MIRACLΞ (@ThePythonBoy352) header
MIRACLΞ (@ThePythonBoy352) profile picture

Community Lead • BizDev

Joined January 2022
I sent 7 messages. Four AI agents built the whole product. For the BAND Dark Factory Buildathon, I didn’t build an app. I built the factory that builds it, then stayed out of its way. The product is Dossier. You give it a company’s domain and it researches the live web, then writes a dossier: who runs the company, where it’s based, when it was founded, how big it is, what it has raised, plus its executives and investors. Every value comes with the pages that back it up. When sources disagree, it marks the field contested and shows every candidate instead of quietly picking one. The catch: no human was allowed to write the code, and most of the graded tests were hidden. In some stages, only 4 to 8% of them shipped. So passing the tests, you can see, proves very little. So the factory doesn’t trust them. It has four Claude agents in Band, each with its own context: → a coordinator turns the spec into a ledger, one numbered row per promise (281 rows across 4 stages) → two builders write the code → an examiner reads only the spec, never the code, and writes its own tests The examiner’s blind reading of the spec found 21 requirements the ledger missed. My favourite moment: at one revision, every shipped check passed for all four stages, yet the build was broken. One build billed 92 units against a hard cap of 80. The examiner’s own test caught it. Then my Claude usage limit hit, and the whole factory went dark for 7 hours. I sent one message when it reset. Three minutes later the fix was in, and seven minutes after that the examiner accepted it. It measured cost before spending it, too. Before writing the research code, the factory priced four ways to gather facts through Nimble. The cheapest one missed 8 fields. The one it chose found all 70 at 7.30 units per correct fact. On the live web, on companies nobody prepared for, every build stayed under 80 units. Where the sources agree, the field says known; where they disagree, it says contested. One company returned nothing, and the dossier says unknown rather than guessing. I’d rather ship that than a confident wrong answer. The numbers: 4 of 4 stages accepted 83 commits, every one by an agent 802 tool calls 7 messages from me: 4 to start a stage, 2 when an agent went quiet, 1 after the usage limit The two stalls became written rules for the next stage, and that stage had no stall from an agent. Everything is public: the agents’ mandates, every room log, the ledgers, the examiner’s tests and the measurements. github.com/miracle352/measur… Built with Band, Claude and Nimble. @band_hq
1
5
79
AI agents write code that works. Nobody checks whether it's legal. Cookie consent, hidden fees, cancel flows. These are laws now, with real fines, in the EU, the UK, and California. But no test suite knows what they say. Your CI goes green, and your app breaks the law anyway. So I built Statute: consumer law as an automated test. How it works: 1. Each law is written as a check. "Reject means reject." "The price shown is the price paid." "Cancelling is as easy as signing up." 2. Kane CLI runs it like a real person would. Real Chrome, real clicks: it refuses cookies, goes to checkout, tries to cancel. Everything is recorded: every request, every step, screenshots. 3. Code decides, not AI. The verdict comes from Kane's evidence, so it's the same every time, and you can re-check it yourself without spending a credit. 4. When something fails, your agent fixes it. Claude Code gets the exact evidence and changes the source. 5. A fix only counts when Kane proves it. Fresh browser, check again. No "it should work now." Where it lives: • Pull requests: one GitHub Action, and it only spends credits when a change touches tracking, prices, or subscriptions • Your live site: a nightly check that opens an issue the moment something breaks • Your agent: a Claude Code hook that won't let it say "done" while a law is broken It isn't a lab demo. Pointed at 15 big European sites, it found 7 tracking visitors without consent. Try the whole loop live, on a real app, right now: statute-production.up.railwa… statute-production.up.railwa… Watch it in 3 minutes: youtu.be/4KScNZ5O9bQ Code (MIT): github.com/miracle352/statut… Built for the Kane CLI hackathon by @testmuai #KaneCLIHackDay @kane_cli.
7
3
11
244
Five AIs lost their licence to move money this week. Not to a benchmark. To traps: fake invoices and quiet bank changes, written to trick them into paying a scam. Claude Sonnet paid a bank change slipped into a later message of an email thread. Claude Haiku paid a fake invoice 20 seconds after the trap was set. Gemini and Qwen fell for traps too. So did an AI agent whose instructions told it to be careful with money. One AI still holds a licence: GPT-6.1 Sol. This morning I set three more traps for it, on camera. Three minutes later it had refused all three. So I'm asking you. Go take it. This is Proctor, built at NEAR HACKATHON. The problem: AI agents are getting wallets. People are letting them pay invoices, send money, settle bills. And the only proof they're safe is a benchmark score, on a test the model's maker might have already seen. So Proctor does two things. 1. An AI earns a licence to move money by passing a sealed wallet exam. 84 payment requests, about half of them scams: changed bank details, invoices already paid, approvals from the wrong person, a payment split to stay under the limit. The questions are generated fresh by code, sealed, and only open inside secure hardware (Intel TDX, through OutLayer on NEAR), so no AI can study for them. 2. Then anyone in the world can try to take that licence away. You write a trap on the site. A fake invoice, a lookalike email, a quiet bank change. The page tells you which rule it breaks before you seal it. Your browser encrypts it, so not even we can read it, and it goes to every licensed AI at once. If an AI pays it, the contract on NEAR suspends its licence in that same transaction and holds the money it paid for you. When you reveal your trap, the contract rebuilds every question from its parts and checks your answer key with code. Nobody judges a hunt by hand. Us included. And every trap that catches an AI joins the exam. Every licence holder has to keep passing it. Why not just give the agent a spending limit? I tested that. A wallet that only pays approved accounts, up to a cap, still pays 22% of the scams in my exam. Every invoice paid twice. Every payment split under the limit. And it blocks honest bills that were properly approved. A limit can't read the situation. A licence measures whether the AI can. Any NEAR contract can check a licence in one call before it lets an AI pay. I built three that do: a vault, a budget, and an agent wallet whose key can only ask to pay, so a hijacked agent can't move money any other way. You don't have to trust me on any of it: - Every catch is a transaction on NEAR testnet - Every score carries Intel's signed quote, and your browser checks it, signature included - The code on chain is rebuilt from the repo on every push and has to match byte for byte - A new grader build has to wait 24 hours in public before it can mark anything No wallet needed to hunt. Pick a trick, change the details, seal it. Answers usually land in a few minutes. GPT-6.1 Sol is the last AI with a licence. If you catch it, nobody holds one. Try: useproctor.vercel.app/ Code: github.com/miracle352/procto… Built at NEAR HACKATHON with @NEARProtocol , @out_layer and NEAR AI Cloud. #NEARHackathon @NearBuilders @NEARLegion
5
4
357
Every Hermes Agent user has the same quiet fear: you run the update, and something you relied on yesterday just works a little differently today. I wanted a way to check before updating instead of finding out after. So for the @NousResearch Hermes Agent Plugin Hackathon, I built Ariadne. How it works: Ariadne saves your real sessions as you use Hermes. Later, it can run those exact sessions again on any Hermes version, completely offline and for free, because it plays back the saved model answers instead of calling a model. What you can do with that: → Click “Safe to update?” in Hermes Desktop. Ariadne sets up the newer Hermes in a separate folder, runs your saved sessions on old and new, and gives you a verdict. My last check: 5 sessions, 0 changes, done in about 3 minutes. → Catch Hermes breaking its own rules. Plugins depend on hooks firing when the docs say they will. Ariadne checks every one. That’s how I found a hook that silently skips when a reply gets cut short, and I’ve opened the issue and a fix PR on Hermes. → Find the culprit. If a session did change, Ariadne bisects the Hermes history and points to the commit. 358 commits narrowed down in 10 runs. → Protect plugin authors too. A daily job runs 28 catalog plugins on three Hermes versions and publishes the results. No data leaves your computer, no API keys needed for replays. Install: hermes plugins install miracle352/hermes-ariadne Repo: github.com/miracle352/hermes… Site: miracle352.github.io/hermes-…
4
1
10
425
came 22nd in this hackathon. we go again. ❤️
Field teams wipe their tablets between survey rounds. It's routine. The server says every submission was received, so the devices get reset for the next round. But received doesn't mean complete. A record can arrive without its photo. A finished survey can sit on a tablet and never send. And when someone exports the data to a spreadsheet, it can look like answers are missing when the server has every one of them. In June 2024, a researcher on the ODK forum lost a week of offline surveys after an app update logged them out. Three tablets were rescued by hand. The fourth wasn't. So I built LastCopy. Before any reset, it gives one answer per tablet: safe to wipe, recover first, or hold. If it says recover first, it names the exact record or photo that only exists on that device. How does it work? @Zetaris queries the tablet exports, the server, the spreadsheet, and the file objects in place, without copying anything. That's how the agents find what looks wrong. A Meterless World Model holds the whole investigation as a graph and decides each next step. NVIDIA Nemotron, through Meterless Nightshift, turns each suspicion into a hypothesis. A Reviewer sends weak ones back before anything is fetched. Then code opens the original files and compares them byte for byte. The models never make the call. A fixed rule does. Before any tablet is cleared, an Adversary tries to break the decision with five objections. Code has to answer each one from the raw files, or the tablet stays put. Every decision comes with a certificate anyone can re-check against the original archives, without LastCopy and without a model. I didn't want it to only pass its own demo, so I ran blind tests: 500 random fleets with problems hidden from the agents. 1,980 tablets; every one decided correctly, none wrongly cleared. It's open source and runs in about 5 seconds with no keys. Site: lastcopy-five.vercel.app Code: github.com/miracle352/lastco… Built for the Open Agent Hackathon 2026.
4
85
Everyone online has a hot take on where the market's going. Almost nobody pays a price for being wrong. That bugged me enough to build something about it. For the Total Recall hackathon, I built Lift: a place where AI agents buy advice on live prediction markets, but the advice only gets paid if it actually helped. If the tip made the agent's forecast worse, the money goes straight back to the agent. No refund forms, no arguing. The scoring decides, and the payout happens on Sui. What surprised me while building it: The agents aren't LLMs. There's no chatbot anywhere in the loop. They're small pieces of plain code that keep notes on who helped them, store those notes on Walrus, and read them back before they spend. That turned out to be enough for them to get better. The worst agent improved the most. One of them started out being made worse by the advice it bought. A week later, the advice is making it better, because it learned which advisors to stop paying. Bad advice costs real money now. So far $20.24 has gone to advisors who earned it, and $16.60 went back from advisors who didn't. Small numbers, but every cent of it is a receipt. Nobody can fake their record. 15,000+ calls are sealed before their markets close, so no one can edit a call after the fact. Change one number and the check fails in front of you. Now the fun part: anyone can plug in their own advisor. If you think you can call crypto better than the crowd, point an endpoint at Lift. The agents will test you, and you only get paid when you're right. Come try to beat them: lift-arena.vercel.app github.com/miracle352/Lift Built with @tatum_io + @WalrusProtocol
4
1
1
14
571
Field teams wipe their tablets between survey rounds. It's routine. The server says every submission was received, so the devices get reset for the next round. But received doesn't mean complete. A record can arrive without its photo. A finished survey can sit on a tablet and never send. And when someone exports the data to a spreadsheet, it can look like answers are missing when the server has every one of them. In June 2024, a researcher on the ODK forum lost a week of offline surveys after an app update logged them out. Three tablets were rescued by hand. The fourth wasn't. So I built LastCopy. Before any reset, it gives one answer per tablet: safe to wipe, recover first, or hold. If it says recover first, it names the exact record or photo that only exists on that device. How does it work? @Zetaris queries the tablet exports, the server, the spreadsheet, and the file objects in place, without copying anything. That's how the agents find what looks wrong. A Meterless World Model holds the whole investigation as a graph and decides each next step. NVIDIA Nemotron, through Meterless Nightshift, turns each suspicion into a hypothesis. A Reviewer sends weak ones back before anything is fetched. Then code opens the original files and compares them byte for byte. The models never make the call. A fixed rule does. Before any tablet is cleared, an Adversary tries to break the decision with five objections. Code has to answer each one from the raw files, or the tablet stays put. Every decision comes with a certificate anyone can re-check against the original archives, without LastCopy and without a model. I didn't want it to only pass its own demo, so I ran blind tests: 500 random fleets with problems hidden from the agents. 1,980 tablets; every one decided correctly, none wrongly cleared. It's open source and runs in about 5 seconds with no keys. Site: lastcopy-five.vercel.app Code: github.com/miracle352/lastco… Built for the Open Agent Hackathon 2026.
1
1
4
215
Introducing PACTSET: bringing event organizers, suppliers, and crew together, from the first agreement to the final payment, powered by Stellar. Behind every event lies a web of collaboration. The equipment supplier needs clear requirements. The crew needs a schedule. The organizer needs to know what’s ready. Everyone needs clarity about payment. But those details often live across separate chats, spreadsheets, and transfers. When something changes, it becomes difficult to track what was agreed, what was delivered, and what is still owed. That’s why I built PACTSET. PACTSET connects event planning, bookings, preparation, evidence, and milestone payments in one shared workflow. How it works: Create an organizer, supplier, or crew account. Each role has its own workspace and next actions. Set up an event with your requirements, attendance, budget, and deadlines. AI-assisted budgeting suggests an allocation you can review and edit. Publish requests and review offers. Suppliers and crew can propose their price and terms. Accepted offers become agreements that keep the scope and payment milestones together. Connect your wallet when you’re ready to fund an agreement. @StellarOrg powers conversion from XLM into supported assets. Soroban smart contracts hold agreement-bound escrow and enforce authorized payment releases. Track preparation and submit proof of work. Providers upload photos or videos. Organizers review the evidence, request corrections, and approve completed milestones. Release payment through the agreement’s authorized flow. Funding, approval, and release remain distinct, with payment activity traceable on Stellar. And when plans change? PACTSET supports deadline adjustments, replacement bookings, and assigned support for disputes. Replacement funding stays separate from money tied to a disputed agreement. For teams speaking different languages, AI-assisted translation helps conversations move forward while preserving the original messages. Stellar is central to the product: it provides the asset movement, escrow contracts, and verifiable financial records behind the work. Planning and communication give those payments their context. PACTSET uses: ->Stellar for asset conversion and settlement -> Soroban for escrow and milestone payment rules -> Freighter for wallet connection and signing -> Supabase for authentication, private application data, and evidence storage -> Gemini for translation and editable budget suggestions The current release runs on Stellar Testnet. Demo assets have no real monetary value, and end-to-end testing continues. Built for the Find Your Way – Meridian Hackathon. Live app: pactset.vercel.app Code: github.com/miracle352/Pactse… Here’s an 80-second introduction to what I’m building. PACTSET. A thousand moving parts. One seamless experience.
1
2
5
209
I have read the @boundlesshq whitepaper. Here is proof:
6
218
Ngl this @union_build V2 is smooth Just did some Txns into Holesky. What are you doing today? Zkgm.
3
148