Why did Endeit build Aletheia as a proprietary sourcing and screening engine?
It was analysed through a ‘buy versus build’ lens: if the market has a solve that meets your needs, then you buy. But for us, something that could source and screen at scale from the perspective of Endeit’s investment thesis just didn’t exist. Actually, a big screening engine like Aletheia that was reliable and had the quality we wanted just wasn’t possible more than 12 months ago. With that reality, we had to build what we needed.
What brief were you given when you started building?
The mandate I was given was to imagine setting up a fund within our fund. So how would I build a fully AI-native fund completely from scratch? How would it behave? What would it need to do? How to realise that was completely up to me - I was given a blank canvas to iterate whatever I thought answered the requirements.
How does Aletheia fit into the wider Endeit ecosystem?
It began as a sourcing and screening engine, but really it’s more than that. It can tackle any problem that requires company screening. It looks at what the market’s like: its growth data, traction, scale, and analyses its AI positioning, and comes up with a weighted average. All of this helps to get a feedback loop going so Aletheia improves. It also observes 70 data points to select companies that we need to follow. It tracks rounds being raised and looks at relevant companies over the last 2 weeks to keep up with what they’re doing, which allows us to be opportunistic if something interesting surfaces.
Aletheia also functions as an extra member on our Investment Committee Council (ICC). Effectively, it acts as an extra partner on the team, giving an opinion on if we progress with investments. As an AI agent, it assesses the key thesis, questioning what needs to be true to return 5x. It then takes two roles: the sponsor and the sceptic - and then advocates for each position based on what it knows. Each role is played by a different LLM to mitigate the biases that exist within LLMs. But it acts as a supporting decision-maker to advise the partners, who always have the final say.
It also looks at deal terms. You could have a great company, but the price isn’t there, so it calculates if the proposed deal terms will deliver on what we modelled. Its role is to provide food for thought or propose a view that perhaps hadn’t been considered. But it’s always the team that decides if it’s important or not.
Were there any existing platforms that you looked at for reference?
In terms of AI, a similar model didn’t exist. Funds had sourcing engines, but they relied on classic machine learning as opposed to LLMs. In the R&D phase for Aletheia, in Q4 of 2025, I tried a machine learning model, but its performance just wasn’t good enough. It took a lot of interviews with the team to define exactly what we wanted its capability to be, and then I discussed my thoughts with Endeit’s venture partner Professor Jan Scholtes, who has decades of experience in artificial intelligence and natural language processing.
Now we have a model that does a full analysis. But instead of doing the whole analysis in one-go, we split it up in pieces to preserve the LLM's reasoning quality; because as the context (number of tokens) in your context window grows, the performance of the LLM worsens. Splitting it counteracts that.
What has been the most challenging issue so far?
The most difficult part has been how to know if the model works well. This is why you set up benchmarks and evaluate, with a classification metric for accuracy. The tough part is setting up the benchmark: how you compose and define it, which companies to use for the benchmark, and augmenting it with the team’s awareness as ‘boots on the ground’ in the industry. It’s tailored towards each team and each segment Endeit operates in. But, of course, our investment thesis evolves, and will depend on our current fund’s lifecycle, so the moving parts are always variable. It’s a constant iterative process.
There’s a whole monitoring system in place to make sure the sourcing and screening system still works in the way it did when we benchmarked. The model might drift and have an error in the analysis, but I built a dashboard for all of this and I check it constantly. It’s important to stay close to the team for their real-world input.
What was the initial reception of Aletheia by Endeit the team?
The initial reception was really good; they were excited to start using it. But as anyone who’s initiated a new process in any company will know, there’s a classic adoption issue, and a discrepancy between what people say and what they do. People get excited but don’t use it enough, which can result in investment opportunities falling through the cracks. Though the reality is this was totally new for the team’s workflow, and they had to get to know and understand Altheia’s features and capabilities.
What have been your learnings in the process?
I had to learn what to optimise for based on the reality of how the team worked. Initially, I made the model let through as many companies as it could, as that seemed to make sense, at the cost of letting through some less-than-ideal ones. But in practice, this made it way too much for the team to look at - and of course some of the selection choices made it seem like the model wasn’t great at evaluating.
After their feedback, I calibrated the balance to be stricter and more exacting, which has worked much better. Though I always tweak the model behind the scenes based on the team’s perspective. Negative feedback on companies is actually more helpful. For example, finding out that a business model is wrong can help fine tune how Aletheia works.
Any positive feedback gets translated into action by the team. We can do trend analysis built on all the data points, and it might resurface some companies we might have forgotten, which the team can revisit to see where they stand now.
Do you think more VCs will build their own platforms?
More tech guys are being hired at VCs, so there seems to be a shift in the need to analyse huge amounts of data, which will continue. VCs that don’t keep up will fall behind, and they’ll eventually end up being at too big a disadvantage.
To truly comprehend what companies are doing in the market with their own AI platforms, it makes sense to understand what you’re investing in, and what it can do. For that reason, funds are now hiring more tech people.
What insights have become clear since deployment?
Since Aletheia was deployed in March this year, it’s conducted 30,620 evaluations, which far exceeds what Endeit would have been able to do previously.
Deal sourcing has always been super repetitive and labour intensive, but regardless of automation, someone still has to build a connection with a company to stand behind it and be an advocate for why we should invest. Human analysis is still required. Questions still need to be answered, and nuance still needs to be addressed, and of course there has to be reasoning behind the numbers. You still have to ‘know the why’ and need full comprehension to understand all aspects of a deal.
What lies ahead now that Aletheia is up and running at Endeit?
Ultimately, Aletheia is a stepping stone towards automating more, so the big picture is that AI runs everything that it makes sense for AI to run. But obviously, not all problems need to be done by AI. Human judgement is, currently, still required in the process, but it can assist in making things more efficient and higher quality across the board. Its role at Endeit is to be an additional team member with a lot of capacity, that’s objective and brutally honest.
