Every bank on Wall Street is racing to ship "AI banking agents." Regulators are not thrilled.
The CFPB's warnings boil down to three failure modes:
• Hallucinated balances and policies
• Customer PII streaming into third-party cloud APIs
• A probabilistic model doing the arithmetic instead of a ledger
A hallucinating support bot is embarrassing. A hallucinating banking agent is a compliance violation.
So today I had MacAttack, my local multi-tier AI engine, build a conversational banking agent + double-entry ledger from scratch. Zero human intervention.
Why it's different from the cloud chatbots in the news:
1/ Zero cloud. Every line of code, every reasoning trace, every DB call ran on an M4 Max with open-weight MLX models. No balance, account number, or transcript ever left the machine.
2/ Language ≠ ledger. The LLM only translates intent ("move $250 checking → savings") into type-checked function calls. Money is handled by a double-entry SQLite ledger with hard invariants:
• Conservation: debit + credit in one atomic txn, nothing vanishes or appears
• Zero overdraft, rejected at the DB layer before funds move
• Idempotency keys: double-tap "Send" and the replay returns DUPLICATE
• Anything over $5k is held REQUIRES_APPROVAL without touching balances
3/ Built by a swarm, not by me:
• Tier 1 (Qwen3-4B) spec'd schemas and I/O contracts, ~37ms TTFT
• Tier 3 (Qwen3-30B-Thinking) stress-tested race conditions and replay attacks
• Tier 2 (Qwen3-Coder-30B) wrote the implementation and adversarial tests inside a default-deny macOS sandbox
81 tests passing. Applied only after full verification.
The future of AI in finance isn't a cloud black box guessing your balance. It's on-device intelligence wrapped in deterministic, zero-trust engineering.
#LocalAI #MLX #AI
@GarbageHuman24 Second-degree assault? That's attempted murder, and everyone who helped is an accomplice! This whole system is ass-backwards. My taxes and hard work are paying for these barbaric animals. I'm done paying for what I shouldn't have to.
@BethanyForTruth Yeah, total bullshit. So does this mean I don't have to pay taxes, I can work 11 tech jobs at once, and I can do a little insider trading when Anthropic goes public?
Yeah. If you look and listen, someone fired a shot; the officers jumped back, and then they used lethal force. We've all seen how this plays out. The big question is who shot first. The bodycam footage will likely tell a different story than the person filming. We'll see.
Also, why are you still here? Aren't you being investigated for something like 200 counts of fraud?
@Goodnewsviki1 They have no regard for life, and life should have no regard for inhumane pieces of shit like them. Is there a verified fund for her baby girl?
I've been laid off 5 times in my tech career. I know how you feel. But don't give up yet, because something new is taking shape.
Several open protocols now let AI agents handle payments, API micro-transactions, and commerce:
- x402 (Coinbase): pay-per-request over HTTP in stablecoins
- MPP (Stripe + Tempo): machine payments with budgeted sessions
- L402 (Lightning Labs): pay-per-call in bitcoin
- AP2 (Google): proves what a human authorized an agent to spend
- ACP (OpenAI + Stripe): agent-to-merchant checkout
- UCP (Google + retailers): discovery to checkout
What engineers should notice: these protocols turn agents into paying customers.
Build an API, MCP server, or tool once, put x402 or MPP in front, and get paid every time an agent uses it: no sales team, no invoicing.
A niche tool that would never justify a startup can now earn per call, worldwide.
The catch: agents won't pay for what they can write in ten seconds. They pay for proprietary data, domain expertise, maintained integrations, and reliability someone stands behind. That's experienced-engineer territory.
The job market is rough. But the next opportunity might be packaging what we know into something agents pay for.
It's early, so now is the time to build. Building is getting cheap. Knowing what's worth building isn't.