Developer Archetypes
Unsupervised clustering of 9,896 top GitHub developers by 22 behavioral features — SkillBench Research Lab, UCSB
Finding the Natural Groupings
We extracted 22 features from each developer's public GitHub profile — value dimensions, activity patterns, skill breadth, AI-era signals, and identity metadata — then ran k-means clustering to find natural groupings. Silhouette analysis peaks at k=4.
9,896 Developers in 2D
PCA projection of the 22-dimensional feature space. Clusters overlap in 2D because the real separation lives in higher dimensions — but the structure is visible.
Cluster Value Profiles
Radar charts showing each cluster's average score across six value dimensions, compared to the global mean.
High pre-AI history (78%), active cadence, broad collaboration. The backbone of open source.
| Login | Name | Followers | Annual Contribs | Pre-AI | Reviews |
|---|---|---|---|---|---|
| torvalds | Linus Torvalds | 287,983 | 3,264 | 87% | 2 |
| karpathy | Andrej | 135,647 | 859 | 69% | 6 |
| yyx990803 | Evan You | 107,039 | 1,902 | 94% | 17 |
| gaearon | dan | 90,422 | 836 | 78% | 42 |
| ruanyf | Ruan YiFeng | 85,507 | 685 | 97% | 0 |
Deep pre-AI foundations (91%) but sporadic current activity. Built things, then stepped back.
| Login | Name | Followers | Annual Contribs | Pre-AI | Reviews |
|---|---|---|---|---|---|
| gustavoguanabara | Gustavo Guanabara | 109,928 | 8 | 57% | 0 |
| peng-zhihui | 空心 | 86,188 | 0 | 97% | 0 |
| rafaballerini | Rafaella Ballerini | 59,052 | 0 | 94% | 0 |
| tj | TJ | 51,571 | 2 | 100% | 0 |
| IDouble | Alp ₿📈🚀🌕 | 50,276 | 24 | 95% | 0 |
Low pre-AI ratio (23%), newer accounts (~5yr), moderate activity. Became visible after AI coding tools emerged.
| Login | Name | Followers | Annual Contribs | Pre-AI | Reviews |
|---|---|---|---|---|---|
| dmalan | David J. Malan | 35,793 | 71 | 0% | 16 |
| elyxdev | Elyx | 26,763 | 63 | 0% | 0 |
| lucast1574 | Lucas Santillan | 21,120 | 377 | 0% | 0 |
| OracleBrain | Aashis Jha | 15,368 | 32 | 7% | 0 |
| george0st | jist | 13,508 | 1,109 | 0% | 0 |
High follower counts but near-zero activity, languages, and domains. Famous but not building publicly.
| Login | Name | Followers | Annual Contribs | Pre-AI | Reviews |
|---|---|---|---|---|---|
| claude | Claude | 46,597 | 0 | 100% | 0 |
| cumsoft | cumsoft | 18,410 | 0 | 100% | 0 |
| metatimeofficial | Metatime | 18,023 | 0 | 0% | 0 |
| PremChapagain | Prem Chapagain | 14,157 | 4 | 33% | 0 |
| ghost | Deleted user | 12,145 | 0 | 0% | 0 |
Sub-Archetypes: AI-Era Arrivals
The AI-Era Arrivals cluster contains developers who joined or became active primarily after AI coding tools. Sub-clustering reveals three distinct sub-types.
Weekly cadence, highest quality in AI-Era group, some collaboration. Genuinely productive — but started in the AI era.
| Login | Name | Followers | Annual Contribs | Pre-AI | Reviews |
|---|---|---|---|---|---|
| george0st | jist | 13,508 | 1,109 | 0% | 0 |
| jrohitofficial | Rohit Jha | 12,721 | 648 | 35% | 0 |
| XiaomingX | Y11 | 11,465 | 3,702 | 0% | 0 |
| LaurieWired | LaurieWired | 11,006 | 208 | 0% | 0 |
| elder-plinius | pliny | 10,922 | 477 | 0% | 0 |
Accounts under 3 years, near-zero pre-AI commits, no code reviews. High social signal, low work-product signal.
| Login | Name | Followers | Annual Contribs | Pre-AI | Reviews |
|---|---|---|---|---|---|
| elyxdev | Elyx | 26,763 | 63 | 0% | 0 |
| lucast1574 | Lucas Santillan | 21,120 | 377 | 0% | 0 |
| pewdiepie-archdaemon | PewDiePie | 13,278 | 130 | 0% | 0 |
| elidianaandrade | Eli | 12,423 | 253 | 5% | 0 |
| meliksahyorulmazlar | meliksahyorulmazlar | 10,711 | 69 | 0% | 0 |
Some pre-AI history (62%) but activity declined. Had foundations, went quiet.
| Login | Name | Followers | Annual Contribs | Pre-AI | Reviews |
|---|---|---|---|---|---|
| dmalan | David J. Malan | 35,793 | 71 | 0% | 16 |
| OracleBrain | Aashis Jha | 15,368 | 32 | 7% | 0 |
| 996icu | 996icu | 8,566 | 4 | 100% | 0 |
| JCSIVO | JCSIVO | 7,512 | 26 | 73% | 0 |
| U7P4L-IN | U7P4L x C0D3R | 7,455 | 2 | 100% | 0 |
Sub-Archetypes: Seasoned Builders
The largest cluster splits into two recognizable types: individual contributors vs. community hubs.
Deep individual contributors, strong foundations, low review/collaboration activity. Domain experts in their lane.
| Login | Name | Followers | Annual Contribs | Pre-AI | Reviews |
|---|---|---|---|---|---|
| torvalds | Linus Torvalds | 287,983 | 3,264 | 87% | 2 |
| karpathy | Andrej | 135,647 | 859 | 69% | 6 |
| yyx990803 | Evan You | 107,039 | 1,902 | 94% | 17 |
| gaearon | dan | 90,422 | 836 | 78% | 42 |
| ruanyf | Ruan YiFeng | 85,507 | 685 | 97% | 0 |
Highest collaboration breadth (30+), active code reviewers, daily cadence. Framework authors and open source maintainers.
| Login | Name | Followers | Annual Contribs | Pre-AI | Reviews |
|---|---|---|---|---|---|
| alesanchezr | Alejandro Sanchez | 2,068 | 12,605 | 86% | 122 |
| punkpeye | Frank Fiegel | 1,709 | 7,995 | 2% | 11 |
Complete Taxonomy
| Archetype | N | % | Quality | Cadence | Pre-AI | Age | Top Exemplar |
|---|---|---|---|---|---|---|---|
| Solo Veterans | 4,116 | 41.6% | 0.76 | 3.1 | 78% | 13y | torvalds |
| Ecosystem Leaders | 2 | 0.0% | 0.61 | 3.5 | 44% | 10y | alesanchezr |
| Pre-AI Veterans (Quiet) | 4,037 | 40.8% | 0.68 | 1.1 | 91% | 12y | gustavoguanabara |
| Active AI-Native Builders | 514 | 5.2% | 0.51 | 2.8 | 14% | 5y | george0st |
| Minimal Evidence Newcomers | 415 | 4.2% | 0.36 | 1.7 | 2% | 3y | elyxdev |
| Faded Practitioners | 339 | 3.4% | 0.40 | 1.1 | 62% | 6y | dmalan |
| Minimal Evidence Profiles | 473 | 4.8% | 0.17 | 0.5 | 31% | 5y | claude |
Key finding: Among the top 10,000 GitHub users by follower count, 12.8% became prominent primarily in the AI era. Within that group, a third show minimal evidence of engineering depth — high social signal with low work-product signal. From public artifacts alone, these profiles are indistinguishable from legitimate developers who simply work in private repos.
The Telemetry Gap
This analysis shows what work products can tell you — and where they hit a wall.
Behavioral archetypes from public signal
Distinct patterns of quality, consistency, collaboration, and AI-era adaptation cluster reliably across ~10,000 profiles.
Decoupled social proof from evidence
High follower counts don't guarantee engineering depth. The data shows where these signals diverge.
Who actually wrote the code
Commits show who pushed, not who authored. A single commit may be 90% AI-generated or 100% handwritten — the git log looks identical.
Learning velocity
A developer who went from zero to competent in 6 months looks the same as one who plateau'd 5 years ago. Static snapshots miss trajectories.
AI delegation patterns
Did they use AI for boilerplate and write the hard parts themselves? Or prompt-engineer the entire feature? GitHub doesn't know.
Skill in context
Knowing Python as a software engineer is different from knowing Python as a business analyst. Repos show the what, not the how or why.
SkillBench closes this gap by instrumenting the development environment itself — turning the process of coding into a legible signal. These archetypes are the boot block. Telemetry turns them into living profiles.