Our Vision
Startup financing should work at the speed of a mortgage.
Home ownership sits near the base of Maslow's hierarchy, and it is genuinely remarkable that in much of the world today, anyone who works hard and pays their debts on time can own a home. A bank quantifies the risk using tangible assets, work history, tax returns, and credit history, then offers a mortgage at a rate that matches the borrower's profile. That is one of the most important problems humanity has ever solved at scale.
Self-actualization sits at the top of the same hierarchy, and for most people, entrepreneurship is the path to it. Building something of your own, on your own terms, is how people reach the highest version of themselves. Yet the financing that makes entrepreneurship possible is still gated, slow, and opaque.
There is a reason for that. Until recently, it was technologically impossible to quantify the things that actually matter in early-stage investing: the quality of a team, the shape of a market, the durability of a business model. When a company has no revenue, traditional finance has no anchor, so the industry defaulted to pattern matching, warm intros, and gut feel. That is not a flaw of individual investors. It is the ceiling of what was measurable.
Vela has spent years breaking through that ceiling. Our foundational research in AI for venture capital, much of it published with the University of Oxford, has made it possible to measure what used to be unmeasurable. We can now quantify a founder's probability of success before there is a product, score a market before there is traction, and assess a business model before there is revenue. In many cases, we can form a view when there is just a person and an idea, or even just a person.
That is the innovation that makes the rest of this vision possible.
Any founder, anywhere, should know where they stand with us within hours, not months. If the probability of success is low, we say so directly and tell them what would change the answer: reach $10K MRR, add a technical co-founder, prove demand in a specific market. If the probability is moderate, we partner at a lower valuation. If it is high, we partner at a higher valuation. And the partnership does not end with the investment decision. Paul Graham writes essays that help founders think. Andreessen Horowitz produces podcasts that help founders learn. We build products that help founders operate. Whether we invest capital or not, every founder we engage with gets free software that works alongside them every day, because we just want entrepreneurs to succeed.
Today, entrepreneurs waste two to six months convincing VCs to invest. They are interviewed, pulled into events, grilled in partner meetings, and often rejected with vague reasons. It is a broken user experience, and the incumbents have not fixed it. We will.
We exist to make entrepreneurial capital as accessible and as fast as mortgage capital, so that more people, in more places, can reach self-actualization through building.
What We Are
Vela Partners is a quant, AI-native, product-led, and scientific VC. These are not taglines. They describe how the firm is built.
Quant. We run probability models on every founder and company we evaluate. Our L1 to L5 founder quality system, market quality rating (VMQ), competition models, and traction signals are quantitative, reproducible, and backtested against thousands of outcomes. Decisions at Vela are scored, not argued. This is what a quant VC looks like in practice.
AI-Native. At the core of our firm is Vela OS, a VC operating system we built from scratch. V, our AI partner, is the master agent that runs Vela OS, orchestrating a team of specialized subagents across sourcing, diligence, portfolio construction, and founder support. Super Analyst is one of those subagents, running market, competition, traction, founder, and data room diligence in parallel. Vela OS is not a black box. Our major LPs are inside the system with us, using the same agents, reading the same diligence, and participating in the process in real time. We were built from day one as a software company that invests, not an investment firm that adopted software later. An AI-native VC is not a VC with AI tools bolted on. It is a VC where AI is the infrastructure.
Product-Led. We build products that entrepreneurs use, not just products that we use internally. Entrepreneur OS is V for founders: a self-service funding path where entrepreneurs can be evaluated, see their probability of success, and move forward in hours instead of months. Founder Mode is a free founder psychometrics tool that helps people decide whether they are ready to start a company at all, mapping them to one of sixteen founder archetypes with a full strategy report. Our products are open to everyone, whether we invest or not. A product-led VC is one whose products stand on their own.
Scientific. Every decision we make is grounded in peer-reviewed research. We do not rely on gut feel or pattern matching. We publish our methodology, open-source our models, and invite the world to audit our thinking, because we want other firms to adopt and copy our approach. The faster the industry becomes scientific, the faster entrepreneurial financing improves for everyone.
Relentless Focus on Scientific Research
Venture capital has produced more mythology than literature. Most of what passes for insight in the industry is anecdote: a famous founder, a legendary bet, a pattern someone noticed once and has been repeating ever since. That approach has served the winners well. It has not served entrepreneurs, LPs, or the truth.
We think capital allocation at this stage of the economy deserves a serious research program, so we built one.
For more than five years, Vela has run an active research partnership with the University of Oxford, producing over fifty papers, typically ten a year, along with a growing portfolio of patents on quantifying decision making in venture capital. The work is not abstract. Every model we publish eventually ends up inside Vela OS, evaluated against real deals and real outcomes. When a paper fails to improve our predictions, we retire it. When it succeeds, it becomes part of how we invest.
The headline number: the base rate for producing a unicorn in the US is about 1.9%. Our models consistently outperform that baseline by roughly 10x, reaching 19% precision on founder success prediction. In a category defined by rare-event outcomes, a 10x lift on the unicorn rate is the difference between a portfolio driven by luck and a portfolio driven by science.
Some of the contributions we are most proud of:
- The multi-agent architecture that became V. Our foundational paper on a multi-agent framework for startup evaluation won Best Poster at the NeurIPS 2025 Workshop on Generative AI in Finance. It is the research origin of V and the system that Vela OS runs today.
- Policy Induction (IEEE 2025). A memory-augmented, in-context learning framework that embeds natural-language policy directly into prompts, making the reasoning interpretable and editable. It reaches more than 20x the precision of random chance and 7.1x the precision of top-tier VC firms.
- Reasoned Rule Mining (Commendation Award, ICIM 2026, Oxford). A Bayesian, precision-optimized framework for founder evaluation that reaches 12.25x the market index in precision at 97.4% accuracy.
- GPTree (patent filed). LLM-powered decision trees for explainable founder success prediction. GPTree was the foundational work for Think-Reason-Learn and for every reasoning-based method we have built since.
- Random Rule Forest (provisional patent). An interpretable ensemble of LLM-generated YES/NO rules that beats zero- and few-shot LLM baselines by 41% relative F0.5 and delivers an 8x lift over random chance.
- VCBench. The world's first AGI benchmark for venture capital, with 9,000 anonymized founder profiles. State-of-the-art LLMs outperform Y Combinator and tier-1 firms on it. Available at vcbench.com.
- Portfolio construction. The worst thing that can happen to a venture fund is to hold zero unicorns across its entire portfolio. Our probabilistic portfolio modeling work adapts the risk frameworks used in bank loan portfolios to the outlier-driven dynamics of venture capital, quantifying individual outlier probability and inter-deal dependence in a unified model. The goal is to minimize the probability of the zero-unicorn scenario and to increase the probability of hitting a target number of outliers at the portfolio level.
Beyond specific papers, we invest heavily in the methods that make these results possible. We use reinforcement learning to train agents that evaluate founders sequentially, under the same constraints and information asymmetries a human investor faces. We build verifiable learning loops so that every model in Vela OS is measured against real-world outcomes as they unfold, not just against historical backtests. Models that drift are retrained. Models that stop working are retired. This is the infrastructure that keeps Vela OS honest.
The generalization of all of this is a framework we call Think-Reason-Learn. It is an open-source, LLM-native machine learning framework that treats reasoning itself as a learnable object, the way scikit-learn treats classifiers. Many of the papers above are implemented as modules inside it. Think of it as the next-gen scikit-learn for the class of problems that used to be considered unquantifiable: strategic choices, qualitative assessments, expert decisions. We built it for venture capital, but the framework generalizes to any domain where humans currently make high-stakes judgment calls with incomplete information. It is available at thinkreasonlearn.com and on GitHub.
All of it is public, or becomes public once the patents are filed. We publish because capital allocation is too important to hide behind mystique, and because we want other firms to read our work, challenge it, and build on it. If our models are right, they should be auditable. If they are wrong, the research community should be the one to tell us. Science compounds faster than secrecy, and entrepreneurial financing, along with every other domain that relies on expert judgment, will improve faster if the whole industry works in the open.
What This Means for You
If you are a founder: You should not spend six months pitching a firm that will reject you with a one-line email. Talk to V through Entrepreneur OS and get an honest read on where you stand, with a clear path forward if the answer is not yet. If you are still deciding whether to start a company at all, Founder Mode will tell you what kind of founder you are in about ten minutes. Both are free. We built them because we want entrepreneurs to succeed, whether we end up investing or not.
If you are an LP: You should not allocate based on narrative. Vela OS is not a black box, and we do not run one for our LPs and a prettier one for the slides. Our major LPs are inside the same system we use every day, reading the same diligence, seeing the same probability scores, and participating in the process in real time. If that is how you want to allocate, we want to talk.
If you are a researcher: Every meaningful question in venture capital is a research question, and we have built the infrastructure to answer them. Our papers are on arXiv, our benchmark is at vcbench.com, and our core framework is open source at thinkreasonlearn.com. If you are working on reasoning, reinforcement learning, rare-event prediction, or interpretability, we want to collaborate. The Oxford partnership is one such collaboration. It does not have to be the only one.
Feel free to drop us a message at engage@vela.partners.