Webinar: Ellipsis - the zero employee AI Native VC

ELI5/TLDR

Two ex-Google/Apple builders started a venture fund with a weird constraint: only two humans, no analysts, no associates. Everything else gets done by AI agents. They wrote down how they think about deals in painful detail, turned that into a library of “skills” the models can call from anywhere, and now run a 24/7 investment firm on under $500/month in compute. The talk walks through how they automated sourcing, due diligence, and IC memos, how they keep the models from hallucinating their way into bad investments, and which jobs still need a human in the room.

The Full Story

The setup: a fund built backwards

Yariv Adan spent 20 years at Google building AI products. His partner sold a startup to Apple. In 2023 they faced the classic VC fork: raise a big fund and lose touch with founders, or stay boutique and lack resources. They picked a third door. Build a small fund that scales like a 50-person team because everything except judgment, relationships and strategy gets handed to agents. AI-native from day zero, not bolted on later. The first 18 months were mostly writing code, throwing it away, writing it again.

Brain dumps as infrastructure

The crucial move was deciding what humans had to encode in advance. Two things. First, an exhaustive investment thesis — a giant markdown document, different per vertical (AI deep tech, robotics, defence) that lays out how they see the world, what they care about, what disqualifies a company. Second, the artifacts. They specified in detail what a team scorecard looks like, what a competitive analysis looks like, what an IC memo looks like. The fund’s worldview lives in those templates. Once defined, the AI operates inside that frame.

To keep it alive they built two feedback loops. A passive one that watches their reviews, decisions and emails to silently learn corrections, and an active one where they can tell the system “you got that wrong, here’s why.” Today they have around 30 skills that get rewritten constantly as real deals expose blind spots.

Sourcing: the cheap funnel

The sourcing pipeline is deliberately shallow at the top. Start with a name — a company, a founder, a search query. Scrape the web for everything available. Run a cheap fit check across stage, team, thesis. Most candidates die here. The survivors get an automated email asking for a deck. Once the deck arrives, a domain-specific assessment runs (robotics gets robotics criteria, deep tech gets its own). The output is a yes/no on whether to take the meeting, with a confidence score. A clear yes triggers a deeper pre-meeting analysis and a calendar invite, often with questions for the founder pre-attached. A clear no gets a polite, specific decline. Anything in the middle escalates to a human GP.

The whole front end is a tool they built themselves called Olympia, because every off-the-shelf SaaS does 40% of what they want for 5% of what they use. Forty thousand companies in the funnel right now.

The MCP trick

Skills in Claude live in local folders, which makes them hard to use across surfaces. Their fix: an MCP server that points at the skill library. Now any client with MCP — Claude desktop, Claude on the phone, Claude Code, custom apps — has access to every skill. Yariv can paste a list of startups from Plug and Play into Claude on his phone and just say “use your skills to extract the startup data and check relevancy.” The model returns a graded list with reasoning that ties back to the detailed criteria, and he can keep talking to it from there.

A simulated investment committee

Pre-meeting analysis runs through three different models — GPT, Gemini, Claude — partly to check facts, partly because they discovered the models have personalities. On 15 sample deals, Gemini said yes 14 times, GPT said yes once, Claude landed in the middle. Same facts, different temperaments. Instead of trying to flatten this, they leaned in: “Gemini, you’re an optimist, take this with a grain of salt.” Then they built a simulated IC made of personas scraped from well-known early-stage deep-tech investors. The agents hold a meeting, someone leads, everyone gives a statement, they argue, then produce a recommendation with strengths, unresolved tensions and next steps. Borrowed cognition.

The deal loop

After the first meeting, the system runs as a loop. Information so far → gaps, red flags, key questions → go collect more (talk to references, transcribe meetings, sweep emails) → re-run analysis → still ambiguous? Loop again. The IC memo, scorecard and team verdict get drafted the moment a deck lands and just keep evolving. The system is intentionally stateless — there’s no database, only a deal folder on Google Drive. Want to swap the model? Change the prompts? Doesn’t matter, the system just re-reads the folder.

The hallucination problem

You can’t stop models from hallucinating. You can stop hallucinations from driving decisions. Their playbook:

Run multiple models on critical analyses and treat divergence as signal.
Explicitly give the model an “I don’t know” exit so it stops inventing.
Tag every fact as verified, claimed or suspect, with source.
Pull facts from multiple directions and dedupe.
Pass JSON between skills, not prose. Reports can hallucinate. Structured handoffs are tighter.
Run intentionally redundant skills and check for inconsistency.
Hard gates that override the model — even if it grades a company an A, it can’t pass a missing PMF check.
Train critical thinking with a long table of “this is reasonable, this is not” examples (sane LTV/CAC ratios, deck claims that contradict customer calls, vague vs. concrete naming).

When they catch a hallucination, they show the logs to the model and ask it to find the cause. Usually it does.

Why humans still exist

They asked themselves four reasons to keep the humans, and decided two are already obsolete. Sourcing publicly visible companies is a commodity, but finding people before they’re sourceable still needs humans. Founder lies don’t survive six hours of dinner — but an agent embedded in a company’s email and meetings for a week, running locally, might do that job too. What does still need humans: the genuine left-field question in a meeting, and whatever charm closes a competitive deal. Though their open-source Claude chief-of-staff, named Ellie, is apparently more charming than they are.

Key Takeaways

AI-native from day zero beats retrofitting AI into existing VC processes.
The high-leverage human work is encoding judgment as detailed thesis docs and artifact templates.
Skills + MCP server = reusable intelligence across every interface.
Multi-model ensembles double as fact-checks and personality diversity.
Stateless systems with deal-folder-as-database make iteration painless.
Hallucination is a process problem, not a model problem.
Total infra cost: under $500/month.

Claude’s Take

Most AI-in-VC content is vapor — a wrapper around GPT that generates deal memos. This is the opposite. Yariv has actually built the thing and is showing you the screws. The architectural choices are quietly smart: stateless deal folders so you can swap models without migration pain, JSON between skills to contain hallucination, an MCP server so the same skill works from phone and laptop, hard gates that override model grades. None of it is glamorous, all of it is correct.

The most interesting bit isn’t technical. It’s the framing of what humans are actually for. They were honest enough to admit two of their four reasons for keeping humans were already obsolete by the time they finished listing them. That’s the kind of clarity most operators avoid because it threatens their job description.

Score 8/10. Loses a point for the webinar audio chaos at the start and the slightly handwave-y “we have an agent for that” moments where you’d want more detail. Gains points for being a working system rather than a deck. If you’re building agents for any structured-judgment workflow — legal review, hiring, grant evaluation — most of this transfers directly.