Agents
Each agent has one job. They run in a fixed order set in code, from the deck read to the memo. Each one below covers what it does, the tools it uses and its edge cases.
Doc Intelligence
Reads the deck and the website and turns them into one structured record that every later agent works from.
- Extracts the problem, product, traction, business model, funding, sector, location, key metrics, deal terms, claimed market size and any competitors the deck names.
- Records which sources it could read.
- On re-analysis, reuses the previous extraction if the deck hasn’t changed.
Firecrawl- No readable content from the deck or the site: the run stops and the deal is held for a person.
- Canva and DocSend links are read by our own extractor, including DocSend folders of several files. Multi-column and scanned PDFs get a slower second pass when the first returns too little.
- No website supplied: the founder’s email domain is tested as the company domain, and used for research only.
Pre-screening gate
Checks the deal against our investment mandate before the slower, costlier research runs. It can recommend a pass, but only a partner can decline a deal.
- Checks the mandate rules: stage (pre-seed to Series A), round size (under $20M), an Australian connection, and whether it’s a technology business rather than direct-to-consumer or e-commerce.
- Notes warning signs, like a solo founder, no revenue yet, a crowded market or a valuation ahead of the stage. A warning sign becomes a question for the founder and lowers the score, but never ends a deal on its own.
- A recommended pass goes to a partner in Slack. If they disagree, the run picks up where it stopped.
- An exceptional founder based overseas. Location is one of the mandate rules, so a partner reviews every pass the gate recommends.
- Too many deals come back as more info. Recalibrated in October 2026, and still being worked on.
Founder Intel
Researches every founder’s career and checks it against what the deck says about them.
- Finds each founder’s professional history and interests, with sources.
- Flags any mismatch between the deck’s story and the public record.
- Runs alongside Market Trends and Enrichment.
Renidly
Web search- Someone else with the same name. A profile that doesn’t match the deck is reported as a mismatch and not used.
- The insider with no conventional CV, whose strength doesn’t show up on a CV.
Market Trends
Sizes the market from outside sources, so the deck’s own number gets tested instead of taken on trust.
- Estimates the total and serviceable market, growth and key trends, with a source on every figure.
- Compares that with the market size the deck claims.
Perplexity
Exa- A company creating a category that doesn’t exist yet finds no market data, and gets flagged as a thin opportunity. A partner can read the same gap as the reason to back it.
- If it can’t work out the sector, it skips the deal.
Competitive Intel
Finds competitors from its own searches, once the market research is done. It starts from the competitors the deck names but doesn’t stop there.
- Finds competitors and their recent funding, with sources.
- Uses at least two of the company name, domain and sector in every search, so it doesn’t drift onto a different company.
PerplexityFirecrawl- With no Doc Intelligence record to search from, it marks itself skipped instead of failing.
Synthesis A
Reads everything the earlier agents produced and drafts the memo.
- Rates every criterion strong, adequate or weak, with the evidence, grouped by category.
- Pulls the five most similar past deals as comparables.
- Lists the red flags, the warning signs, what it couldn’t verify and an interview agenda, and recommends proceed, pass or more info.
- Code, not the model, turns the ratings into the score.
pgvector
Supabase- If founder, market and competitor research all failed, it doesn’t run, and the run is logged as failed.
- Given a middle option, a model takes it, so too many deals come back as more info. Not solved yet.
Synthesis B
Checks the draft memo against its sources before it’s saved, so a partner never reads an unchecked memo.
- Re-reads the draft against the original documents and every agent’s output.
- Flags any claim the sources don’t support, like a citation pointing to a page that doesn’t say what the memo claims.
AI committee
In an investment committee the outlier view gets talked down to the median one, and at seed the outlier is often the deal. This committee argues every deal from four opposed positions, some modelled on the reasoning of investors we respect.
- The sceptic hunts for the reason to pass. The optimist looks for the upside. The futurist asks where the world goes if it works. The operator checks whether it can be built.
- Returns where they agree, the dissent that didn’t resolve, and a next step.
Meeting briefing
Lays the analysis out as the memo a partner reads. It formats what the earlier agents found and adds no reasoning of its own.
- Writes the memo into Attio as a Company Profile and Briefing Notes, with its score. Every earlier draft is kept.
- Formats it for the stage: discovery call, founder interview, or the investment recommendation report for committee.
- Posts it to Slack.
Attio
SlackNew context
When new material lands on a deal, it pulls out the facts, so the rest of the run can weigh them against the deck.
- Turns call transcripts, notes, emails, attached files and the deal team’s Notion notes into structured facts.
- Merges them with last time’s, or reuses the earlier result if nothing has changed. Then market and competitor research are skipped, except before investment committee.
Attio
Notion- The founder’s account, the research and our own read often disagree. Every claim keeps its source and a partner decides who to believe.
Enrichment
Finds the founders’ contact details, so the record is complete before a partner picks it up.
- Checks a series of data providers in turn and stops at the first good answer.
FullEnrich