms), F0.5 = 34% top baseline (IEEE CSCloud'25, New York). On Vela's non-anonymized founder dataset: 40.5% precision on 1.96% base rate, a 20x lift and 7.1x tier-1 VC firms (arXiv:2505.21427).
- [SSFF / Multi-Agent V](https://vela.partners/research/multi-agent-v): The first multi-agent framework for startup evaluation and the research origin of V, Vela's quant VC master agent. 67.8 weighted F1 on a 10% prevalence test set (3.7x vanilla GPT-4o-mini). Best Poster Award, NeurIPS 2025 Workshop on Generative AI for Finance. arXiv:2405.19456.
- [Founder-GPT](https://vela.partners/research/founder-gpt): Self-play and tree-of-thought for quant VC founder evaluation. The methodological root of Vela's Multi-Agent Systems research. Introduced explicit founder-idea fit scoring, simulated VC analyst self-play, and critique-based refinement. arXiv:2312.12037, December 2023.
- [VCBench](https://vela.partners/research/vcbench): The first AGI benchmark for venture capital. 9,000 anonymized founder profiles, F0.5 metric, public leaderboard at vcbench.com. Vanilla GPT-4o reaches 29.1% precision (3.2x the index, above Y Combinator and tier-1 VC firms). arXiv:2509.14448.
- [LLM-AR](https://vela.partners/research/llm-ar): Neural-symbolic reasoning for quant VC founder evaluation. LLM proposes rules, ProbLog executes them deterministically. 59.5% precision (5.9x random) on a 4-fold 6,000-founder evaluation, beats every vanilla LLM tested. arXiv:2510.22034.
- [GPT-HTree](https://vela.partners/research/gpt-htree): Hierarchical clustering and LLM personas for quant VC. Eight founder archetypes with 22x top-to-bottom success spread; best cluster 9x the 1.9% baseline. CTGAN resampling separates VC-experience subclusters from 3.9% to 50%. arXiv:2501.13743.
- [Rare-Event Prediction](https://vela.partners/research/rare-event-prediction): LLM-powered feature engineering combined with an XGBoost + Random Forest + Linear Regression ensemble for quant VC founder-success prediction. 10.3x precision at 36% recall on a 10,825-founder dataset; removing LLM features halves the model (4.6x). arXiv:2509.08140.
- [Verifiable Reasoning](https://vela.partners/research/verifiable-reasoning): LLMs as deterministic code generators for quant VC founder screening. One LLM call per iteration writes Python predicates that run across all founders. 99% cost reduction vs per-sample LLM evaluation. 37.5% precision on VCBench (4.17x random, beats GPT-4o). arXiv:2603.13287.
- [Learning What to Ask](https://vela.partners/research/cost-aware-sequential-evaluation): Cost-aware sequential founder evaluation for quant VC. RL policy decides which attribute to query next and when to stop; LLM supervises under uncertainty. F0.5 = 37.1% on VCBench (highest reported), using only 54% of the available founder information. Beats Policy Induction by +3.1 F0.5 points. MIDAS 2026 Workshop, ECML-PKDD 2026.
- [When Career Data Runs Out](https://vela.partners/research/structured-features-signal-limits): Engineered career features beat asking an LLM directly, then hit a wall: most successful founders look like failures on paper.
- [Portfolio Outliers](https://vela.partners/research/portfolio-construction): Latent-factor probabilistic model for quant VC portfolio construction. Adapts Vasicek / Koyluoglu credit-risk machinery to venture. Computes P(zero unicorns) accounting for sector, geography, and founder-type correlation. For a 40-deal portfolio of identical 4% deals, P(zero unicorns) rises from 19.6% under independence to 32.5% under full correlation. arXiv:2602.07761.
- [Beyond Picking Winners: Correlation-Driven Tail Risk](https://vela.partners/research/correlation-tail-risk): Gaussian-copula model for quant VC portfolio construction. Holding deal-level success probabilities constant, sweeping inter-deal correlation thickens both tails of the fund return distribution and leaves expected return unchanged. Two GPs with identical pick quality can produce wildly different return distributions if their correlation structures differ. Companion paper to Portfolio Outliers. arXiv:2604.23087.
- [CoFEE: Reasoning Control for LLM-Based Feature Discovery](https://vela.partners/research/cofee): Cognitive scaffolding for LLM feature engineering. Enforces backward chaining from outcomes, subgoal decomposition, leakage and observability verification, and explicit backtracking. Compared to vanilla prompting: 15.2% higher success, 29% fewer features, 53.3% reduction in cost. The LLM-era counterpart to Featuretools and autofeat for unstructured inputs. Provider-agnostic (GPT-4, Claude, Gemini). Upstream feature engineering layer for the Think-Reason-Learn family. arXiv:2604.21584.
- [Think-Reason-Learn](https://vela.partners/research/think-reason-learn): Vela's open-source LLM-native ML framework for quant VC. Scikit-learn with reasoning built in. MIT-licensed Python library. Includes GPTree, Random Rule Forest, Reasoned Rule Mining, and Policy Induction under a unified interface. Install with "pip install think-reason-learn". Docs at thinkreasonlearn.com.
- [Entrepreneur OS](https://vela.partners/entrepreneur-os): AI-powered startup evaluation. Get an investment decision in minutes.
- [Startups assessed by Vela](https://vela.partners/entrepreneur-os/startups): Companies Vela has researched and assessed: what each builds, who is behind it, and what the public record shows. Every company listed asked to be.
- [Founder Mode](https://enterfoundermode.com/): Free founder psychometrics tool. Sixteen founder archetypes.
- [Blog](https://vela.partners/blog): Technical writing on AI agents, graph databases, and quantitative VC.
- [Contact](https://vela.partners/contact): Write to Vela. Founders, investors and researchers; replies come from the team.

## Research projects
- [Benchmarks: VCBench](https://vela.partners/research/benchmarks): The first AGI benchmark for venture capital. 9,000 anonymized founder profiles, F0.5 as the primary metric, public leaderboard at vcbench.com. Vanilla GPT-4o reaches 29.1% precision, 3.2x the index.
- [Simulation: GP-Bench](https://vela.partners/research/simulation): In development. A long-horizon environment in which an AI agent runs a $50M pre-seed fund for 40 quarters, with deal flow, costly information, follow-ons and exits. Every episode is generated fresh, so nothing can be memorized.
- [Harness: VCRouter](https://vela.partners/research/harness): Model router for venture tasks, releasing alongside VC Arena. One API call sends any venture task to whichever model currently leads it in the arena, and blends across models. vcrouter.com.
- [Predictions: Think-Reason-Learn](https://vela.partners/research/predictions): Vela's open-source, MIT-licensed library of LLM-native prediction methods (GPTree, Random Rule Forest, Reasoned Rule Mining, Policy Induction) with the fit / predict pattern of scikit-learn. thinkreasonlearn.com.

## Featured blog posts
- [What is a Quant VC?](https://vela.partners/blog/what-is-a-quant-vc): Plain-English guide to quantitative, AI-native, and scientific venture capital. Reference post for the quant VC term.
- [AI Coach for Founders](https://vela.partners/blog/ai-coach-for-founders): Vela Coach is the open-source AI coach for founders. Reads the last 90 days of meeting transcripts and surfaces leadership, fundraising, and team-dynamics patterns a $500-an-hour executive coach would surface after a month. Browser-only, free, self-hostable. The end-of-Claude-Code-sprint pre-pitch ritual.
- [How Founders Use an AI Coach](https://vela.partners/blog/how-founders-use-ai-coach): Field guide to using Vela Coach beyond pitch prep. Specific patterns the AI surfaces in sales calls, interview rounds, and weekly one-on-ones, with the recommended weekly cadence and event triggers.
- [Think-Reason-Learn Launches FounderBrain](https://vela.partners/blog/founderbrain-launch): FounderBrain, the first pre-trained model from Think-Reason-Learn (Vela Research and the University of Oxford), is a free MCP server that gives founders feedback on how they speak in pitches and meetings. It compares a meeting transcript with 773 public interviews of founders, CEOs and operators and returns a speaking type, four speaking styles, the well-known founders who speak most like you and three takeaways. Works in Claude, ChatGPT, Claude Code and Codex at https://mcp.thinkreasonlearn.com/mcp. Free, no account, transcripts not stored.
- [Contrarian vs Consensus in Venture Capital](https://vela.partners/blog/contrarian-vs-consensus-venture-capital): Most legendary venture deals (OpenAI 2015, Wiz at seed, Anthropic at seed) were not contrarian. They were access bets. The right framework is not contrarian versus consensus; it is expected value, calculated honestly: check size x expected multiple x probability of success.
- [Emerging Open-Source AI Infrastructure Trends 2026](https://vela.partners/blog/emerging-open-source-ai-infrastructure-trends-2026): 19 open-source AI infrastructure trends identified from 854 GitHub repositories with 90-day momentum.
- [KuzuDB AI Agent Memory](https://vela.partners/blog/kuzudb-ai-agent-memory-graph-database): Vela's fork of KuzuDB with concurrent multi-writer support for multi-agent AI workloads. 374x faster path queries than Neo4j.

## Research and open source
- [VCBench](https://vcbench.com): World's first AGI benchmark for venture capital. 9,000 anonymized founder profiles.
- [Think-Reason-Learn](https://thinkreasonlearn.com): Open-source LLM-native ML framework. GitHub at vela-research/think-reason-learn.
- Multi-agent framework that became V: https://arxiv.org/abs/2405.19456 (Best Poster, NeurIPS 2025 Workshop on Generative AI in Finance).
- Probabilistic portfolio modeling for VC: https://arxiv.org/abs/2602.07761.

## Context
- [Full firm context (natural language)](https://vela.partners/llms-full.txt)

## Sitemaps
- [XML sitemap](https://vela.partners/sitemap.xml)