𝕆𝕙-𝕁𝕒π•ͺ (@Oh_Jay0) header
𝕆𝕙-𝕁𝕒π•ͺ (@Oh_Jay0) profile picture

crypto x ai builder, devrel. building @contri_ng. content partner @raydium

Joined December 2015
Photo by 𝕆𝕙-𝕁𝕒π•ͺ (@Oh_Jay0)Photo by 𝕆𝕙-𝕁𝕒π•ͺ (@Oh_Jay0)Photo by 𝕆𝕙-𝕁𝕒π•ͺ (@Oh_Jay0)Photo by 𝕆𝕙-𝕁𝕒π•ͺ (@Oh_Jay0)Photo by 𝕆𝕙-𝕁𝕒π•ͺ (@Oh_Jay0)Photo by 𝕆𝕙-𝕁𝕒π•ͺ (@Oh_Jay0)Photo by 𝕆𝕙-𝕁𝕒π•ͺ (@Oh_Jay0)Photo by 𝕆𝕙-𝕁𝕒π•ͺ (@Oh_Jay0)Photo by 𝕆𝕙-𝕁𝕒π•ͺ (@Oh_Jay0)Photo by 𝕆𝕙-𝕁𝕒π•ͺ (@Oh_Jay0)
your stablecoin earns nothing while corporate credit markets pay real dividends. companies send cash to shareholders every month, none of it reaches onchain holders. @apyx_fi turns that cash into onchain yield, live now on solana. let's see how digital credit actually works.πŸ”»
89
13
131
9,107
the part of receipt i care about most is what it refuses. on pull request #4 i fed it a cheating fix that edits the test instead of the bug. the guardrail blocked it and nothing got pushed. a fixer that makes tests pass by weakening them is worse than no fixer. full demo: youtu.be/mU3JYazU7XM github.com/ohjay-official/re… built with @kane_cli by @testmuai #KaneCLIHackathon #KaneCLI
25
3
29
419
𝕆𝕙-𝕁𝕒π•ͺ retweeted
This is so cool. Love that it blocks the fixer from touching the tests.
kane cli caught my demo shop charging $29 at checkout for a $19 plan, a bug i planted on purpose, and it caught it on the preview before anything got merged. the same run caught a launch10 coupon that took nothing off and a pay button you could not reach on a 390px phone layout. previews get merged because they look fine. nobody clicks through checkout on every pull request, so a wrong total ships and a customer finds it first. so i built receipt, a github action. when a pull request opens it waits for the vercel preview, then kane cli runs a plain english test contract against it. if kane catches something, a gemini fixer can patch the shop code and nothing else, because it is fenced to shop/** with at most 3 attempts. every fix makes a new preview, kane checks it again, and only then does the pull request get a receipt comment with the verdict for each step and a link to the evidence. the proof is public. on pull request #3 kane went red on run 43698574333, the fixer pushed commit 186ecdc that touched only prices.ts and checkout.ts, and kane went green on run 43699955024. all three bugs were fixed on the first of 3 attempts. the part i care about most is what it refuses. on pull request #4 i scripted a cheating fix that edits the test instead of the bug. the guardrail refused it on run 43700368892 and nothing got pushed. a fixer that makes tests pass by weakening them is worse than no fixer, so receipt blocks any change to the tests, the contract, the workflow or its own code. the first commit went in on nov 15 at 4:33 am, and by about 11 that morning the action worked and all five proof pull requests were up. i built it from my phone on $0, free tiers only. after that i ran a 94 test ship audit on it, and all 13 code fixes landed 35 minutes after the audit started. the hero video went from 10.4mb to 517kb. don't merge a preview without a receipt. github.com/ohjay-official/re… receiptci.vercel.app built with @kane_cli by @testmuai #KaneCLIHackathon #KaneCLI
1
3
19
kane cli caught my demo shop charging $29 at checkout for a $19 plan, a bug i planted on purpose, and it caught it on the preview before anything got merged. the same run caught a launch10 coupon that took nothing off and a pay button you could not reach on a 390px phone layout. previews get merged because they look fine. nobody clicks through checkout on every pull request, so a wrong total ships and a customer finds it first. so i built receipt, a github action. when a pull request opens it waits for the vercel preview, then kane cli runs a plain english test contract against it. if kane catches something, a gemini fixer can patch the shop code and nothing else, because it is fenced to shop/** with at most 3 attempts. every fix makes a new preview, kane checks it again, and only then does the pull request get a receipt comment with the verdict for each step and a link to the evidence. the proof is public. on pull request #3 kane went red on run 43698574333, the fixer pushed commit 186ecdc that touched only prices.ts and checkout.ts, and kane went green on run 43699955024. all three bugs were fixed on the first of 3 attempts. the part i care about most is what it refuses. on pull request #4 i scripted a cheating fix that edits the test instead of the bug. the guardrail refused it on run 43700368892 and nothing got pushed. a fixer that makes tests pass by weakening them is worse than no fixer, so receipt blocks any change to the tests, the contract, the workflow or its own code. the first commit went in on nov 15 at 4:33 am, and by about 11 that morning the action worked and all five proof pull requests were up. i built it from my phone on $0, free tiers only. after that i ran a 94 test ship audit on it, and all 13 code fixes landed 35 minutes after the audit started. the hero video went from 10.4mb to 517kb. don't merge a preview without a receipt. github.com/ohjay-official/re… receiptci.vercel.app built with @kane_cli by @testmuai #KaneCLIHackathon #KaneCLI
37
3
1
40
663
𝕆𝕙-𝕁𝕒π•ͺ retweeted
I think one of the most obvious yet slept-on ways to change your life over the next year is becoming the guy people go to for objective AI model comparisons. maybe I'm wrong, but there's still no real "Marques Brownlee" of AI models - and even if there is, there's 100% room for more. best name I could think of for X is @AlexFinn - but that's it. 10 scrolls on Instagram as a young tech person and you'll probably find 3 videos telling you the "best AI for …", but so much of it feels like people regurgitating whatever ChatGPT or Claude told them rather than actually using these models extensively and comparing them from real user experience. and you can niche down endlessly. if I was a motion designer trying to understand how AI could make me better at my job, the easiest click of my life would be following someone who constantly tests the best models, workflows and tricks specifically for producing excellent motion designs. same for developers. researchers. writers. students. designers. literally anything. I genuinely think whoever becomes trusted for this type of work is sitting on the frontline of what could be legendary content creation over the next few years. the reason I'm writing all of this though is because I want to try. and unfortunately, I need help. this is probably one of the most embarrassing things I've ever written. I feel ashamed even typing it because I've experienced firsthand what it feels like to try and try and still not get the result. after a while, you begin looking down on everything you've ever done. my life really started falling apart after an incident where I was kidnapped and extorted N4.6M by Nigerian police officers (Sept 26 last year) this happened roughly a day after I received money from the Plasma airdrop - money I had honestly earned being one of the biggest voices for stablecoins in my region at the time something about me changed psychologically after that. I entered this phase of thinking: "if I can hustle for over a year and someone can just take half of it from me, why shouldn't I spend the rest on myself before something else takes it?" I still invested some of what remained into MegaETH, but far less than I would have with a clearer head and more capital. you'd think it ends there but roughly two months later, another disaster happened.
55
120
9
310
15,594
on @supabase, two different bugs can both look like "the app returned nothing", and they need different fixes. supabase tightened its defaults this year. new public tables aren't exposed to the data api automatically anymore, so a missing grant now fails loudly with 42501 and postgres never runs your policies. a wrong rls policy still fails silently: no error, every row filtered, or an update that changes 0 rows. that's easy to mix up, for people and for ai agents acting on someone's behalf. so i built policy trace, a lab that shows whether a postgres row was stopped by a grant or filtered by a policy. you pick a scenario, switch who you are, edit the policy, and the trace explains which grant or policy decided each row. in the first lab, alice's select fails with 42501 and the trace says grant layer, not policy. tap the silent empty select chip and the same query succeeds with no error, every row filtered, and the trace says policy layer. the other four labs cover an ai agent using alice's login, a using (true) policy leaking rows, a silent update and team access. it uses the publishable key only, never runs arbitrary sql, and rolls back every run. the final qa round ran 150 runs against the real database and every one matched, and 44 automated tests run on every push. no signup, any run can be shared as a read only link, and it's an independent, mit licensed community project. repo: github.com/ohjay-official/po… start here: policy-trace-gray.vercel.app…
87
6
78
6,174
one thing the video does not linger on: 42501 and an empty result are not the same bug. 42501 means the role has no grant on the table, so postgres stops before any policy runs. an empty result with no error means the grant was there and a policy filtered every row. same looking app bug, two different fixes. the first lab shows both, starting as alice: policy-trace-gray.vercel.app…
78
𝕆𝕙-𝕁𝕒π•ͺ retweeted
I started clipping on Content Rewards, and my first clip pulled 56K+ views. The campaign pays $1 per 1,000 views, so that’s already $50.78 in rewards from one clip. I’m currently sitting at #2 on the leaderboard. If you’ve been looking for a reason to start clipping, this is your sign. Get clipping on Content Rewards.
People keep arguing about cameras at conventions, but the problem isn't the cameras. It's how you use them. Stop, let her pose, take the shot, and nobody has a problem. Walk by at hip height and skip all of that, and you've got one. You're allowed to film in public. That doesn't make every angle okay. #CosplayAndCameras #FilmingEtiquette
117
59
354
62,416
𝕆𝕙-𝕁𝕒π•ͺ retweeted
1
3
3
53
𝕆𝕙-𝕁𝕒π•ͺ retweeted
Official Submission for the Near Hackathon. THE PROBLEM Every trade on Hyperliquid is public: which wallet holds it, how big it is, and the exact price that wipes it out. Hunters push the market to that price, your trade gets wiped out, and the price bounces back. On 13 November, 61 of the last 91 liquidations over $100K on Hyperliquid came straight back within the hour. So I built @unhuntable_xyz HOW UNHUNTABLE FIXES IT You place one trade, say $35 of ETH. Unhuntable then: 1. Splits it into small pieces, each opened from its own new Hyperliquid account. 2. Funds those accounts from your private balance on NEAR, at random times and sizes, so nothing links them to you or to each other. 3. Keeps your stop off Hyperliquid. It is sealed in your browser and held in a secure enclave, which closes your pieces only when independent prices confirm your stop was hit. So a hunter looking at Hyperliquid sees a few small strangers, no big position and no stop. There is nothing worth pushing the price for. It is live on near mainnet now. In the demo video: my real $35 ETH trade goes in, and Hyperliquid sees two small strangers and no stop. βœ…οΈCopy trading, without showing your trades: β–ͺ️Good traders can now sell their trading. A copy key copies every trade into the follower's own accounts, at their own size and timing. A viewing key shows closed trades only. β–ͺ️Followers pay by the day in private transfers. Nobody, followers included, ever sees an open position, so following a trader can never expose them. Every trader's record is signed inside an enclave over every account they used, so no losing trade can be hidden. βœ…οΈBuilt for AI agents too. An agent gets its own key with limits, trades the same hidden way, and can never move money or read its stop back. Crypto on Hyperliquid, and US stocks like NVIDIA on trade.xyz. βœ…οΈEleven parts of NEAR, each doing a real job: - Chain Signatures: every account is controlled through NEAR. No private key exists to steal, not even ours. - NEAR Intents: your money sits in a private balance and funds each account with no trail. - OutLayer enclave: holds your sealed stop and fires it only when independent prices agree. - NEAR AI Cloud: an attacker AI hunts your own accounts in a secure enclave and signs its answer. - x402 on NEAR: agents pay per call in USDC, settled before the call runs. PingPay lets them pay from other chains. - 1Click: deposit from Ethereum, Base, Arbitrum, Solana or NEAR. - Aurora Intents: a second private way in. - OutLayer VRF: random sizes and timing anyone can check. - OutLayer wallets with onchain policies: your NEAR account sets where money can go, not our server. - Sputnik DAO: a team's account moves only on a passed vote. - NEAR wallets: sign in and approve with Meteor, HOT or MyNearWallet. We never hold your money. Your 12 words alone can always bring your money home. Open a real trade yourself, no wallet needed: tryunhuntable.site/judge Code: github.com/Jennycruzy/unhunt… Traders: what would you hide first, your stop or your size? Built at NEAR HACKATHON @NEARProtocol @NEARLegion @NearBuilders #NEARHackathon
16
7
53
1,791
𝕆𝕙-𝕁𝕒π•ͺ retweeted
walk through a full circle in the demo at contri.ng/demo, no wallet needed. when you’re ready, start a real one at contri.ng. if anything confuses you, reply here. we read every one.
3
3
4
96
𝕆𝕙-𝕁𝕒π•ͺ retweeted
I just woke up to a -99% loss after copytrading a whale’s wallet And this is a mistake many crypto traders make We blindly follow profitable wallets without understanding the data behind their trades Only to end up losing! Here’s how to do it right with @LerovaTech πŸ”» #LEROVA
36
27
1
86
1,578
𝕆𝕙-𝕁𝕒π•ͺ retweeted
before the lock we ran a full circle on mainnet with real money on two phones. both pots paid out and both deposits came back. the contract also has 98 unit tests and five review rounds, with every finding and fix published in the repo.
2
3
3
66
𝕆𝕙-𝕁𝕒π•ͺ retweeted
Stock Tokens on Robinhood Chain trade 24/7. Their official price feeds don't. When the US market closes, the Chainlink 24/5 feeds freeze, nobody can mint or redeem until Monday, and on a first-come-first-served sequencer whoever is fastest sets the price. So I built Tutup ("closed" in Indonesian): sealed-bid batch auctions for the hours the market is shut, bonded in $WIF. How it works: 1. Commit: seal a hash of your quantity and price, plus a $WIF bond 2. Reveal: open the bid after commits close, and the bond comes back. Never reveal and it burns 3. Settle: the highest bids fill first and every winner pays one clearing price Every listing burns 100 $WIF. What's different: - Your bid receipt is drand-timelocked inside the commit calldata, so a lost salt no longer costs you your bond - Every sale writes its price on chain behind a Chainlink-compatible feed. WIF/WETH and NVDA/WETH are live - A keeper agent bids, reveals and settles by itself, woken by a Quicknode Stream - Tutup is an MCP server and ERC-8004 agent #25584, so AI agents can bid too - A real NVDA lot is up for auction this weekend Built for @WIFonRH's World is Flat hackathon on Robinhood Chain, with @Quicknode (RPC, Stock Token API, Streams, ERC-8004 API), @getblockio, @chainlink, @RelayProtocol, @Lighter_xyz and @Uniswap. Live on mainnet: tutup.vercel.app Code: github.com/PugarHuda/tutup
4
1
1
8
607
𝕆𝕙-𝕁𝕒π•ͺ retweeted
yo @grok , when saying "please" to an AI. do I waste tokens?
2
1
8
541
𝕆𝕙-𝕁𝕒π•ͺ retweeted
The 2026 Hackathon Boom: Why Shipping Fast Isn’t Enough Anymore A few years ago, showing up to a weekend hackathon meant competing against 5 or 10 other teams. If you managed to deploy a working link by Sunday afternoon, you had a real shot at the podium. Today, those same events are pulling 200 to 300+ submissionsβ€”and the competition has gotten brutal. One veteran builder recently called 2026 a "black year" after entering more than 15 hackathons without a single win. Passion for building has never been higher. But the rules of winning have completely changed. Why Every Hackathon Suddenly Has 300+ Entries Two massive shifts turned cozy weekend code sprints into hyper-competitive arenas: β€’ The collapse of the technical barrier: With AI code editors, autonomous agents, and "vibe coding" workflows, you no longer need a four-person engineering team to ship a full-stack app in 48 hours. Solo designers, PMs, and first-time builders are shipping functional MVPs over a weekend, multiplying submission counts by 10x to 30x. β€’ Aggressive ecosystem sponsorships: AI labs, cloud providers, and Web3 protocols have shifted huge chunks of their marketing budgets into developer bounties. With five- and six-figure prize pools, API credits, and recruiting pipelines on the line, every major online hackathon now draws a global crowd. The "Black Year" Effect: Why Veterans Are Losing When 250+ projects land in a judge’s queue over a single weekend, two things happen: 1. Judge fatigue sets in fast. Complex backend architectures get overlooked if the value isn't obvious in the first 30 seconds. Meanwhile, dozens of flashy, surface-level AI wrappers flood the feed and dilute attention. 2. A working prototype is now the floor, not the ceiling. Shipping fast used to be a moat. Now it’s just the entry ticket. Veterans are no longer competing on coding speedβ€”they’re competing on product polish, distribution, and storytelling. The New Playbook for Standing Out Builders who are still consistently winning in 300-entry pools have narrowed their strategy down to three rules: β€’ Build one bulletproof end-to-end flow: Stop trying to cram five half-baked features into a weekend build. Scope down to a single user workflow with a sharp "before and after." One reliable, zero-bug journey beats a sprawling prototype every time. β€’ Prioritize storytelling over technical readmes: Judges care less about your stack and more about why the problem matters. A crisp, 2-minute demo video showing real utility beats dense technical documentation. β€’ Show proof of real demand: Hackathons increasingly reward early validationβ€”actual signups, user feedback, or live transactionsβ€”over theoretical pitch decks. Case Study: The $100K "Profit Over Prototypes" Sprint Nowhere is this shift clearer than in Marc Lou and Jack Friks’ $100K Hackathon (run under their Ship or Die community and sponsored by Higgsfield). Instead of judging weekend code, the 30-day sprint (running November 16 to December 16, 2026) bans existing projects and judges builders on one metric: net profit. Participants have to launch a brand-new software or mobile app from zero and verify real revenue via TrustMRR and manual audits to compete for $100,000 in cash, an iPhone Duo, and a MacBook Neo. Even the launch reflected the tension of the 2026 builder boom. Lou initially set a $99 entry fee to deter spam and fake revenue numbersβ€”only to reverse course within 24 hours after pushback from indie builders in developing economies and beginners trying to make their first dollar online. He refunded everyone, made entry free, and watched signups surge past 1,600 builders ahead of kickoff. The Hackathon Meta: Then vs. Now 2022–2024 β€’ Pool Size: 10–50 teams per event β€’ Bottleneck: Writing enough code in time β€’ What won: A working prototype β€’ Format: Local venues & niche portals 2026 β€’ Pool Size: 200–1,500+ solo builders β€’ Bottleneck: Standing out from 300 similar AI apps β€’ What wins: Sharp UX, distribution & real revenue β€’ Format: Free global sprints with $100K+ prizes The barrier to building has dropped to zeroβ€”which means the barrier to winning just moved to everything after the code is written.
5
1
6
847
𝕆𝕙-𝕁𝕒π•ͺ retweeted
a circle’s name and members are nobody’s business. names are encrypted on your phone, seats carry no names, and your passkey makes a separate near account for each circle, so nothing on chain links one member’s circles together.
1
2
3
56
𝕆𝕙-𝕁𝕒π•ͺ retweeted
built dioramic, a Hermes plugin that turns scene descriptions into vertical 3D clips using headless Blender. describe a scene, and it renders. want to make a change? just say β€œmake the cube blue” and it re-renders with the update. from text to 3D, with natural language as the interface. github.com/faze39/dioramic @NousResearch
21
7
29
1,752
𝕆𝕙-𝕁𝕒π•ͺ retweeted
Say no to flat videos 🫡🏾 make objects move with Keyframes. Mma’s Editing Tutorials ep. 3! 🫢🏾
Mma’s Editing Tutorials ep. 2!☺️ Showing you how to recreate this effect I did in my @Kimi_Moonshot video (check my highlights if you haven’t already seen it) Save for later and share, you’re welcome!
31
27
1
158
2,533
𝕆𝕙-𝕁𝕒π•ͺ retweeted
Ask an AI to close a land survey and you'll get a confident guess. So I built survey-desk for the Nous Research Hermes Agent Plugin Hackathon, a plugin that gives Hermes a proper survey desk that does the maths exactly. I'm a final-year Surveying and Geoinformatics student, so I know how a traverse goes. You walk the boundary, write down a bearing and a distance for every side, and the numbers have to agree with each other. When they don't close, you end up hunting for the one value you wrote wrong, and that's exactly the kind of arithmetic a language model shouldn't be trusted with. So you paste in a field book and survey-desk computes the traverse, adjusts the misclosure, works out the area two different ways and checks that both answers match, then draws the plan with every bearing and distance on it. It exports the plan as an image and a PDF, a DXF for CAD, a KML you can open in Google Earth and GeoJSON for GIS. The part I like most is what happens when the numbers are wrong. In the video I swap two digits in one distance, 87.05 becomes 78.05, and the traverse stops closing by about 9 metres. survey-desk points at leg D-E and says the distance should probably be 87.087. The real value was 87.05, so it landed within 4 centimetres. I tried five typical typos like that on the demo field book and it named the right leg every time. It labels these as ranked candidates and not proof, because parallel legs can fool it, so you re-measure before you change anything. I also wanted it to be honest about its limits. When it converts from the Minna datum to WGS84 it tells you the published datum shift is around 15 metres, so you know how far to trust the spot on Google Earth, and it warns you if your points sit outside the zone you picked. For now it handles closed traverses in metres only, the default tolerance of 1:5000 is a placeholder you should set to whatever your authority requires, and every plan says it's a computing aid and not a licensed survey plan. Under the hood it's four tools, a /survey command and a skill, with 38 tests passing and no network calls, and it passes hermes plugins validate and doctor. The catalog PR is open and the code and README are here: github.com/ferzerz5-lab/surv… github.com/NousResearch/herm… @NousResearch
2
3
4
213
𝕆𝕙-𝕁𝕒π•ͺ retweeted
someone in every circle needs money before their turn comes. in contri, turns are nft seats, so you can buy an earlier turn from a member who is willing to wait. the circle stays whole and the timing gets flexible.
2
2
4
52