Bit Primitive

Bit Primitive is a software and AI company: data tools, AI that learns from your own data, web3 and crypto analytics, and self-hosted boxes like Estate Box. Plus open datasets.

Bit Primitive is a software and AI company. We build data tools, AI that learns from your own data, web3 and crypto analytics, and self-hosted boxes that keep all of it at home. You run them yourself; nothing passes through us.

What we build

• Data tools: Collect raw data and organise it into clean, documented, tested tables in a warehouse you own.
• AI that learns from your data: Small, transparent models trained on your own data every night, kept only when they measurably improve.
• Web3 and crypto analytics: Blockchain data as star schemas, and crypto moves explained, forecast and alerted on from every angle.
• Self-hosted boxes: A warehouse and an AI brain on a small computer at home. Your data never has to leave it.

Estate Box. AI built in.

A small, self-hosted box that brings your statements, exchange activity and market feeds into a warehouse you own, and runs an AI brain that trains on that data every night. It packages Data Estate, the reference build that runs daily on a Raspberry Pi.

• Trains on your data: Models are trained from scratch on the box itself. No downloaded weights, no external AI service.
• Champion and challenger: Each night a new version competes with the current one on held-out data, and wins its place or is dropped.
• Forecasts that are graded: Every forecast has a range and is scored against what really happened once the date arrives.
• Answers with sources: Ask in plain words. Answers come from facts in your warehouse, with their source. It never invents a number.

Every night, on the box

1. Collect new statements and feeds
2. Rebuild and test the warehouse, then swap it in
3. Train a challenger for every model
4. Score it against the champion on unseen data
5. Promote only what measurably wins
6. Publish forecasts, insights and answers to your phone

Meet Estate Box

How it works

1. Collect (cryo · dlt · APIs): From statements, exchanges, blockchains, news and public data sources, through their own APIs, gently and within their terms.
2. Organise (dbt · DuckDB · Postgres): Into documented tables with tested keys and joins, in a warehouse on your own machine, server or box.
3. Learn (self-training models): Models trained on your own data, retrained as it grows, and promoted only when they beat the current version on data they have never seen.
4. Answer (apps · SQL · open data): Forecasts with ranges, insights and plain-language answers on your phone, SQL for the deep dive, and open datasets anyone can use.

Try chainstar live: query its real output in your browser.

Try sigstar live: query its real output in your browser.

Projects

• Estate Box: Your own data warehouse and an AI that learns from your data, in one self-hosted box. Nothing leaves it.
• chainstar: Raw blockchain data to clean dimensional tables in your own warehouse, from one command.
• sigstar: Collect market, on-chain, news and macro data, and explain, forecast and alert on crypto price moves from every angle.
• Whittle: Describe a part in plain words and get a verified, printable file. Runs fully local and offline.
• Data Estate: A personal data warehouse with a self-training AI brain, running every day on a Raspberry Pi. The reference build behind Estate Box.
• Orbit: See everything Claude Code does, live on a map in your browser. Replay sessions, read past chats and track cost. Runs on your own computer.
• Lore: A data file that explains itself. Parquet that carries its own meaning, privacy flags and an AI-readable brief.
• Tap & Tank: When is the water off, how full are the dams, and what will fuel cost next month? Alerts for South Africa, starting in the Western Cape.
• Till & Stack: Every loyalty card in one offline wallet, with optional crypto cashback through licensed partners. No custody, ever.
• SynthDeck: An AI music studio for DJs. Generate tracks from a prompt, split stems, auto-mix and remix on two decks, all on your own machine.

Public data

• Base ERC-20 transfers (100-block sample): Every ERC-20 Transfer event in 100 consecutive Base mainnet blocks, decoded, with each token's symbol, name and decimals joined in.
• Bitcoin transaction outputs (3-block sample): Every output of every transaction in 3 consecutive Bitcoin blocks: who received how much, the output's script type, and the sender when every input came from a single address. OP_RETURN and other outputs without an address are kept, with no address.
• ERC-20 token directory (Ethereum and Base): Every ERC-20 contract seen in the Ethereum and Base transfer samples, with the name, symbol and decimals it reports and how many transfers it had. Useful for decoding amounts.
• Ethereum ERC-20 transfers (200-block sample): Every ERC-20 Transfer event in 200 consecutive Ethereum mainnet blocks, decoded, with each token's symbol, name and decimals joined in. A sample of what chainstar produces from a single command.
• Ethereum transaction fees (200-block sample): The execution fee of every transaction in the same 200 Ethereum blocks: gas used, the price paid and the base fee, in wei and in ETH. Blob fees are not included.
• Solana DEX swaps (300-slot sample): Every swap in 300 consecutive Solana slots through Raydium, Orca Whirlpool, Meteora DLMM, Jupiter, Pump.fun or PumpSwap: what the signer gave and got, with a USD value from prices derived from the same on-chain swaps.
• Solana token and SOL transfers (20-slot sample): Every SPL, Token-2022 and native SOL transfer in the successful transactions of 20 consecutive Solana slots, inner instructions included, with each mint's standard and decimals.
• Solana token prices from on-chain swaps (300-slot sample): Hourly and daily USD prices for every Solana mint traded in the same 300 slots, derived only from on-chain swaps against USD stablecoins and SOL: outliers rejected, each price with its method, observations, volume and confidence. Low-confidence prices are left NULL, with the observed median kept.