01 · Roasts
Burst-and-Ghost Developer
565 commits this year but weeks 1–30 are basically a tundra. Your GitHub looks like a seismograph — flatline, flatline, EARTHQUAKE (ds4 launch), flatline again. Real programmers commit on Tuesdays in February too.
63% Graveyard Curator
108 public repos and 63% are abandoned. That's 68 digital ghost towns. At some point the portfolio stops being 'prolific' and starts being a museum of half-shipped ideas you'd rather not talk about at conferences.
Solo Artist, No Band
soloPct = 100% across every single repo. 28,277 followers hang on your every commit and you've never once let them open a PR that mattered. Redis built a community; antirez apparently builds a monastery.
Three-Week Framework Guy
ds4: created May 6, last push May 27 — 12,000 stars in 21 days. Impressive. Also, llama.cpp-deepseek-v4-flash: 30 commits in 2 days then silence. The world is watching you ship one specific model family on one specific chip config. Very narrow hallway for a legend.
Tests? Only When The Compiler Cries
linenoise: no tests. gguf-tools: no tests. bme680: no tests. bplog: no tests. ZOT actually has 288 unit tests — a beautiful anomaly. The rest of the portfolio runs on vibes and ZEXALL blessings.
Built using
Zoral
Shadows one worker for a week, then takes over their job with zero extra setup. Behaves exactly like the original.
zoral.ai
02 · Category breakdown
- Impact25% weight83A
- Consistency20% weight65C
- Quality20% weight77B
- Depth15% weight75B
- Breadth10% weight65C
- Community10% weight75B
03 · Stats
365-day commit heatmap
100 active days
Language distribution
- C++40%
- C31%
- Python7%
- Cuda5%
- Objective-C3%
- HTML3%
- Other11%
04 · Numbers
Owned repos
non-fork
93
Commits
last 12 months
565
Followers
28,277
Joined GitHub
Mar 2009
05 · Top repos
antirez /
ds4
Specialized DeepSeek V4 inference engine in C with Metal/CUDA backends; 12k+ stars, production-ready with comprehensive testing, typed code, and detailed architecture. Narrow focus (one model family) limits ecosystem impact vs. broader frameworks.
antirez /
linenoise
Widely-adopted ~1600-LOC C library providing minimal readline alternative; used in Redis, MongoDB, Android. Well-documented README, clean API design, extensive UTF-8/emoji support, but lacks automated tests and CI.
antirez /
ZOT
Clean-room Z80/Spectrum/CP/M emulator written in C with zero stdlib deps. 60 stars, ~6500 LOC, shipped with types, tests, CI, comprehensive docs. Demo of modern indie emulation craft and AI-assisted development.
antirez /
gguf-tools
C library + CLI for GGUF file manipulation with show/compare/split-mixtral tools. Well-structured codebase with mmap-based file parsing, detailed tensor handling, and practical ML model inspection utilities for the llama.cpp ecosystem.
antirez /
llama.cpp-deepseek-v4-flash
Experimental DeepSeek v4 Flash fork of llama.cpp with 2-bit quantization support. Ships with docs/, design.md, ARCHITECTURE.md; HAS_README, HAS_TESTS, HAS_CI; 317MB codebase; fresh burst of 30 commits in 1 day (created Apr 26, last push Apr 27)
antirez /
bme680-pure-mp
MicroPython BME680 sensor driver: clean, typed untyped but structured single-file implementation with comprehensive API documentation in README. Minimal scope (10 KB, 1 file), 5 recent commits, no tests/CI. Solves real hardware interfacing problem.
antirez /
bplog
New reverse-engineered BLE client for Hilo/Aktiia blood pressure cuff. Minimal adoption (16 stars), thin output (64 KB), one-week sprint with comprehensive protocol docs and working Python + Android implementations.
06 · Timeline
- Mar 21, 2009Joined GitHub
- Mar 20, 2010Created linenoise — A small self-contained alternative to readline and libedit
- Dec 24, 2023Created gguf-tools — GGUF implementation in C as a library and a tools CLI program
- Feb 27, 2024Created bme680-pure-mp — Pure MicroPython Bosch BME680 sensor driver
- Feb 23, 2026Created ZOT — ZOT is a Z80, ZX Spectrum 48k and CP/M emulator
- Mar 22, 2026Created bplog — Hilo (TM) / Aktiia (TM) open source client implementation for the bluetooth cuff
- Apr 26, 2026Created llama.cpp-deepseek-v4-flash — Experimental implementation of DeepSeek v4 flaash in llama.cpp
- May 6, 2026Created ds4 — DeepSeek 4 Flash local inference engine for Metal and CUDA
- May 27, 2026Most recent push to ds4
07 · Compare
08 · Rubric
How this score was produced
Overall = Σ (category × weight) + gentle top-end curve
Tier thresholds
▸ How the pipeline works
- 01Scrape.Pull every non-fork repo pushed in the last 90 days, plus your contribution calendar, followers, and language byte counts — straight from GitHub's REST & GraphQL APIs.
- 02Triage.A small model reads every repo's file tree + README and picks the 20 files per repo that actually reveal how you code.
- 03Grade each repo. All repos run in parallel through a fast scoring model that reads the picked files and rates each one independently on Impact, Quality, and Depth — with evidence citations.
- 04Aggregate. A larger reasoning model combines the per-repo scores with server-computed stats (heatmap, commit cadence, language entropy, follower count) to produce the 6-dimension profile score + roasts.
- 05Correct.Deterministic server-side checks enforce anchor-scale floors (e.g. a profile with 2,000+ public commits can't score 30 Consistency) and recompute the final verdict.
~90 seconds per profile, ~$0.25 in compute. Total of ~240 files read across your top-12 repos. One rating per GitHub account per day.
▸ Data sources & caveats
- Heatmap & commit totals: GitHub GraphQL
contributionsCollection— covers the last 365 days, includes private repos when the user has opted in (default). - Language %: byte totals across the top 30 owned non-fork repos.
- Curve: a small upward nudge centered on raw score ≈ 70, capping at 100. Prevents specialists from being unfairly penalised for narrow breadth.
- Anchor corrections: when server-measured signals (e.g. privateWorkLikely, multiRepoVolume, follower count) mandate a minimum category score, the aggregation step enforces it. These are signal-conditional, not identity-based floors.