I read other people's code until I find the crack. Scientific Python, ML training paths, and spec test suites — the places where a bug stays quiet because the output still looks correct. 38 merged upstream PRs across sktime, SageMath, JSON Schema, NVIDIA NeMo and others, plus 73 issues filed upstream.
Currently studying machine learning — the theory, the libraries, and the LLM tooling on top. Reading production ML code until something looks wrong turns out to be a fast way to learn it.
| Project | PRs | Focus |
|---|---|---|
| sktime / skpro | 6 merged · 5 open | Probabilistic regression internals — MultiIndex alignment in quantile prediction, and type validation that never rejected anything. Latest: #1034 — [DOC] Fix broken link in dependencies.rst |
| SageMath | 5 merged · 1 open | Matroid isomorphism segfaults, PARI matrix conversion on zero-column input, docstrings and docs across the org. Open: #42752 — Validate the documented parameter domains of the generalized Tamari constructors |
| Tracer Cloud | 3 merged | AI-SRE toolkit — Alertmanager alerts and silences as report evidence, provider retyping, Telegram approval dispatch. Latest: #5762 — feat(alertmanager): map alerts and silences into report evidence |
| NVIDIA NeMo | 10 merged · 8 open | Dataset tokenization paths where batching options silently no-op, plus agent docs that ship CI-failing instructions. Latest: #4456 — fix(speculative): save the effective config in draft checkpoints |
| OpenTelemetry | 1 merged | Docs infrastructure and team-link integrity across members-only orgs. Latest: #10457 — docs: add "(members-only)" to team links in README |
| JSON Schema | 11 merged | Spec test-suite coverage that landed in one draft and was never backported. Latest: #1220 — Check every subschema for unknown keywords in the unified checker |
| Sakana AI | 1 open | Private evaluation aborting when a failed run is followed by an identical solution. Open: #6 — Fix mnist_pool: honor kernel_size in forward/backward tasks |
| c2siorg/Scan8 | 4 open | docs: fix documentation links returning 404 · fix: point the worker HPA at the worker and use millicores for cpu Open: #149 — docs: fix documentation links returning 404 |
| c2siorg/Codelabz | 4 open | fix: write notification_id when the notification is created · fix: scope the notifications page read to the signed-in user Open: #466 — fix: write notification_id when the notification is created |
| PrithvijitBose/Proofly | 1 merged | Security: authenticate profile publish against GitHub identity Latest: #19 — Security: authenticate profile publish against GitHub identity |
| mynextproject-portfolio/github-flow-onboarding-kabirvashisht4-glitch | 1 merged | feat: personalize greeting with my name Latest: #2 — feat: personalize greeting with my name |
| agentregistry-dev/agentregistry | 1 open | docker: run the server image as a non-root user Open: #703 — docker: run the server image as a non-root user |
| Cloud-CV/EvalAI | 1 open | fix #4870: add hover effect to .text-light-gray class Open: #4900 — fix #4870: add hover effect to .text-light-gray class |
Live snapshot of open pull requests in upstream repos (auto-refreshed).
- NVIDIA-NeMo/Automodel#4171 — ci: lint the PR merged into its target branch
- agentregistry-dev/agentregistry#703 — docker: run the server image as a non-root user
- SakanaAI/robust-kbench#6 — Fix mnist_pool: honor kernel_size in forward/backward tasks
- NVIDIA-NeMo/Automodel#3918 — fix(loss): stop masked losses from overwriting the caller's labels
- c2siorg/Scan8#149 — docs: fix documentation links returning 404
- NVIDIA-NeMo/Automodel#3804 — fix(loss): make ChunkedCrossEntropy reductions independent of chunk_len
- NVIDIA-NeMo/Automodel#3797 — fix(loss): return zero, not NaN, when a batch has no supervised tokens
- NVIDIA-NeMo/Automodel#3789 — fix(models): declare and honor logits_to_keep on Mistral3 VLM forward
- c2siorg/Codelabz#466 — fix: write notification_id when the notification is created
- sagemath/sage#42752 — Validate the documented parameter domains of the generalized Tamari constructors
- c2siorg/Codelabz#464 — fix: scope the notifications page read to the signed-in user
- c2siorg/Scan8#145 — fix: point the worker HPA at the worker and use millicores for cpu
- NVIDIA-NeMo/Automodel#3775 — fix(launcher): honor --nproc-per-node over the local device count
- NVIDIA-NeMo/Automodel#3768 — fix(training): honor max_steps when num_epochs is unset
- c2siorg/Codelabz#462 — fix: show the newest notifications instead of an arbitrary fifty
- NVIDIA-NeMo/Automodel#4455 — Speculative recipes save the pre-override config, so auto-resume skips their own checkpoint
- json-schema-org/JSON-Schema-Test-Suite#1213 — regex: only one valid pattern is tested, so a checker that rejects alternation or quantifiers passes
- sagemath/sage#42935 — Meson subprojects tracking upstream HEAD (mpfi, singular, flint) break builds without any change in Sage
- NVIDIA-NeMo/Automodel#4170 — ci: a PR that is green against a stale base can turn main red and fail other contributors' PRs
- json-schema-org/JSON-Schema-Test-Suite#1209 — regex: only one syntax error is ever rejected, so a bracket counter passes the whole file
- NVIDIA-NeMo/Automodel#4130 — Default uv.lock pins torchao 0.14.0, so peft raises ImportError on LoRA injection
- NVIDIA-NeMo/Automodel#3948 — nemo_run executor overrides silently ignore keys the executor has no field for
- NVIDIA-NeMo/Automodel#3933 — default_collater reshapes the caller's example tensors from [S] to [1, S]
- NVIDIA-NeMo/Automodel#3917 — Masked CE losses overwrite the caller's labels in place, zeroing a reused batch's loss
- json-schema-org/JSON-Schema-Test-Suite#1197 — idn-email: the domain half is never checked — no draft has a single invalid-domain case, unlike email
- Widgeto — one coding streak across every platform you use. Zero-dependency streak engine, web onboarding, Flutter app. Live
- Neiro — local-first audio player and visualiser. Decoded, equalised and rendered entirely in the browser tab: no account, no upload, no server.
- Earth 2.0 Explorer — pulls the NASA Exoplanet Archive and filters 5,000+ exoplanets down to those 60–100% similar to Earth. Live
- SIEM prototype — multi-threaded security event simulation, correlating brute-force attempts and malware downloads into a live attack campaign.
- nextjs-eslint-plugin-crash-repro — minimal reproduction proving four
@next/eslint-plugin-nextrules throw a TypeError instead of linting.
Python · JavaScript · TypeScript · Node · Express · MongoDB · React · Next.js · PyTorch · scikit-learn · pandas · NumPy · Dart · Flutter · Docker · Git · GitHub Actions
Auto-updated 2026-10-23 01:44 UTC · workflow