TWIN
AI trading twin · Solana · pump.fun

Copy the taste,
not the trade.

TWIN reads a wallet's on-chain history and learns what it buys: market cap, token age, momentum, volume and timing at every entry. Then it scans new launches for the tokens that wallet would pick, before it buys them instead of seconds after.

Trained on real on-chain entriesValidated before it speaksLive pump.fun and Solana poolsRuns in your browser
Training data

The public Solana RPC rate-limits hard; a free Helius endpoint works well. Price history comes from GeckoTerminal (about 30 requests a minute), so training 40 tokens takes around a minute and a half.

How it learns

A model of one wallet

TWIN does not guess from a prompt. It measures the wallet's real entries against the market it was trading in, and only shows a score once it has proven it can tell the two apart.

step 1

Find every entry

Parsed swaps from the wallet's own transactions. The first buy of each token is an entry; its exit gives the outcome.

step 2

Rebuild the moment

Market cap from the pool price in the buy itself. Age, momentum and volume from the pool's minute candles before the buy.

step 3

Learn the taste

Kernel densities for the wallet's entries and for the live market. Their likelihood ratio, feature by feature, is the taste.

step 4

Prove it, then scan

Leave-one-out test on its own entries. Then every new and trending SOL pool is scored, with the reasons and similar past buys.

# x = a token right now: market cap, age, 5m and 1h momentum, 1h volume, curve or graduated, hour
twin_score(x) = σ( Σ_f clip( log p_wallet(x_f) − log p_market(x_f) ) )
accuracy      = P( score(real entry) > score(market token) )   # leave-one-out AUC, 0.5 = no edge
past_buys(x)  = outcome of the 5 most similar entries in its history
The token

$TWIN

Launching on pump.fun. Training a twin and scanning the radar stay free. Holders get the tools that keep the twin watching when they are not.

CA: not deployed yet

Twin alerts

A message the moment a new launch scores above your threshold.

soon
Blended twins

Merge several wallets into one taste and see where they agree.

soon
Twin league

Public twins of top wallets, scored weekly on their picks.

soon
Built with SHADOW

Every pick shows what a copier would really keep after fees and latency.

soon
FAQ

Questions

Is the twin score a buy signal?

No. It measures how much a token looks like the wallet's past entries, compared with the rest of the market. A wallet's taste can be bad: check the "similar past buys" line under each pick.

What does twin accuracy mean?

It is a leave-one-out AUC. Each real entry is hidden from the model, then scored against market tokens. 0.50 means the wallet's choices look random; above 0.75 the wallet has a clear, learnable style.

How many trades does it need?

At least 10 entries to train, 25 or more for a stable profile. Wallets that only trade graduated coins or only snipe at launch are learned fastest.

Why is the market reference "now" and not historical?

Public APIs do not list every token that existed at each past moment. TWIN uses the current set of new and trending SOL pools as the market, and treats time of day separately so the snapshot hour does not bias it.

Do you store wallets or keys?

No. Everything runs in your browser against the RPC you choose, and TWIN never asks for a signature.