# Jev Atlas rankings

Snapshot: September 18, 2026. Source posts and public metadata reviewed; implementations were not reproduced.

## Method

Same editorial dimensions as Astra Atlas: **Creativity** (specific, worthwhile or unexpected application), **Complexity** (constraints and interacting systems), and **Integration** (handoffs of state, data or intent). Each is scored 1–5. These are judgments of the described workflow, not benchmark results. Overall rank weighs usefulness, reusable architecture, novelty and strength of the available evidence; it is deliberately not a sum of the three scores.

59 entries received editorial reviews; the top 15 are highlighted as picks. The 12 roundup-only projects remain unscored rather than receiving zero. Source text and metadata were reviewed; no complete video, thread or implementation audit is implied.

Reach is separate: primary-post views divided by the author’s follower count at retrieval. It does not measure quality, unique visitors or followers at posting time. No thread views are summed and no best-performing source is substituted. Roundup-only entries do not inherit the roundup’s reach.

## Editorial order

| Rank | Project | Creativity | Complexity | Integration |
|---:|---|---:|---:|---:|
| 1 | [Fraud-email screening with a Kimi fallback](https://x.com/nutlope/status/2100614659690713543) | 4 | 4 | 5 |
| 2 | [Bannerbear template-to-data field mapping](https://x.com/yongfook/status/2100801037192024478) | 4 | 3 | 5 |
| 3 | [Jamcat voice-driven UI](https://x.com/JonPTaylor/status/2100736122502390211) | 5 | 4 | 5 |
| 4 | [YouTube sponsor-skipping extension](https://x.com/tdinh_me/status/2100793777103466615) | 5 | 4 | 5 |
| 5 | [Firstmate production agent dispatch](https://x.com/kunchenguid/status/2100468943853085061) | 4 | 4 | 5 |
| 6 | [Rubik’s Cube with coded solving stages](https://x.com/redp314/status/2100489858951073858) | 5 | 4 | 5 |
| 7 | [Browser Use flight-search agent](https://x.com/gregpr07/status/2100411066966749359) | 4 | 4 | 5 |
| 8 | [Minecraft with Astra planning and Jev reactions](https://x.com/wuyang_zhou/status/2100727660875808913) | 5 | 5 | 5 |
| 9 | [PR risk checks from a diff](https://x.com/redp314/status/2100585126652481915) | 4 | 4 | 5 |
| 10 | [Classify 20,000 emails, Slack messages and transcripts](https://x.com/ZacGawn/status/2100704691869098295) | 4 | 4 | 4 |
| 11 | [Analyze 724 live advertisements](https://x.com/TheMattBerman/status/2100654891756589230) | 4 | 4 | 4 |
| 12 | [Intent-aware file launcher](https://x.com/dabit3/status/2100756930054504776) | 5 | 3 | 4 |
| 13 | [Organize 1,018 AI research papers](https://x.com/nutlope/status/2100426999546184123) | 4 | 3 | 5 |
| 14 | [Natural-language X feed filter](https://x.com/marcelpociot/status/2100520134481735729) | 4 | 3 | 4 |
| 15 | [Discord moderation bot](https://x.com/brainstormity/status/2100471987860553931) | 4 | 4 | 5 |
| 16 | [Audit 3,282 X posts for growth patterns](https://x.com/iannuttall/status/2100668908227162567) | 4 | 4 | 4 |
| 17 | [Jev + AXe iOS simulator control](https://x.com/camsoft2000/status/2100648648434434298) | 3 | 4 | 4 |
| 18 | [Desktop control through agent-desktop](https://x.com/mdlahfir/status/2100359236924637349) | 3 | 4 | 5 |
| 19 | [Voice-controlled browser](https://x.com/moritzkremb/status/2100577979021832365) | 4 | 3 | 4 |
| 20 | [Stagehand browser automation](https://x.com/kylejeong/status/2100622054945095934) | 3 | 4 | 4 |
| 21 | [Agent action-safety classifier](https://x.com/fazxes/status/2100300097695232164) | 3 | 4 | 4 |
| 22 | [Route coding tasks across Claude Code, Codex and OpenCode](https://x.com/mdlahfir/status/2100314182201802811) | 3 | 3 | 4 |
| 23 | [Rank 1,000 posts for a useful AI-agent tip](https://x.com/daniel_mac8/status/2100620339097026633) | 4 | 3 | 4 |
| 24 | [SuperX post scoring](https://x.com/robj3d3/status/2100631889585606959) | 4 | 4 | 4 |
| 25 | [Conversation checks and proactive-message decisions](https://x.com/MichaelLee04/status/2100003037150683593) | 4 | 4 | 4 |
| 26 | [Non-pausing driving simulator](https://x.com/SigGravitas/status/2100325221932958134) | 4 | 5 | 4 |
| 27 | [JevPilot simulated self-driving](https://x.com/jpschroeder/status/2100347770867458384) | 4 | 4 | 4 |
| 28 | [500 agents in a 3D environment](https://x.com/crislenta/status/2100457614073327754) | 4 | 5 | 4 |
| 29 | [Trading bot with real order execution](https://x.com/jarrodwatts/status/2100356151468585346) | 4 | 4 | 5 |
| 30 | [Semantic ad-blocking extension](https://x.com/iam_zachi/status/2100529273186472318) | 4 | 4 | 4 |
| 31 | [Pac-Man with Astra strategy](https://x.com/daniel_mac8/status/2100335929273524541) | 4 | 4 | 5 |
| 32 | [Jev Review coding-agent feedback loop](https://x.com/niazmorshed_/status/2100465662867218857) | 3 | 3 | 4 |
| 33 | [Resume-to-job relevance scoring at HiringCafe](https://x.com/h_nilforoshan/status/2100409794276520341) | 3 | 4 | 3 |
| 34 | [Wikipedia navigation races](https://x.com/_shubhankar/status/2100402012194218231) | 3 | 3 | 4 |
| 35 | [Debate rhetoric meter](https://x.com/chetaslua/status/2100473581251748216) | 4 | 3 | 3 |
| 36 | [Synthetic product-adoption panel](https://x.com/ytiskw/status/2100474943154827344) | 4 | 3 | 3 |
| 37 | [Pictures from parallel pixel decisions](https://x.com/anshuc/status/2100246929611411501) | 5 | 2 | 2 |
| 38 | [Restaurant-kitchen simulation](https://x.com/yoheinakajima/status/2100777327005511709) | 4 | 3 | 3 |
| 39 | [Subway Surfers and parallel game sessions](https://x.com/_MaxBlade/status/2100634359099232678) | 4 | 4 | 4 |
| 40 | [Doom decision loop](https://x.com/CompleteSkeptic/status/2099925687465570372) | 3 | 4 | 4 |
| 41 | [Minecraft on an Orgo cloud computer](https://x.com/nickvasiles/status/2100670497818313175) | 3 | 4 | 4 |
| 42 | [Slay the Spire 2 player](https://x.com/coolish/status/2100570517954838897) | 3 | 4 | 3 |
| 43 | [Tetris player](https://x.com/marcus_lowe/status/2100315518930661861) | 3 | 3 | 3 |
| 44 | [Super Mario Bros. player](https://x.com/faadilhshaik/status/2100086301894881578) | 3 | 3 | 3 |
| 45 | [Request-to-model router](https://x.com/ephraimduncan/status/2100454070536351824) | 2 | 2 | 3 |
| 46 | [Replace an agent’s tool-selection reasoning](https://x.com/oviniciuslana/status/2100457622407168509) | 3 | 3 | 4 |
| 47 | [Security-pipeline classification](https://x.com/grichadev/status/2100437998571860087) | 3 | 3 | 3 |
| 48 | [Replace hard-coded decisions in existing SaaS products](https://x.com/illyism/status/2100489521557123265) | 3 | 2 | 3 |
| 49 | [Blitz chess against frontier models](https://x.com/aimlapi/status/2100372930282573876) | 3 | 3 | 3 |
| 50 | [Flappy Bird model comparison](https://x.com/anshnanda/status/2100611596859093082) | 3 | 3 | 3 |
| 51 | [Four-task comparison with larger models](https://x.com/anshnanda/status/2100717739556167722) | 3 | 4 | 3 |
| 52 | [Live viral-post analyzer](https://x.com/rileybrown/status/2100425868053008758) | 3 | 2 | 3 |
| 53 | [Post-virality simulator with a shared feed](https://x.com/leojrr/status/2100470174130250127) | 3 | 2 | 3 |
| 54 | [Structured trading-signal decision demo](https://x.com/_trou3/status/2100481938016669917) | 3 | 2 | 2 |
| 55 | [Ask Jev judgment playground](https://x.com/waynesutton/status/2100487878992388279) | 3 | 2 | 3 |
| 56 | [Text from a fixed word menu](https://x.com/hi_im_isaac_/status/2100408276949385668) | 4 | 2 | 2 |
| 57 | [Fast tool-history filtering for Claude](https://x.com/tamarajtran/status/2100694549362553153) | 4 | 2 | 2 |
| 58 | [Bulk email classification](https://x.com/rileybrown/status/2100404532119269426) | 2 | 2 | 2 |
| 59 | [Car-parking experiment](https://x.com/theokramer_/status/2100424419877515295) | 3 | 2 | 2 |

## Why the top 15 stand out

### 1. Fraud-email screening with a Kimi fallback

This is the clearest reusable architecture in the collection: cheap screening first, then spend on the cases whose confidence is low. The reported 96/100 result gives a concrete test to inspect instead of only a speed claim.

**Takeaway:** Measure the entire cascade, including fallback cost and remaining errors.

**Evidence limit:** Jev screens 100 emails; predictions below 95% confidence go to Kimi K3. The combined test reportedly scored 96/100 for about $0.07.

### 2. Bannerbear template-to-data field mapping

The narrow scope is the strength. Matching a customer’s fields to a template removes a real setup chore and is already described as live in Bannerbear.

**Takeaway:** Look for small repeated setup decisions where a user can immediately inspect the result.

**Evidence limit:** Match differently named fields such as photo→avatar, name→full_name and company_name→business in one click.

### 3. Jamcat voice-driven UI

A useful division of labor: one model handles the commands while another handles the conversation. The handover is the product experience.

**Takeaway:** Separate action selection from open-ended conversation and make the handover explicit.

**Evidence limit:** Use Pipecat to hand off between Jev for interface commands and PhoneLLM for generated responses.

### 4. YouTube sponsor-skipping extension

This changes behavior at the moment it matters: a judgment about content becomes a player action while the video is running.

**Takeaway:** Align decisions with playback state and provide an obvious way to recover from a wrong skip.

**Evidence limit:** Detect sponsor segments and skip them during playback, with optional audio capture. Open-source BYOK prototype; reported cost about $0.005 per video.

### 5. Firstmate production agent dispatch

A measured production-shaped example. The author compares the full dispatch operation as well as the very cheap Jev call, which is the right boundary for assessing the benefit.

**Takeaway:** Evaluate whole-workflow time and cost rather than only the replacement model call.

**Evidence limit:** Choose a harness, model and reasoning effort from user preferences. The author reports matching Fable on 25 evaluated tasks, with 71% lower dispatch cost and 90% less wall time.

### 6. Rubik’s Cube with coded solving stages

The most instructive puzzle demo: code owns the solving method, Jev recognizes the case, and code checks the choice. It gives the model a well-defined job.

**Takeaway:** Use learned judgments inside a verifiable algorithm rather than asking them to replace the algorithm.

**Evidence limit:** Code implements the beginner method; Jev picks the case at each step and code checks the choice. Reported result: 94 moves and about four seconds of model time.

### 7. Browser Use flight-search agent

Browser actions are a good test of fast decisions because every step changes the valid next choices. The explicit text-generation fallback makes the architecture more believable.

**Takeaway:** Regenerate the action space after each interaction and reserve generation for tasks that need it.

**Evidence limit:** Choose browser actions from DOM state, with a small LLM for typing. The author reports a flight search in seven seconds for $0.0039.

### 8. Minecraft with Astra planning and Jev reactions

The planning/reacting split is easy to grasp here: one model thinks ahead while another handles immediate threats. It is ambitious, although the post alone does not reveal the reliability of the coordination.

**Takeaway:** Define what triggers a new plan and which actions the fast controller may choose.

**Evidence limit:** Pair a slower planner with fast decisions; the demo claim includes fighting multiple zombies.

### 9. PR risk checks from a diff

The interesting part is not producing 14 scores; it is turning them into explicit outcomes and escalating ambiguous critical risks.

**Takeaway:** Keep merge policy in code and measure false approvals separately from unnecessary blocks.

**Evidence limit:** Ask 14 typed questions about secrets, injection, authentication, API changes, tests and other risks. Code routes uncertain critical checks to a person or larger model. Six PRs are reported in the demo.

### 10. Classify 20,000 emails, Slack messages and transcripts

Retagging existing communications is a stronger business case than inventing a new chat interface. The reported completed batch is distinct from the author’s proposed future analyses.

**Takeaway:** Start with questions the business can act on and keep the original records available for retagging.

**Evidence limit:** Tag business records for upsells, complaints, missed follow-ups and other categories. The author reports seven minutes and $1.45. The later call-centre hypothesis workflow is a proposed next use.

### 11. Analyze 724 live advertisements

A useful conversion from an unstructured ad library to comparable features. The next challenge is checking whether labels support better decisions, not merely producing more labels.

**Takeaway:** Validate a small sample of labels before drawing conclusions across the whole collection.

**Evidence limit:** Label hooks, formats, offers, calls to action, awareness stages and landing-page mismatches across 37 brands. Reported run: 40 seconds and $0.09; StealAds/MCP availability was still forthcoming.

### 12. Intent-aware file launcher

The interface becomes sensitive to intent rather than literal filenames. This is an imaginative use of low latency in an everyday action.

**Takeaway:** Pass a bounded candidate list and stabilize results as the user keeps typing.

**Evidence limit:** Rank launcher results as the user types requests such as finding a recently downloaded PDF; the author reports roughly 100ms responses.

### 13. Organize 1,018 AI research papers

A clear multi-model workflow with separate cost accounting. The author also distinguishes running a test from replacing the live site’s classifications.

**Takeaway:** Measure each stage separately and validate new labels before replacing existing ones.

**Evidence limit:** DeepSeek produces summaries; Jev assigns topics from 24 choices. Reported classification cost: $0.08, separate from $3.99 for summaries. The author was still evaluating Jev before replacing the live site’s labels.

### 14. Natural-language X feed filter

A specific personal policy becomes a fast, local interface change. The hard part is protecting useful posts from silent false hides.

**Takeaway:** Make hiding reversible and audit false hides, not just the number of removed posts.

**Evidence limit:** Hide or collapse posts according to a written filtering instruction in a browser extension.

### 15. Discord moderation bot

Moderation needs more than a score. This project describes escalation steps, logging and admin controls around the classifier, making the integration worth inspecting.

**Takeaway:** Treat administrator corrections as evidence to test, not proof that future mistakes are impossible.

**Evidence limit:** Classify messages and apply progressive moderation, audit logs and admin confirmation. The author estimates $0.30/month for 1,000 messages/day; the claim that pardons prevent all repeat errors is unverified.

## Reach order

| Rank | Project | Views | Followers | Views / follower |
|---:|---|---:|---:|---:|
| 1 | [Super Mario Bros. player](https://x.com/faadilhshaik/status/2100086301894881578) | 519,450 | 162 | 3206.48× |
| 2 | [Structured trading-signal decision demo](https://x.com/_trou3/status/2100481938016669917) | 101,482 | 151 | 672.07× |
| 3 | [Car-parking experiment](https://x.com/theokramer_/status/2100424419877515295) | 20,447 | 42 | 486.83× |
| 4 | [Resume-to-job relevance scoring at HiringCafe](https://x.com/h_nilforoshan/status/2100409794276520341) | 284,355 | 881 | 322.76× |
| 5 | [Route coding tasks across Claude Code, Codex and OpenCode](https://x.com/mdlahfir/status/2100314182201802811) | 121,405 | 437 | 277.81× |
| 6 | [Text from a fixed word menu](https://x.com/hi_im_isaac_/status/2100408276949385668) | 256,263 | 1,093 | 234.46× |
| 7 | [Agent action-safety classifier](https://x.com/fazxes/status/2100300097695232164) | 366,251 | 1,724 | 212.44× |
| 8 | [Minecraft with Astra planning and Jev reactions](https://x.com/wuyang_zhou/status/2100727660875808913) | 75,031 | 443 | 169.37× |
| 9 | [Desktop control through agent-desktop](https://x.com/mdlahfir/status/2100359236924637349) | 71,841 | 437 | 164.40× |
| 10 | [PR risk checks from a diff](https://x.com/redp314/status/2100585126652481915) | 225,873 | 1,542 | 146.48× |
| 11 | [Conversation checks and proactive-message decisions](https://x.com/MichaelLee04/status/2100003037150683593) | 487,296 | 3,767 | 129.36× |
| 12 | [Blitz chess against frontier models](https://x.com/aimlapi/status/2100372930282573876) | 447,596 | 3,713 | 120.55× |
| 13 | [Replace an agent’s tool-selection reasoning](https://x.com/oviniciuslana/status/2100457622407168509) | 94,824 | 873 | 108.62× |
| 14 | [Rubik’s Cube with coded solving stages](https://x.com/redp314/status/2100489858951073858) | 161,945 | 1,542 | 105.02× |
| 15 | [Fast tool-history filtering for Claude](https://x.com/tamarajtran/status/2100694549362553153) | 991,081 | 10,657 | 93.00× |
| 16 | [Security-pipeline classification](https://x.com/grichadev/status/2100437998571860087) | 78,866 | 1,168 | 67.52× |
| 17 | [Browser Use flight-search agent](https://x.com/gregpr07/status/2100411066966749359) | 1,951,669 | 29,186 | 66.87× |
| 18 | [Discord moderation bot](https://x.com/brainstormity/status/2100471987860553931) | 22,919 | 607 | 37.76× |
| 19 | [JevPilot simulated self-driving](https://x.com/jpschroeder/status/2100347770867458384) | 455,196 | 13,082 | 34.80× |
| 20 | [Semantic ad-blocking extension](https://x.com/iam_zachi/status/2100529273186472318) | 156,837 | 4,599 | 34.10× |
| 21 | [Tetris player](https://x.com/marcus_lowe/status/2100315518930661861) | 75,212 | 2,455 | 30.64× |
| 22 | [Jev Review coding-agent feedback loop](https://x.com/niazmorshed_/status/2100465662867218857) | 42,369 | 1,419 | 29.86× |
| 23 | [Trading bot with real order execution](https://x.com/jarrodwatts/status/2100356151468585346) | 892,283 | 31,864 | 28.00× |
| 24 | [500 agents in a 3D environment](https://x.com/crislenta/status/2100457614073327754) | 55,874 | 2,145 | 26.05× |
| 25 | [Analyze 724 live advertisements](https://x.com/TheMattBerman/status/2100654891756589230) | 225,677 | 11,971 | 18.85× |
| 26 | [Pictures from parallel pixel decisions](https://x.com/anshuc/status/2100246929611411501) | 283,852 | 15,943 | 17.80× |
| 27 | [Request-to-model router](https://x.com/ephraimduncan/status/2100454070536351824) | 97,114 | 6,690 | 14.52× |
| 28 | [Wikipedia navigation races](https://x.com/_shubhankar/status/2100402012194218231) | 18,593 | 1,282 | 14.50× |
| 29 | [Jev + AXe iOS simulator control](https://x.com/camsoft2000/status/2100648648434434298) | 48,173 | 3,976 | 12.12× |
| 30 | [Doom decision loop](https://x.com/CompleteSkeptic/status/2099925687465570372) | 1,182,258 | 103,428 | 11.43× |
| 31 | [Synthetic product-adoption panel](https://x.com/ytiskw/status/2100474943154827344) | 188,815 | 24,810 | 7.61× |
| 32 | [Subway Surfers and parallel game sessions](https://x.com/_MaxBlade/status/2100634359099232678) | 144,983 | 22,595 | 6.42× |
| 33 | [Classify 20,000 emails, Slack messages and transcripts](https://x.com/ZacGawn/status/2100704691869098295) | 8,561 | 1,678 | 5.10× |
| 34 | [Stagehand browser automation](https://x.com/kylejeong/status/2100622054945095934) | 40,249 | 8,030 | 5.01× |
| 35 | [Post-virality simulator with a shared feed](https://x.com/leojrr/status/2100470174130250127) | 102,294 | 21,916 | 4.67× |
| 36 | [Debate rhetoric meter](https://x.com/chetaslua/status/2100473581251748216) | 152,289 | 34,397 | 4.43× |
| 37 | [Jamcat voice-driven UI](https://x.com/JonPTaylor/status/2100736122502390211) | 4,907 | 1,176 | 4.17× |
| 38 | [Pac-Man with Astra strategy](https://x.com/daniel_mac8/status/2100335929273524541) | 85,672 | 30,502 | 2.81× |
| 39 | [Flappy Bird model comparison](https://x.com/anshnanda/status/2100611596859093082) | 28,634 | 10,679 | 2.68× |
| 40 | [Slay the Spire 2 player](https://x.com/coolish/status/2100570517954838897) | 152,650 | 64,664 | 2.36× |
| 41 | [Firstmate production agent dispatch](https://x.com/kunchenguid/status/2100468943853085061) | 84,272 | 35,901 | 2.35× |
| 42 | [Voice-controlled browser](https://x.com/moritzkremb/status/2100577979021832365) | 144,315 | 71,892 | 2.01× |
| 43 | [Organize 1,018 AI research papers](https://x.com/nutlope/status/2100426999546184123) | 145,619 | 100,074 | 1.46× |
| 44 | [Ask Jev judgment playground](https://x.com/waynesutton/status/2100487878992388279) | 78,771 | 68,073 | 1.16× |
| 45 | [SuperX post scoring](https://x.com/robj3d3/status/2100631889585606959) | 64,612 | 60,892 | 1.06× |
| 46 | [Non-pausing driving simulator](https://x.com/SigGravitas/status/2100325221932958134) | 46,832 | 49,460 | 0.95× |
| 47 | [Natural-language X feed filter](https://x.com/marcelpociot/status/2100520134481735729) | 60,633 | 70,317 | 0.86× |
| 48 | [Four-task comparison with larger models](https://x.com/anshnanda/status/2100717739556167722) | 8,530 | 10,679 | 0.80× |
| 49 | [Bulk email classification](https://x.com/rileybrown/status/2100404532119269426) | 183,526 | 243,851 | 0.75× |
| 50 | [Minecraft on an Orgo cloud computer](https://x.com/nickvasiles/status/2100670497818313175) | 9,629 | 15,959 | 0.60× |
| 51 | [Fraud-email screening with a Kimi fallback](https://x.com/nutlope/status/2100614659690713543) | 42,374 | 100,074 | 0.42× |
| 52 | [Intent-aware file launcher](https://x.com/dabit3/status/2100756930054504776) | 74,984 | 194,208 | 0.39× |
| 53 | [Rank 1,000 posts for a useful AI-agent tip](https://x.com/daniel_mac8/status/2100620339097026633) | 11,405 | 30,502 | 0.37× |
| 54 | [Audit 3,282 X posts for growth patterns](https://x.com/iannuttall/status/2100668908227162567) | 29,690 | 81,019 | 0.37× |
| 55 | [Replace hard-coded decisions in existing SaaS products](https://x.com/illyism/status/2100489521557123265) | 9,300 | 28,815 | 0.32× |
| 56 | [Live viral-post analyzer](https://x.com/rileybrown/status/2100425868053008758) | 49,743 | 243,851 | 0.20× |
| 57 | [Restaurant-kitchen simulation](https://x.com/yoheinakajima/status/2100777327005511709) | 6,484 | 126,220 | 0.05× |
| 58 | [YouTube sponsor-skipping extension](https://x.com/tdinh_me/status/2100793777103466615) | 6,785 | 201,677 | 0.03× |
| 59 | [Bannerbear template-to-data field mapping](https://x.com/yongfook/status/2100801037192024478) | 2,632 | 171,758 | 0.02× |