spent the last month building a personal AI research infrastructure on a mac mini. no cloud. no saas. here's what’s running
full bitcoin node. bitcoin core v29, 944k blocks, full txindex. powered by
@umbrel home. also running a local
@mempool instance,
@mononautical's bitfeed, and
@w_s_bitcoin's quantum exposure dashboard to track P2PK address exposure
bitcoin analytics DB. postgresql ingesting every block in real time. per-block fee rate percentiles, hash rate, segwit %, inscription counts, miner IDs, transaction pattern classification, and address type breakdown (P2PK through P2TR). daily aggregates: puell multiple, mayer multiple, NVT, supply issuance, MVRV, SOPR, URPD, and many more. 5,700 days of price history and blockchain data. a lot of what i'm building is directly credited to or expands on work by
@checkmatey,
@TXMCtrades ,
@nic_carter,
@willywoo,
@w_s_bitcoin, and many others. will share more about what these look like in the future
OFAC sanctions and known criminal address monitor. 518 sanctioned BTC addresses from treasury's SDN list, data from open source attribution sets, scanning every block and unconfirmed mempool tx. instant telegram alert the moment a sanctioned address moves
large PnL alerts. monitors inputs ≥10 BTC for realized gain/loss vs. cost basis. fires when >40% move and >$1M. "address bc1q...xyz moved 847 BTC at $62M profit"
obsidian vault, 2,200 documents. full bitcoin optech archive, delvingbitcoin posts, bitcoin-dev threads, every satoshi email and forum post, galaxy research and podcast transcripts, SEC/CFTC/fed filings, GENIUS and CLARITY act text, other congressional legislation and press releases. all also ingested into the kuzu graph. morning digest to telegram daily at 8am
LLM wiki. a modified version of
@nvk's implementation of
@karpathy's llm-wiki pattern. the AI doesn't just index documents, it reads and maintains a persistent cross-linked knowledge base. new source in, wiki updates. contradictions flagged. knowledge compounds. extremely useful already
kuzu knowledge graph. thousands of documents cross-referenced by entity, topic, and source. semantic search in seconds
lightning network. using my node, LND, and LNbits to give clem (my AI assistant) full lightning capabilities. he can create and pay invoices on my behalf. not sure what i'll use this for yet but it's live. maybe just to easily send and receive upon my instruction and later to help manage an expanded lightning node
what's next: bitcoin transaction tracing, address clustering, entity attribution and behavioral pattern matching (chainalysis-style tooling, self-hosted). macro and fed data ingestion. a scientific research library. and eventually the same on-chain stack extended to ethereum and solana (hardware permitting), primarily to track defi and stablecoin flows
all of this on a mac mini M4, 48GB RAM, 2TB SSD. clem coordinates everything on signal 24/7, routing to my local models as needed and appropriate (mostly gemma4 26B and deepseak R1 32B). zero access to my icloud, email, contacts, or personal data of any kind. operates entirely within a sandboxed research workspace
~$100/mo total: $20 in claude and chatGPT subscriptions, ~$60 in anthropic API fees.. about $100/mo. hardware was one-time purchase
everything else: self-hosted, open-source, mine
what else should i build? this year has been the most exciting time for me building personal technology in years
if you aren’t building with AI, what are you even doing?