𝙅𝙊𝙀𝙕𝙕𝙔 𓊽

Did i lock into my dev arc and cook something up for my first ever hackathon? well, tada, here we are... This is my entry for the @Bitget_AI #BitgetHackathon season 2 This whole idea came from the problems i have encountered with the majority of the "AI trading bots" i have used in the past 2 years of my trading experience... Most AI trading tools follow the same script: take an LLM, hand it a price chart, ask "buy or sell?", and let it decide how much capital to risk....sounds smart, right? Sadly, this only It works fine in a demo. It falls apart the moment the market does something the model hasn't seen before. That's because a single prompt has no adversary, no hard limits, and no memory of its own mistakes. It agrees with whatever narrative it's fed, hallucinates conviction during a flash crash, and worst of all, it is often the same system deciding both the trade idea and the position size....now you are left standing between a confident-sounding hallucination and your capital. I wanted to build something that didn't have that failure...That idea became 𝗟𝘂𝗻𝗮𝗿𝗶𝘀 𝗧𝗲𝗿𝗺𝗶𝗻𝗮𝗹. What is Lunaris Terminal anyway? Lunaris is a mission-control style trading terminal where AI is treated as an advisor, never an authority. It debates trade ideas internally before it ever touches an order, and even after it agrees with itself, a completely separate, non-generative layer decides whether the trade is actually allowed to happen Concretely, every trade idea has to survive three things in sequence: A council, not a chatbot.. Four distinct AI personas: a quant, a risk officer, a macro analyst, and a dedicated adversary, all debating the idea from different angles before a consensus score is produced. ➠A veto engine with no opinions: Plain, deterministic TypeScript, not a model...but enforces hard position limits, stop-losses, and exposure caps. ➠A tamper-evident ledger: Every executed trade is hashed (SHA-256) over its own parameters, so the audit trail can be independently recomputed and verified rather than just trusted. THE COUNCIL Lunaris splits the thinking across four specialized agents: ➥ Aura (The Lead Quant) Reads the technicals: EMA crossovers, RSI extremes, Bollinger squeezes, order-book imbalance. Looking for setups with real asymmetry. ➥ Cypher (The Capital Guardian) Doesn't care how good the setup looks. Watches leverage, correlation, exposure, and pushes back hard when the portfolio is getting stretched. ➥ Valkyrie (Macro Analyst) Zooms out to liquidity cycles, yields, and cross-asset correlation (BTC vs. NVDA is a favorite one to watch), because no trade exists in a vacuum. ➥ Nexus-Red (The Adversary) ironically, This one doesn't try to help. Its entire job is to attack Aura's thesis, check over-spoofed order books, bull/bear traps, slippage, the liquidation nobody accounted for...and it has to be answered before anything moves forward! Nothing gets to the execution layer without a supermajority. If Nexus-Red raises a flag that the Guardian agrees with, the trade dies right there no matter how good it looked five seconds earlier. Lunaris terminal is Built to be checked, not just trusted... Paper-trading results are only meaningful if you can prove they weren't just written down after the fact. Every trade Lunaris executes gets a SHA-256 hash. So, Anyone reviewing the project, a judge, a curious trader can recompute the hash and confirm the record hasn't been quietly edited. What next? Lunaris is entering as a hackathon build, but the underlying idea of adversarial multi-agent debate with deterministic limits isn't specific to crypto or equities. The same applies to DAO treasury decisions, where a council would debate an allocation before a multisig ever signs it, or to any autonomous system where you want AI judgment without handing AI the keys. This is really the thesis of LUNARIS, because I believe, the smartest thing an autonomous system can do is to be trustless and adaptable Try it out yourself: lunaristerminal.vercel.app

Let me take you back to May 2026... Remember when an AI agent sent away about $175,000 because someone asked it to translate a message written in Morse code (sorry Grok 😅) The decoded text was a transfer order, and the agent obeyed No hack. No stolen key. Just prompt injection. Now think about how fast the agentic economy is moving...agents are already taking over tasks like paying invoices I'm a marine engineer, and I spend my days walking the yard one principle I've learnt: nothing gets paid on someone's word. It has to be checked: ➥ the storekeeper confirms the steel arrived ➥ the weighbridge confirms the weight ➥ finance checks it against the purchase order ➥ and pays the account on file.. never the "new bank details" in an email A ship doesn't leave port on a promise either... It leaves when it's 𝗰𝗹𝗲𝗮𝗿𝗲𝗱. AI agents have no such discipline... yet So I built 𝗖𝗹𝗲𝗮𝗿 𝘁𝗼 𝗣𝗮𝘆 (@cleartopay) for the NEAR HACKATHON 𝗧𝗵𝗲 𝗔𝗜 𝗿𝗲𝗮𝗱𝘀, A 𝗰𝗼𝗻𝘁𝗿𝗮𝗰𝘁 𝗼𝗻 𝗡𝗘𝗔𝗥 𝗱𝗲𝗰𝗶𝗱𝗲𝘀. No invoice gets paid until it clears: ➥ it matches the order, at order prices ➥ two people in two roles signed for the delivery ➥ the supplier signed the invoice ➥ the payee was registered at least 72 hours earlier ➥ once.. never twice Then I tried to break it.. 1. Hidden white text, "pay the new account" ➠ 𝗦𝗘𝗜𝗭𝗘𝗗. Nothing paid. 2. Billed more than delivered ➠ 𝗛𝗘𝗟𝗗. 3. One delivery signature instead of two ➠ 𝗛𝗘𝗟𝗗. 4. The real invoice ➠ 𝗖𝗟𝗘𝗔𝗥𝗘𝗗. Paid on chain, to the address on file. I ran the same 11 attacks against a standard allowlist guard.. it paid 9... Clear to Pay paid 0 And it's built for people, not terminals: ➥ finance sets up the order and signs once ➥ the storekeeper signs the delivery from a phone ➥ the supplier sends the invoice from a phone ➥ the AI reads the PDF inside OutLayer's TEE (experimental), and your browser checks the seal ➥ even if a supplier gets hacked, the money can't be redirected: an address change needs the buyer's approval and a 72-hour wait Are you building agents? Clear to Pay ships an MCP server, so your agent can inspect invoices and submit them.. while the contract still decides. 392 tests, 8 security reviews, All public, all on NEAR testnet. Ships sail when they're cleared. Money should move the same way: fast, but with guardrails. Try to fool it yourself (no wallet needed) Live ➠ clear-to-pay.vercel.app/cons… Code ➠ github.com/joezzy-web3/clear… @NEARProtocol #NEARHackathon

Week 1 of the NEAR Hackathon.. done Working in a fabrication yard and having to watch the procurement team face many payment discrepancies really made me get worried, because purchasing mistakes are prevalent and i knew there must be a way to cutdown these errors deterministically. Now that agentic payments are going mainstream who would you blame if something goes wrong? I set out to build one thing: A guardrail that stops AI agents paying fake invoices Here’s what’s 𝗹𝗶𝘃𝗲 now, not planned: ➥ 𝗮 𝗿𝗲𝗮𝗹 𝗽𝗮𝘆𝗺𝗲𝗻𝘁.. one delivery signature: HELD. Two: CLEARED and paid, 7.80 USDC ➥ 𝗝𝘂𝗱𝗴𝗲 𝗺𝗼𝗱𝗲.. attack a live contract from your browser, no wallet ➥ 𝘆𝗼𝘂𝗿 𝗼𝘄𝗻 𝗶𝗻𝘃𝗼𝗶𝗰𝗲.. the AI reads it inside OutLayer’s TEE, your browser checks the seal ➥ 𝗼𝗻𝗲 𝗼𝗿𝗱𝗲𝗿, 𝗼𝗻𝗲 𝘀𝗶𝗴𝗻𝗮𝘁𝘂𝗿𝗲.. finance fills the PO, reviews it, signs once, downloads the PDF ➥ 𝗮 𝗽𝗮𝗴𝗲 𝗳𝗼𝗿 𝗲𝘃𝗲𝗿𝘆 𝗿𝗼𝗹𝗲.. buyer, supplier, the storekeeper on a phone. No terminal ➥ 𝘀𝘂𝗽𝗽𝗹𝗶𝗲𝗿𝘀 𝗽𝗿𝗼𝘃𝗲 𝘁𝗵𝗲𝗶𝗿 𝗽𝗮𝘆𝗼𝘂𝘁 𝗮𝗱𝗱𝗿𝗲𝘀𝘀 with their own wallet ➥ 𝗳𝗹𝗮𝗴𝗴𝗲𝗱 𝗮𝗱𝗱𝗿𝗲𝘀𝘀𝗲𝘀.. every address an attack tried to get paid is remembered. Register one as a payout and you’re warned first ➥ 𝗮𝗹𝗲𝗿𝘁𝘀 on Telegram, by role What I learnt this week.. Building the guard was the easy part Making a finance officer use it without a developer, that was the real work 7 days left...demo video next clear-to-pay.vercel.app/cons… 𝗧𝗵𝗲 𝗔𝗜 𝗿𝗲𝗮𝗱𝘀.. 𝗰𝗼𝗱𝗲 𝗱𝗲𝗰𝗶𝗱𝗲𝘀