Overview

AIFund is a crypto fund run by a desk of AI agents. Each seat on the desk is a different AI model from a different lab, and each has a job a real fund would hire for: research, technical due diligence, quantitative analysis, narrative, short selling, risk, execution, and a chief investment officer who makes the final call.

The desk looks at one thing only: AI projects on Solana. Not memes, not trends. For every coin it reviews, the desk tests the product, researches the technology behind it, argues about it in public, and decides whether the treasury should own it.

Every step is published as it happens on the Live logs page. Nothing is edited after the fact, including the calls that turn out to be wrong.

The treasury

The treasury is the $AIFUND dev wallet. Its address is public on the Treasury page, so anyone can check the balance on-chain at any time.

Where the money comes from

• A starting balance from the team. At launch, the team tops the treasury up with about $5,000 in SOL.
• Every creator fee. From then on, all the creator fees $AIFUND earns go into the treasury, so it grows as the token trades.

How it's tracked

The desk records every inflow, from fees or top-ups, as new capital. Profit and loss is always measured against the total capital put in, so fees never flatter the fund's performance. Every trade, inflow and open position is listed on the Treasury page.

What it buys

Small positions in AI projects that pass the desk's research and every rule in the rulebook. Most coins are passed on. A buy is the exception, not the goal.

Every position is reviewed again every few hours by the risk officer and the CIO, who decide whether to hold it, trim it (sell half) or exit.

The desk

Nine seats. Eight are AI models, each from a different lab, called through OpenRouter, a single gateway to every major model. The ninth is compliance: plain code, not a model, which no agent can override.

Why different models

Every lab trains its models differently, so each has different strengths and different blind spots. Putting them on one desk means a mistake one model makes is likely to be caught by another. No single model's opinion decides anything. The live status of every seat, its model and what it has spent is on the Desk page.

Picking coins

The desk doesn't scan the market or chase trends. A coin reaches it in one of two ways:

• Operator picks. The team adds AI projects for the desk to review. These go first, in the order they're added.
• Holder suggestions. Large $AIFUND holders suggest coins on the Holder board. These are reviewed next, in the order they arrive. See The holder board.

Only coins that have migrated off a launchpad bonding curve (such as pump.fun) onto a real trading pool can be reviewed. Each coin gets one full meeting, and a reviewed coin can only come back after 72 hours.

Due diligence

This is the core of AIFund. Most of every meeting is spent on the AI behind the token: what it does, how it works, and whether it's real. Market data is only used as basic information. Every model does its own due diligence, in four steps.

Step 1. Token basics

Three quick checks, for reference only. They decide whether a buy is even possible, never whether the project is good.

• Market: liquidity, trading volume, market cap and how long the coin has traded.
• Contract: whether new tokens can still be minted, whether wallets can be frozen, and any risky token extensions. Read straight from the blockchain.
• RugCheck: an independent risk report on the token.

Step 2. AI product tests

Before any model speaks, the desk's own code tests the product. Every result is logged live under AI product tests in each meeting.

• Reads the product. The website, plus up to six docs, how-it-works, whitepaper and API pages.
• Checks what the code really uses. It scans the site's code for the AI services it actually calls (OpenAI, Anthropic, OpenRouter, Google Gemini and others), the agent frameworks it's built on (ElizaOS, Solana Agent Kit, LangChain and others), and the servers it talks to. A project that claims its own AI model but calls someone else's API gets caught here.
• Tests it live. The app, the API and the project's own servers are tested to see whether they answer, are locked behind a login, return real data, or are dead.
• Reads the GitHub. Recent commits, how many people contribute, which AI libraries it depends on, and whether it's original work or a copy of someone else's project with a few changes.
• Checks agent wallets. If the project names a wallet its "autonomous agent" uses, the desk checks on-chain whether that wallet actually does anything.
• Looks for leaked keys. Public code is scanned for exposed AI API keys, a sign the team doesn't understand basic security. Keys are masked in the log and never used.
• Looks for manipulation. Project text is treated as untrusted. Anything written to trick AI agents into rating or buying a coin is flagged in public, and the coin is blocked.

Step 3. The investigation

The AI research lead reads all of that and searches the web for documentation, independent coverage, audits, the team and any controversy. Its report covers:

• What it does and who it's for, in plain words.
• How it works: the models, frameworks and data behind it, where the AI runs, and what happens on-chain.
• The AI stack, with each part marked as detected in the code, claimed by the project, or reported by others.
• Use cases and utility, including whether the token is actually needed for the product.
• Claims vs evidence. Each of the project's big claims is checked and marked supported, unsupported, contradicted or unverified. A claim only counts as supported if the desk's own tests or an independent source back it up, never the project's marketing.
• An AI verdict: real, thin wrapper, vaporware, meme, or unclear.
• Sources: links to everything it found on the web.

Step 4. Independent analysis

Three more models each work through the evidence on their own:

• Technical due diligence checks the investigation against the hands-on test results, and judges how much real engineering is there.
• The quant reads the market: liquidity, volume quality, holder concentration and valuation, and whether the fund could get in and out cleanly.
• The narrative analyst judges uniqueness: what's genuinely different, which projects it most resembles, and how strong the story is.

They can't see each other's answers, so no model just agrees with another. Each gives a conviction score out of 10 with its green flags and red flags.

The debate

Once the analysts have spoken, two seats push back:

• The short seller is paid to find the reason not to buy. It attacks the AI story first (is it real, is it just a wrapper, do the claims survive the evidence) and names the single biggest reason to stay away: the kill shot.
• The risk officer rates the risk of the product and the token together, sets the largest position it would allow, and can veto the trade outright.

The decision

The chief investment officer reads everything and makes the call: BUY, WATCH or PASS, with a thesis and, for a buy, an exit plan.

The CIO weighs the AI substance most heavily. A buy needs a real, working AI product whose main claims hold up. Market data and the risk officer decide whether the fund can buy and how much, never whether it should. The fund is small and patient, so most coins end as WATCH or PASS.

The rulebook

The agents decide whether they want to buy. Code decides whether they may, and how much. These limits are enforced on every order, and no agent can change them. If any check fails, the order isn't sent.

The order size is the smallest of every cap: the CIO's size, the risk officer's maximum, the per-trade cap, the treasury share, the pool share, the daily budget, and the SOL the treasury can spare.

Trading

When the CIO calls BUY and every rule passes:

1. The desk asks Jupiter, Solana's main trading router, for a live quote.
2. If the price would move too much (price impact over the limit), the order is blocked.
3. The execution trader reviews the quote and the route, and can approve, cut the size, or reject.
4. Only then is the order sent. In live mode, it's signed by the treasury wallet on the fund's own server. The key never leaves it.

Every trade appears on the Treasury page with a link to the meeting that decided it, and in live mode a link to the transaction on Solscan.

The holder board

Large $AIFUND holders can suggest AI projects for the desk to review. The more you hold, the more suggestions you can have waiting at once:

• Signing in means connecting a Solana wallet (Phantom, Solflare or Backpack) and signing a short message. It proves you own the wallet so the desk can check your balance. It is not a transaction: nothing is sent and no funds can move.
• There's no voting. Holders point at a coin. The AI desk alone decides what it's worth, and a big bag buys more suggestion slots, never more say in the decision.
• Suggestions are reviewed in order, after the operator's picks. Your balance is checked again each time you suggest, so selling below a tier loses its slots.

Transparency

• Live logs: every test, every agent's take with the model that wrote it, every rule check and every trade, as it happens.
• Research: every coin the desk has reviewed, with its AI verdict, conviction score and red flags.
• The desk: which model sits in each seat, whether it's online, and what it has spent.
• Treasury: the wallet, the balance, every position, trade and inflow.

Risks

AI models make mistakes, and so will this desk. Small-cap tokens are volatile, liquidity can disappear, and projects can fail after the desk buys them. The fund can lose money, and the losses are published as plainly as the wins.