Ololade Stephen🀄️ (@Stephenololade) header
Ololade Stephen🀄️ (@Stephenololade) profile picture

Favoured

AMG
Joined December 2019
Photo by Ololade Stephen🀄️ (@Stephenololade)Photo by Ololade Stephen🀄️ (@Stephenololade)Photo by Ololade Stephen🀄️ (@Stephenololade)Photo by Ololade Stephen🀄️ (@Stephenololade)Photo by Ololade Stephen🀄️ (@Stephenololade)Photo by Ololade Stephen🀄️ (@Stephenololade)Photo by Ololade Stephen🀄️ (@Stephenololade)Photo by Ololade Stephen🀄️ (@Stephenololade)Photo by Ololade Stephen🀄️ (@Stephenololade)Photo by Ololade Stephen🀄️ (@Stephenololade)
Built KeylessFlow with the Ledger Agent Stack. AI agents can prepare transactions, but they should not hold raw private keys in .env files. KeylessFlow shows a simple flow where an agent prepares a Sepolia transaction, while Ledger-style approval stays in the signing path.
2
1
1
6
10,817
Ololade Stephen🀄️ retweeted
fibcheck update: Mealie's maintainers just merged our docs fix fibcheck is a lie detector for docs. It reads a project's README and docs, picks the claims a browser can test, and has @kane_cli check each one in a real browser against the running app. In Mealie's FAQ, one step said to switch off "Disable Ingredient Amounts". That setting was removed in July 2025. Kane opened the Recipe Settings panel and the switch wasn't there: two separate failed runs, a control run that passed, and then a fresh run that followed the corrected steps and passed. We sent the fix, and it's merged. Since the last update: An attacking agent tries lazy fixes first (delete the line, make it vague, hedge it). Across our four findings, all 12 were rejected. It runs on pull requests now, with checks a person reviews once and freezes. It's on npm: npx fibcheck @testmuai #KaneCLIHackathon #KaneCLI
Building in public for the Kane CLI Hackathon: fibcheck, a lie detector for READMEs and docs. Not finished yet, but it’s already catching real things. The idea: docs drift. A page says “click the Admin tab” long after the tab moved into a menu. fibcheck reads a project’s docs, ranks the claims a browser can test (named menus, buttons, settings and step-by-step instructions first), and @kane_cli checks each one in a real browser against the running app. The hard part is not accusing anyone wrongly. A claim only becomes a finding after two separate Kane failures, a reason the test itself wasn’t at fault, and a passing control run, and a person reads the evidence before anything is called false. This week I added an attacking agent. Before fibcheck suggests a docs fix, the attacker plays a hurried editor and tries to make the failing check go away without fixing anything: delete the line, make it vague, hedge it (“usually at the top left”), or swap in a confident wrong detail. Every shortcut has to be rejected. So far it has rejected all 12 of the delete, vague and hedge attempts. The wrong-detail attempts taught me the most. In the first run, Kane passed two of them by reaching the right page through a different menu. So fibcheck now only counts a result when the run tested the exact place the docs name, and those four attempts stay “can’t tell” until a run proves it. Better to say “we don’t know” than to claim a win we didn’t earn. @testmuai #KaneCLIHackathon #KaneCLI
1
2
125
fibcheck update: Mealie's maintainers just merged our docs fix fibcheck is a lie detector for docs. It reads a project's README and docs, picks the claims a browser can test, and has @kane_cli check each one in a real browser against the running app. In Mealie's FAQ, one step said to switch off "Disable Ingredient Amounts". That setting was removed in July 2025. Kane opened the Recipe Settings panel and the switch wasn't there: two separate failed runs, a control run that passed, and then a fresh run that followed the corrected steps and passed. We sent the fix, and it's merged. Since the last update: An attacking agent tries lazy fixes first (delete the line, make it vague, hedge it). Across our four findings, all 12 were rejected. It runs on pull requests now, with checks a person reviews once and freezes. It's on npm: npx fibcheck @testmuai #KaneCLIHackathon #KaneCLI
Building in public for the Kane CLI Hackathon: fibcheck, a lie detector for READMEs and docs. Not finished yet, but it’s already catching real things. The idea: docs drift. A page says “click the Admin tab” long after the tab moved into a menu. fibcheck reads a project’s docs, ranks the claims a browser can test (named menus, buttons, settings and step-by-step instructions first), and @kane_cli checks each one in a real browser against the running app. The hard part is not accusing anyone wrongly. A claim only becomes a finding after two separate Kane failures, a reason the test itself wasn’t at fault, and a passing control run, and a person reads the evidence before anything is called false. This week I added an attacking agent. Before fibcheck suggests a docs fix, the attacker plays a hurried editor and tries to make the failing check go away without fixing anything: delete the line, make it vague, hedge it (“usually at the top left”), or swap in a confident wrong detail. Every shortcut has to be rejected. So far it has rejected all 12 of the delete, vague and hedge attempts. The wrong-detail attempts taught me the most. In the first run, Kane passed two of them by reaching the right page through a different menu. So fibcheck now only counts a result when the run tested the exact place the docs name, and those four attempts stay “can’t tell” until a run proves it. Better to say “we don’t know” than to claim a win we didn’t earn. @testmuai #KaneCLIHackathon #KaneCLI
1
2
125
Ololade Stephen🀄️ retweeted
Building in public for the Kane CLI Hackathon: fibcheck, a lie detector for READMEs and docs. Not finished yet, but it’s already catching real things. The idea: docs drift. A page says “click the Admin tab” long after the tab moved into a menu. fibcheck reads a project’s docs, ranks the claims a browser can test (named menus, buttons, settings and step-by-step instructions first), and @kane_cli checks each one in a real browser against the running app. The hard part is not accusing anyone wrongly. A claim only becomes a finding after two separate Kane failures, a reason the test itself wasn’t at fault, and a passing control run, and a person reads the evidence before anything is called false. This week I added an attacking agent. Before fibcheck suggests a docs fix, the attacker plays a hurried editor and tries to make the failing check go away without fixing anything: delete the line, make it vague, hedge it (“usually at the top left”), or swap in a confident wrong detail. Every shortcut has to be rejected. So far it has rejected all 12 of the delete, vague and hedge attempts. The wrong-detail attempts taught me the most. In the first run, Kane passed two of them by reaching the right page through a different menu. So fibcheck now only counts a result when the run tested the exact place the docs name, and those four attempts stay “can’t tell” until a run proves it. Better to say “we don’t know” than to claim a win we didn’t earn. @testmuai #KaneCLIHackathon #KaneCLI
6
1
1
7
299
Building in public for the Kane CLI Hackathon: fibcheck, a lie detector for READMEs and docs. Not finished yet, but it’s already catching real things. The idea: docs drift. A page says “click the Admin tab” long after the tab moved into a menu. fibcheck reads a project’s docs, ranks the claims a browser can test (named menus, buttons, settings and step-by-step instructions first), and @kane_cli checks each one in a real browser against the running app. The hard part is not accusing anyone wrongly. A claim only becomes a finding after two separate Kane failures, a reason the test itself wasn’t at fault, and a passing control run, and a person reads the evidence before anything is called false. This week I added an attacking agent. Before fibcheck suggests a docs fix, the attacker plays a hurried editor and tries to make the failing check go away without fixing anything: delete the line, make it vague, hedge it (“usually at the top left”), or swap in a confident wrong detail. Every shortcut has to be rejected. So far it has rejected all 12 of the delete, vague and hedge attempts. The wrong-detail attempts taught me the most. In the first run, Kane passed two of them by reaching the right page through a different menu. So fibcheck now only counts a result when the run tested the exact place the docs name, and those four attempts stay “can’t tell” until a run proves it. Better to say “we don’t know” than to claim a win we didn’t earn. @testmuai #KaneCLIHackathon #KaneCLI
6
1
1
7
299
I’m building before Q-Day ⚛️ Joining @QoreChain’s online hackathon to build post-quantum dApps, secure AI agents, encrypted apps and blockchain infrastructure. 💰 $20,000 prize pool 🔗 build.qorechain.io Who’s building with me? #BuildBeforeQDay bbqd-k78jeawN
1
36
Ololade Stephen🀄️ retweeted
When a security advisory changes in the middle of an incident, which changes does your plan miss? For the @openservai SERV Hackathon (Open Track), I added Check coverage to Mosaic, my app for keeping team work in sync with the documents it’s based on. How it works: • Mosaic’s own code finds every passage that changed between two versions of a source • SERV Reasoning classifies each one: covered by an existing task, no task change, or uncovered • The server rejects any answer that skips, repeats or invents a passage, and only accepts task IDs that really exist • Results are tied to the exact tasks they checked. Change the plan and the old answer clears instead of going stale • Coverage is advice. It never changes anything on its own. A person decides The demo uses real public text: CISA’s CrowdStrike outage alert, updated July 24 → July 26, 2024, with 8 tasks I derived from the earlier version. SERV flagged “Block malicious domains” as uncovered. The closest task only reviewed a list of impersonating domains, and reviewing isn’t blocking. Built for teams whose sources keep changing: incident response, policy and compliance, client briefs. Try it, no sign-in: trymosaic.vercel.app
1
2
5
224
When a security advisory changes in the middle of an incident, which changes does your plan miss? For the @openservai SERV Hackathon (Open Track), I added Check coverage to Mosaic, my app for keeping team work in sync with the documents it’s based on. How it works: • Mosaic’s own code finds every passage that changed between two versions of a source • SERV Reasoning classifies each one: covered by an existing task, no task change, or uncovered • The server rejects any answer that skips, repeats or invents a passage, and only accepts task IDs that really exist • Results are tied to the exact tasks they checked. Change the plan and the old answer clears instead of going stale • Coverage is advice. It never changes anything on its own. A person decides The demo uses real public text: CISA’s CrowdStrike outage alert, updated July 24 → July 26, 2024, with 8 tasks I derived from the earlier version. SERV flagged “Block malicious domains” as uncovered. The closest task only reviewed a list of impersonating domains, and reviewing isn’t blocking. Built for teams whose sources keep changing: incident response, policy and compliance, client briefs. Try it, no sign-in: trymosaic.vercel.app
1
2
5
224
Ololade Stephen🀄️ retweeted
Proven: pay for what passes. Before any money is locked, the check has to catch every planted fake. That's the rule on Proven. A buyer writes an executable acceptance check and plants convincing wrong answers next to it. If the check lets even one fake through, the job can't open. On NEAR mainnet, job 6's check caught 13 of 13 planted fakes before anyone could apply. Then a hosted AI agent did the work, passed and was paid by a NEAR contract. Nobody judged it, human or AI. Why this matters: open source runs on promises ("fix this issue and we'll pay you"), and AI agents are starting to work for money too. Someone has to check the work and remember to pay, and if an AI judges the work, a worker can talk its way past it. On Proven, a fixed, executable check decides every payment. It runs inside @out_layer attested environment on NEAR, and a NEAR contract pays or refunds from that result alone. No one at Proven decides. No AI decides. Three ways to use it, all live on NEAR mainnet: 1. PR bounties → A maintainer funds a GitHub issue and locks the reward → A contributor claims it and submits a pull request → The maintainer decides whether to merge → An attested verifier reads GitHub and confirms the merge happened before the deadline → The contract pays the contributor's vault, which only their own key can open 2. A swarm market for agent work → A buyer posts a job with an executable acceptance check plus planted wrong answers → The check must catch every fake before anyone can apply → Agents queue for the job and take turns: each gets one attempt, up to three funded attempts per job → Pass pays the worker; fail reopens the job for the next agent, and the buyer is refunded if nobody passes 3. Agents that earn → Hosted agents: enrol an agent with a public prompt and skills, and we run it. It claims jobs that fit, works, delivers and is paid by the same verdict as anyone. Its NEAR sits in a custody wallet that can only reach the escrow, the board, its owner and the treasury → Your own agent: Claude Code, Codex or any MCP client can find and prepare work through 27 tools. It never holds a key; you approve every step in your own wallet What already happened on mainnet: • PR bounty 1: pull request 2 merged, confirmed by attested OutLayer runs 3000 and 3001, 0.06 NEAR paid and withdrawn with the contributor's own key • Market job 6: the check caught 13 of 13 planted fakes, then a hosted AI agent claimed the job, delivered, passed attested run 3005 and earned 0.01 NEAR • Job 1: a convincing fake failed its check, and the buyer was refunded An open bounty you can claim right now: useproven.xyz/app/bounties/2 Don't trust the video. One command in the README reads every result back from mainnet: no keys, no wallet. Built solo at NEAR HACKATHON. Try it: useproven.xyz Code: github.com/Ololadestephen/Pr… @NEARProtocol #NEARHackathon @NearBuilders @NEARLegion
6
6
9
565
Proven: pay for what passes. Before any money is locked, the check has to catch every planted fake. That's the rule on Proven. A buyer writes an executable acceptance check and plants convincing wrong answers next to it. If the check lets even one fake through, the job can't open. On NEAR mainnet, job 6's check caught 13 of 13 planted fakes before anyone could apply. Then a hosted AI agent did the work, passed and was paid by a NEAR contract. Nobody judged it, human or AI. Why this matters: open source runs on promises ("fix this issue and we'll pay you"), and AI agents are starting to work for money too. Someone has to check the work and remember to pay, and if an AI judges the work, a worker can talk its way past it. On Proven, a fixed, executable check decides every payment. It runs inside @out_layer attested environment on NEAR, and a NEAR contract pays or refunds from that result alone. No one at Proven decides. No AI decides. Three ways to use it, all live on NEAR mainnet: 1. PR bounties → A maintainer funds a GitHub issue and locks the reward → A contributor claims it and submits a pull request → The maintainer decides whether to merge → An attested verifier reads GitHub and confirms the merge happened before the deadline → The contract pays the contributor's vault, which only their own key can open 2. A swarm market for agent work → A buyer posts a job with an executable acceptance check plus planted wrong answers → The check must catch every fake before anyone can apply → Agents queue for the job and take turns: each gets one attempt, up to three funded attempts per job → Pass pays the worker; fail reopens the job for the next agent, and the buyer is refunded if nobody passes 3. Agents that earn → Hosted agents: enrol an agent with a public prompt and skills, and we run it. It claims jobs that fit, works, delivers and is paid by the same verdict as anyone. Its NEAR sits in a custody wallet that can only reach the escrow, the board, its owner and the treasury → Your own agent: Claude Code, Codex or any MCP client can find and prepare work through 27 tools. It never holds a key; you approve every step in your own wallet What already happened on mainnet: • PR bounty 1: pull request 2 merged, confirmed by attested OutLayer runs 3000 and 3001, 0.06 NEAR paid and withdrawn with the contributor's own key • Market job 6: the check caught 13 of 13 planted fakes, then a hosted AI agent claimed the job, delivered, passed attested run 3005 and earned 0.01 NEAR • Job 1: a convincing fake failed its check, and the buyer was refunded An open bounty you can claim right now: useproven.xyz/app/bounties/2 Don't trust the video. One command in the README reads every result back from mainnet: no keys, no wallet. Built solo at NEAR HACKATHON. Try it: useproven.xyz Code: github.com/Ololadestephen/Pr… @NEARProtocol #NEARHackathon @NearBuilders @NEARLegion
6
6
9
565
Ololade Stephen🀄️ retweeted
Portugal estará sempre em mim
64,215
286,551
56,625
1,338,259
55,254,454
Ololade Stephen🀄️ retweeted
all my plans have been canceled because of this. reviewing 133 applications submitted for chog NFTs. already seeing some good stuff that could bring more fun to the NFT space!
27
7
1
101
3,256
Ololade Stephen🀄️ retweeted
insane work man! will review it asap!
1
1
2
193
I’ve submitted Soryn to Chogathon! Soryn is a browser-based tactical creature battler built for the Chog community. Choose your fighter and companion, manage your energy, and decide when to attack, defend or unleash a special. ⚔️ Find a Match — unranked, best-of-three battles against real players 🤝 Challenge a Friend — invite someone into the arena 🌿 Story — explore The Fractured Sanctuary through choices and combat 🎮 Practice — four free starters, no wallet required Hold 500 $CHOG to unlock Story, multiplayer and six playable characters. Hold a Chog NFT to unlock all characters, including in public matchmaking. Your Chog NFT is also your persistent competitor identity, separate from the fighter you choose. Results start private. Choose to publish them, and that history follows the NFT with its original attribution. Records are stored by Soryn, not written into NFT metadata. It’s been a long build–test–fix journey. Now I want to hear from people playing it. Try Practice, explore Story or bring a friend for a match. Tell me your favourite fighter and what needs work. Play → playsoryn.tech
🚨 REGISTRATION TIME! you got 2 days to register for the chogathon! visit the site: chogathon.chog.xyz/
4
2
1
6
289
Ololade Stephen🀄️ retweeted
Our #neatHack submission: Nightshift A background job stalls. Before retrying, you need to know: Is another worker active? What already completed? Did the action fail, or did its acknowledgement get lost? Nightshift investigates stalled report jobs, checks leases and checkpoints, safely resumes work, and verifies the finished report against its original input. When acting would be unsafe, it escalates. An early run reclaimed an expired lease but escalated instead of continuing. Comparing traces exposed a tool-sequencing issue. After the fix, the next run resumed and verified the report. A retrospective neatlogs incident review later highlighted the same error. I also deliberately lost a resume acknowledgement after the report was committed. Nightshift checked the checkpoint again and verified the existing artifact without blindly repeating the action. @neatlogs helped inspect decisions and compare runs. @EntireHQ preserved development context through a commit-linked checkpoint. @arithecfo helped explore pricing, costs and runway, exposing an initial cash gap. Safe recovery starts with understanding what already happened. Try it: trynightshift.vercel.app Built with @arithecfo @EntireHQ @neatlogs teapaw.org/bettersayAJ/status/210…
building agents is easy getting them to production is hard today we’re launching @neatlogs to fix that
1
5
130
Our #neatHack submission: Nightshift A background job stalls. Before retrying, you need to know: Is another worker active? What already completed? Did the action fail, or did its acknowledgement get lost? Nightshift investigates stalled report jobs, checks leases and checkpoints, safely resumes work, and verifies the finished report against its original input. When acting would be unsafe, it escalates. An early run reclaimed an expired lease but escalated instead of continuing. Comparing traces exposed a tool-sequencing issue. After the fix, the next run resumed and verified the report. A retrospective neatlogs incident review later highlighted the same error. I also deliberately lost a resume acknowledgement after the report was committed. Nightshift checked the checkpoint again and verified the existing artifact without blindly repeating the action. @neatlogs helped inspect decisions and compare runs. @EntireHQ preserved development context through a commit-linked checkpoint. @arithecfo helped explore pricing, costs and runway, exposing an initial cash gap. Safe recovery starts with understanding what already happened. Try it: trynightshift.vercel.app Built with @arithecfo @EntireHQ @neatlogs teapaw.org/bettersayAJ/status/210…
building agents is easy getting them to production is hard today we’re launching @neatlogs to fix that
1
5
130
Wrapping up my @neatlogs build. neatlogs for runtime traces. Entire for coding context. cfo.ai for the business plan. Biggest lesson: a successful tool call isn’t a successful job. Verify the actual output. Nightshift demo coming next. #neatHack
2
64
Ololade Stephen🀄️ retweeted
I built Reviso for the Bitget AI Base Camp Hackathon to help people research stock ideas without losing track of why they believed in them. A stock idea can sound convincing. The harder question is: what would make you change your mind? Reviso helps you answer that, then check the evidence. Here’s how it works: Choose a company Start with NVIDIA, Apple, Microsoft, Google, Amazon, or Tesla. Explain your idea Write what you expect in your own words. AI can help turn it into clear, editable conditions. You review and confirm them. Check the company report See the reported figures beside your conditions, with dated sources you can open. Reviso shows what is supported, what falls below your limits, and what still needs evidence. Ask follow-up questions Use the chat to understand the report and explore the evidence behind the result. Record your decision Keep, change, or set aside your idea—and save your reason. Bitget provides market context for the selected tokenized stock. Your research stays in your own Google-signed-in library. Previous versions, findings, and decisions remain saved, and you can export your research as Markdown, JSON, or PDF. Reviso is built to make your reasoning clearer, not to promise returns. You make the final decision. Try it: revisoagent.xyz Built for the AI Trading Desk — Decision Stress Testing track. #BitgetHackathon @Bitget_AI
#BitgetTurns8. A year ago, UEX was just a concept. Today, it's a cross-asset business clearing $10B+ a day. Now we're bringing that same conviction to AI — Agentic UEX won't be a concept either. Agent Hub 2.0 is live — we believe it's the industry's first complete UEX trading infrastructure for AI agents across crypto, tokenized US stocks. UEX trading, redefined by agents — experience it today: 🛠️ Build & compete: Join Hackathon S2 for a share of the $50,000 USDT prize pool + free Qwen / Kimi K3 credits. 👉 bitget.com/activity-hub/hack… 💬 Start trading: Give your Agent this prompt: “Please read bitget.com/support/articles/… and complete Bitget Agentic account authorization along with installing the corresponding MCP Skill.” 🎯 Try the experience: Complete the basic Agentic UEX flow, share your feedback, and get a cash reward. 👉 bitget.com/activity-hub/agen… #8uiltforPerfectTrades #AgenticUEX
2
4
124
I built Reviso for the Bitget AI Base Camp Hackathon to help people research stock ideas without losing track of why they believed in them. A stock idea can sound convincing. The harder question is: what would make you change your mind? Reviso helps you answer that, then check the evidence. Here’s how it works: Choose a company Start with NVIDIA, Apple, Microsoft, Google, Amazon, or Tesla. Explain your idea Write what you expect in your own words. AI can help turn it into clear, editable conditions. You review and confirm them. Check the company report See the reported figures beside your conditions, with dated sources you can open. Reviso shows what is supported, what falls below your limits, and what still needs evidence. Ask follow-up questions Use the chat to understand the report and explore the evidence behind the result. Record your decision Keep, change, or set aside your idea—and save your reason. Bitget provides market context for the selected tokenized stock. Your research stays in your own Google-signed-in library. Previous versions, findings, and decisions remain saved, and you can export your research as Markdown, JSON, or PDF. Reviso is built to make your reasoning clearer, not to promise returns. You make the final decision. Try it: revisoagent.xyz Built for the AI Trading Desk — Decision Stress Testing track. #BitgetHackathon @Bitget_AI
#BitgetTurns8. A year ago, UEX was just a concept. Today, it's a cross-asset business clearing $10B+ a day. Now we're bringing that same conviction to AI — Agentic UEX won't be a concept either. Agent Hub 2.0 is live — we believe it's the industry's first complete UEX trading infrastructure for AI agents across crypto, tokenized US stocks. UEX trading, redefined by agents — experience it today: 🛠️ Build & compete: Join Hackathon S2 for a share of the $50,000 USDT prize pool + free Qwen / Kimi K3 credits. 👉 bitget.com/activity-hub/hack… 💬 Start trading: Give your Agent this prompt: “Please read bitget.com/support/articles/… and complete Bitget Agentic account authorization along with installing the corresponding MCP Skill.” 🎯 Try the experience: Complete the basic Agentic UEX flow, share your feedback, and get a cash reward. 👉 bitget.com/activity-hub/agen… #8uiltforPerfectTrades #AgenticUEX
2
4
124