Skip to content
View SID-6921's full-sized avatar
🎯
Focusing
🎯
Focusing

Highlights

  • Pro

Block or report SID-6921

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
SID-6921/README.md

Hero banner

Profile views

Biomedical intelligence systems for real clinical use.

About

I am currently pursuing a Master's in Biomedical Engineering.

My work centers on biomedical signal intelligence, diagnostic modeling, and production-ready health-tech workflows. Cybersecurity was a strong undergraduate hobby and still informs my approach to reliability, privacy, and resilient deployment.

Research Philosophy

  • Build beyond prototypes: prioritize deployability from day one.
  • Keep models clinically interpretable, not only statistically strong.
  • Design for reliability, reproducibility, and patient-context constraints.

Current Directions

  • Biomedical signal processing for diagnostics and monitoring
  • Machine learning pipelines for healthcare prediction tasks
  • Embedded and VLSI-assisted sensing architectures
  • Reproducible health-tech experimentation and evaluation

Academic Trajectory

  • Now: advancing biomedical ML workflows for practical clinical utility
  • Next: translating validated prototypes into robust healthcare tools

Open Source Contributions

I find bugs by reading code and comparing sibling functions, not by picking issues off a tracker. I reproduce each fix below locally before submitting it.

Snapshot: 30 repos where I am a credited contributor through merged commits, plus 1 co-authored credit counted separately below. 74 merged PRs. ~57 open PRs under review.

📈 Trending this run: google/elephant-goldfish-model gained +41 ⭐ since the last check (now 21.1k ⭐ total). Recomputed daily from a real snapshot, not a guess.

Contributor repos (merged)

MakazhanAlpamys/Soup ⭐ 10.6k 🔥 +3 today — LLM fine-tuning CLI

13 merged PRs. Security work covers SSRF predicate consolidation. Training-reliability work covers checkpoint resume, failure-boundary widening, and nonce-based worker verification. Data-path work covers format validation and stripe-root re-validation on every shard write.

  • Impact: The SSRF fix (#625) closed a loopback and private-host bypass across the whole request path, not one call site. The stripe-recheck fix (#1663) turned a silent data-corruption window into a surfaced, refused failure, through a maintainer-found volume-mount-point edge case I reproduced and fixed live.
  • Scope: The largest body of work here by PR count. Several of the 13 are substantial test and fixture reworks.
  • Severity: 4/5. The SSRF fix is a real exploit path, not a style issue. The stripe-recheck fix prevents silent data corruption, not a mere crash.

Project-MONAI/MONAI ⭐ 8.8k — medical imaging DL framework

#8956 merged: fixed a divide-by-zero in the pydicom affine computation for single-slice volumes.

  • Impact: Single-slice DICOM series, like localizers and scouts, are routine in practice. Before this fix, loading one silently crashed instead of producing a usable affine.
  • A second PR (#9134, open) fixes MeanIoU's ignore_index to mask voxels instead of zeroing a whole channel. It disagreed with compute_dice() on identical input, 1.0 versus 0.667, and had shipped untested since PR #8757.
  • Severity: 4/5. Single-slice DICOM series are routine, not an edge case. The crash hit real clinical data.

alphaXiv/OpenResearch ⭐ 8.2k 🔥 +6 today — research-paper reading platform

#351 merged: added a full Hindi (hi) locale, +1,413/−229 across 6 files.

  • Impact: Covers UI strings end-to-end instead of a partial translation. It was the largest locale contribution on the repo at the time.
  • A follow-up cleanup (#585, open) removes a statusColor helper that drifted out of sync with StatusBadge, confirmed unused through a repo-wide grep including wildcard re-export paths.
  • Severity: 2/5. This is a feature gap, not a correctness bug. The product was unusable in Hindi, not broken.

nipy/nibabel ⭐ 810 — neuroimaging I/O library

2 merged PRs. #1553 generalized rescale_affine() to non-4x4 affines. It had been hardcoded to affine[:3,:3] despite being documented for arbitrary (N,N). #1562 added shape validation to Nifti1Header.set_sform.

  • Impact: #1553 fixed a function that raised on any non-square-4 input despite its own docstring promising general support, a gap between the docs and the implementation.
  • Found both through sibling-function comparison: rescale_affine versus voxel_sizes. One generalized to N-d, the other did not.
  • Severity: 3/5. A hard crash on documented-valid input, limited to users already outside the common 4x4-affine case.

neuralinkcorp/datarepo ⭐ 208 — data-table/query library

2 merged PRs, both first-pass clean. #75 added missing is null/is not null filter support on ParquetTable, already present for Clickhouse and Delta. #76 replaced a bare dict-subscript KeyError with a clear error for unsupported ROAPI partition column types.

  • Impact: #75 closed a feature gap between three backends meant to share one filter contract. #76 turned an opaque internal exception into an actionable error.
  • A third PR (#78, open) fixes unescaped SQL interpolation in a codegen path, an injection-shaped bug in generated code.
  • Severity: 2/5 for #75 and #76, a missing feature and an opaque error, annoying but not dangerous. The open #78 rates 4/5 on its own once merged, since unescaped SQL interpolation is injection-shaped.

InsightSoftwareConsortium/ITK ⭐ 1.7k — medical image processing toolkit

#6932 merged same day as opened: array_view_from_vnl_vector aliased the deep-copy function instead of the documented view/no-copy function.

  • Impact: Silently broke a documented no-copy contract. Callers relying on in-place mutation through the view got a disconnected copy, with no error raised.
  • Two more opened the same week: #6936 fixes a TypeError in transform_from_dict for composite, multi-step transforms. #6937 fixes image_from_xarray() assigning origin and spacing to the wrong axes on 4D images.
  • Why this repo matters: ITK started as a US National Institutes of Health (NIH)-funded project in 1999. It is now a standard medical-imaging toolkit in academic and clinical research worldwide, including groups across the US and India. A silent bug here does not stay theoretical. It sits underneath someone's real imaging pipeline.
  • Severity: 4/5. No error raised means no one knows it happened. A silent broken contract is worse than a crash.

keras-team/keras ⭐ 64.4k 🔥 +1 today

#23860 merged: ops.ndim returned a symbolic placeholder for dynamic-batch Functional-model inputs, crashing circle/CircleLoss with a cryptic backend error.

  • Impact: Broke a documented loss function for an entire class of models. Any dynamic-batch Functional model hits this, not an edge case.
  • Two earlier PRs (#23215, path-traversal hardening, and #23216, container weight-path stabilization) were reviewed and closed unmerged.
  • Severity: 4/5. Not an edge case. Every dynamic-batch Functional model hitting this loss function would crash.

unstablebuild/rune ⭐ 1.8k 🔥 +2 today — GPU-accelerated IDE and terminal multiplexer

#219 merged: BuildPluginArgv's sh -c fallback joined task args with a plain space, so an empty-string argument collapsed into nothing and silently shifted every later argument one position left.

  • Impact: Reachable from ordinary task config. internal/ide/idetask builds a task's command directly from user-authored Args []string with no guard against an embedded empty string, so a config with one empty arg ran the wrong command with no error.
  • Severity: 3/5. A silent argument shift a user would blame on their own config rather than a crash, but reachable from normal usage, not a contrived input.

MRtrix3/mrtrix3 ⭐ 359 — MRtrix3 provides a set of tools to perform various advanced diffusion MRI analyses, including constrained spherical deconvolution (CSD), probabilistic tractography, track-density imaging, and apparent fibre density

3 merged PRs, each a distinct silent-correctness bug in a different CLI tool. #3450 fixed an off-by-one bounds check in toggle_axis() that indexed one past the vector instead of growing it. #3449 fixed mrfilter silently swapping the user's -zupper/-zlower threshold values. #3448 fixed phaseencoding raising a wrong "unrecognized" error on a numeric axis specifier documented as valid for dwifslpreproc -pe_dir.

  • Impact: #3449 silently inverted a user-supplied cleaning threshold whenever upper and lower differ, with no error. #3448 rejected documented-valid CLI input with a misleading error message.
  • Severity: 3/5. Two of the three are silent wrong-behavior bugs rather than crashes, each scoped to one specific CLI path.

NVIDIA/Model-Optimizer ⭐ 5.6k 🔥 +2 today — A unified library of SOTA model optimization techniques like quantization, distillation, pruning, neural architecture search, speculative decoding, etc. It compresses deep learning models for downstream deployment frameworks like TensorRT-LLM, TensorRT, vLLM, etc. to optimize inference speed.

3 merged PRs in quantization and distillation. #2645 fixed MFTLoss raising on the ordinary (batch, seq_len, vocab) logit shape a language model produces, despite its own docstring allowing it. #2583 fixed a negative ONNX DequantizeLinear axis, legal per the ONNX spec, being mis-handled when transposing a per-channel-quantized weight. #2554 fixed weight-quantization conversion crashing when a Transpose sits between DequantizeLinear and its consumer, as in MaxViT's attention blocks.

  • Impact: #2583 matters most. On a square weight, a negative axis converted silently to a model that runs but computes the wrong output (max diff 4.09 in the PR's own check), with nothing to flag it. #2645 blocked Minifinetuning-style distillation (arXiv:2506.15702) from working at all on real LM-shaped tensors.
  • Severity: 4/5. #2583's silently-wrong-output case on a routine shape, a square weight, is the highest-risk kind of bug in a quantization library, since nothing downstream would catch it.

galaxyproject/galaxy ⭐ 1.9k — Data intensive science for everyone.

#23058 merged: malformed matched-input tool requests now raise a clear RequestParameterInvalidException at the API/tool-expansion boundary instead of reaching async job prep.

  • Impact: Per issue #22884, malformed matched-input requests were reaching Celery async job prep and surfacing as noisy internal errors instead of a clean request-validation error.
  • Severity: 2/5. Turns an opaque internal error into a clear one. Does not change what's ultimately rejected.

google/elephant-goldfish-model ⭐ 21.1k 🔥 +41 today — Open-source plugin for designing and reviewing software with agents

#439 merged: the bot's own ping-abuse guard wraps bare @user/@org/team mentions in backticks before posting a PR-review comment. A stray unpaired backtick inside a fenced code block bypassed it, since the stray backtick paired across the fence with an unrelated one and let a real @mention sitting between them slip through un-neutralized.

  • Impact: A crafted PR comment, prose mentioning a shell backtick for example, got the bot to ping someone irrelevant to the PR, defeating the guard's purpose.
  • Severity: 3/5. A real bypass of an explicit anti-abuse guard, limited to unwanted notifications rather than data exposure.

jordan-gibbs/hyperresearch ⭐ 4.9k 🔥 +6 today — Convert Claude Code or Codex into the most intelligent Deep Research Agent. Collect, search, and synthesize web research into a persistent, searchable wiki that builds on itself. Hosted API + MCP: hyperresearch.ai

2 merged PRs. #202 fixed the citation-density ship-gate counting every bracketed reference in a report's own Sources/References listing, not only its body. #131 vectorized semantic_search's per-candidate cosine scoring into a single numpy matmul, per the maintainer's own ask in #59.

  • Impact: #202 meant a report whose body made claims with few or no real citations still cleared the citation-density floor purely because its bibliography had many numbered entries, the exact failure mode the gate exists to catch.
  • Severity: 3/5 for #202, a quality gate silently passing what it exists to catch. 2/5 for #131, pure performance with no behavior change.

nilearn/nilearn ⭐ 1.5k — Machine learning for NeuroImaging in Python

#6709 merged: clean_img reshaped its output using the raw sample_mask argument's shape instead of clean()'s actual output shape.

  • Impact: Crashed with ValueError: cannot reshape array whenever a boolean sample_mask was passed without mask_img, a combination clean()'s own docs describe as valid.
  • Severity: 3/5. A hard crash on documented-valid input, scoped to one specific argument combination.

pydicom/pydicom ⭐ 2.2k 🔥 +1 today — Read, modify and write DICOM files with python code

2 merged PRs. #2428 mirrored DSfloat's NaN-equality fix onto DSdecimal, so two NaN-valued DS elements compare equal under config.use_DS_decimal = True too. #2382 fixed apply_voi()/apply_modality_lut() casting LUT indices into a too-narrow dtype before clipping, so out-of-range values wrapped instead of being clamped.

  • Impact: #2428 broke structural dataset-equality checks (ds == deepcopy(ds)) for the decimal backing class specifically. #2382 made an 8-bit LUT with 4096 entries, the routine way to display 12-bit data on an 8-bit display, silently return wrong pixel values with no warning.
  • Severity: 4/5. #2382 is silently-wrong image data on a routine clinical-display path, not an edge case. #2428 is a narrower equality-comparison gap.

scverse/scanpy ⭐ 2.6k — Single-cell analysis in Python. Scales to >100M cells.

#4401 merged: Ingest fell back to a hardcoded settings.N_PCS (50) whenever the reference adata's stored neighbor params recorded neither use_rep nor n_pcs, exactly the state pp.neighbors() leaves with its own defaults.

  • Impact: Whenever more than 50 PCs were computed, for example sc.pp.pca(adata, n_comps=100), Ingest silently built its KNN index in a different, truncated representation than the one that produced the reference neighbor graph. No error, only a quietly wrong embedding and label mapping for sc.tl.ingest.
  • Severity: 4/5. Silent wrong output on a default, common code path, found through sibling-branch comparison with the function that computes the real representation.

udgamlabs/gargi ⭐ 3 — Gargi: a 384M-parameter English language model trained from scratch on ~17B tokens. Docs, runnable examples and checksums; weights on Hugging Face.

#1 merged: the Ollama quick-start comment said thinking was off by default, the opposite of what the model's own Modelfile and example script document.

  • Impact: A new user following the README as written would run the wrong command and have no documented way to disable thinking.
  • Severity: 1/5. Documentation-only, no code behavior changed.

udgamlabs/hinglish-commands-0.6b ⭐ 0 — English and Romanized Hinglish commands to structured JSON. A small Qwen3 adapter, reproducible benchmarks, and a local demo.

#1 merged: validate_model_card() dynamically loaded a helper script that isn't present in the repo or its history, failing with a raw FileNotFoundError from inside importlib.

  • Impact: Using the documented --model-card-dir flag crashed with an opaque internal error instead of a clear, actionable one.
  • Severity: 1/5. Error-message clarity only, does not change what's ultimately rejected.

udgamlabs/shruti ⭐ 13 — Shruti: 29.5 MB speech-to-text for Hindi, English and Hinglish that runs offline on a phone, a laptop or in the browser. Built on Cactus Whistle. Docs, runnable examples and checksums; weights on Hugging Face.

#1 merged: the browser demo's getModel() only checked a 4-byte magic number on a Cache Storage hit, skipping the SHA-256 check that the Python/shell paths always run.

  • Impact: The model URL points at a mutable main ref, so a returning browser user with the old file cached would keep silently running a stale, mismatched model with no warning if the upstream model were ever replaced.
  • Severity: 2/5. A real silent-staleness gap, conditional on the upstream model changing.

udgamlabs/udgam-384m ⭐ 0 — Udgam-384M: a 384M-parameter English model for on-device tool calling, trained from scratch. Docs, runnable examples and checksums; weights on Hugging Face.

#1 merged: transformers_tool_call.py was the one of three tool-calling examples that executed the model's chosen tool name without checking it against an allowlist first, unlike its two siblings.

  • Impact: An example meant to demonstrate safe tool-calling practice itself skipped the "never trust the model's choice blindly" guard the other two examples use.
  • Severity: 2/5. Example and demo code, not a library path, but the kind of omission that matters once copied into real usage.

udgamlabs/vemana ⭐ 9 — Vemana: a 278M Telugu language model trained from scratch, by Udgam Labs

#1 merged: llama_greedy and llama_ppl ran their llama.cpp subprocess without checking its exit status, unlike their near-identical sibling llama_ids.

  • Impact: A crashed or missing llama.cpp binary silently produced empty output, which the caller then reported as a spurious greedy-text or perplexity regression, indistinguishable from an actual model problem.
  • Severity: 2/5. A false-signal diagnostics bug that wastes debugging time rather than causing data loss.

NeuroTechX/moabb ⭐ 1.1k — Mother of All BCI Benchmarks

#1361 merged: MotorImagery.used_events collected every event in a dataset regardless of n_classes, while its sibling FilterBankMotorImagery.used_events correctly stopped once it reached n_classes.

  • Impact: A user who set n_classes on MotorImagery without an explicit events list silently got however many classes the dataset had, not the number requested, since the unfiltered result fed directly into EventsToLabels with no truncation downstream.
  • Found through sibling-function comparison between the two paradigm classes.
  • Severity: 3/5. Silent wrong behavior on a parameter the user sets directly, scoped to the events=None path.

SpikeInterface/spikeinterface ⭐ 872 — spike sorting pipelines

#4943 merged: compute_amplitude_cutoffs()'s own num_histogram_bins default stayed at 500, while its own docstring, its sibling AmplitudeCutoff.metric_params, and the low-level helper it wraps all default to 100.

  • Impact: With amplitudes_bins_min_ratio=5, the stale default silently raised the spike-count floor a unit needs before amplitude_cutoff stops returning NaN, from 500 spikes to 2500 spikes, for anyone calling the function directly without an explicit override. A prior fix (#4415) corrected the sibling defaults after a regression but missed this one.
  • Severity: 3/5. Silent NaN results on a routine quality-metrics call, likely to be mistaken for a data problem rather than a stale default.

udgamlabs/chitra-tiny ⭐ 0 — Chitra Tiny: a 212M-parameter text-to-image model (256 px) trained from scratch

#1 merged: app.py decided between a local checkpoint and the hub repo by checking only for model.safetensors, while from_pretrained requires both that file and config.json once it takes the local-path branch.

  • Impact: A local folder with weights but no config, from a partial download or a hand-copied checkpoint, crashed the app with an unhandled FileNotFoundError instead of falling back to the hub download.
  • Severity: 2/5. A crash on an incomplete local checkout, with no graceful fallback where one already existed for the normal case.

udgamlabs/laghu ⭐ 1 — Laghu: a 152M chat model trained from scratch for $146

Found during an audit pass over the training, eval, and data-prep scripts, bundled into one PR: #1 fixed four separate issues.

  • Impact: The most severe is unsandboxed code execution. gsm_eval.py ran model-generated code with a raw exec() and real builtins, inside a forked copy of the live training process, on every evaluation step by default, despite the repo already having a sandboxed execute_code helper built for exactly this. The other three: a contamination check that only marked one of several duplicate positions in a sorted probe array, an audit statistic that ignored the resumed-shard row offset, and an SFT document scanner with no deduplication, unlike its sibling scanner.
  • Severity: 4/5. Running untrusted model output with no sandboxing during ordinary training is a real execution-boundary gap.

udgamlabs/lilavati ⭐ 0 — Lilavati: a 95M model that solves maths word problems by writing Python

2 merged PRs. #1 fixed eval/gsm8k.py never closing or flushing its results file, so an interrupted run, a Ctrl-C, an OOM-kill, a dropped SSH session, left the file truncated or empty even after most problems had been scored. #2 fixed the sandboxed-execution structure check rejecting any solution using try/except ZeroDivisionError, the most natural defensive pattern for a math problem, even though the sandboxed runtime already exposes that exception for this purpose.

  • Impact: #1 silently lost real eval progress on any interrupted long run. #2 scored correct, safe solutions as failures purely for using a guard clause the runtime already supports.
  • Severity: 3/5 for #1, silent data loss on interruption. 2/5 for #2, correct programs scored as wrong, not a security gap.

udgamlabs/phonegpt ⭐ 0 — train a small language model on an Android phone's own CPU

#1 merged: drop_rng, the dropout-mask RNG stream, stayed at a hardcoded static seed while its two siblings, the weight-init and batch-sampling RNGs, were both correctly reseeded from -seed.

  • Impact: With dropout enabled, as the README's own training command uses, two runs with different seeds still drew identical dropout masks every step, an undisclosed shared source of correlation between runs meant to be independent, such as seed-variance ablations.
  • Severity: 2/5. A reproducibility gap in experiment design, not a crash or data-correctness bug.

udgamlabs/drishya-tiny ⭐ 2 — Drishya Tiny by Udgam Labs: a 212M video generator trained from scratch (text or photo to 2-second video)

#1 merged, two fixes: make_clip_ids.py wrote its output CSV directly to the real --out path while iterating shards, and only checked the --expect row count after all rows were already written. vbench_generate.py's run_jobs recovered a job's batch position with chunk.index(j), which falls through to tuple equality.

  • Impact: A failed --expect check in make_clip_ids.py left a wrong or partial CSV sitting at the real output path, so a downstream trainer would load it with no indication anything was wrong. In vbench_generate.py, two i2v jobs sharing the same prompt_en but pointing at different source images made chunk.index(j) fall through to comparing image tensors directly, and PyObject_IsTrue on a multi-element tensor crashed the entire run with "Boolean value of Tensor with more than one element is ambiguous."
  • Severity: 3/5. The CSV fix prevents silent partial-data corruption on a failed precondition. The indexing fix turns a real, reachable crash (not a contrived input, since VBench-I2V's own metadata file reuses prompts across images) into correct behavior through an atomic write and an enumerate-based index.

Co-authored credit (not in contributor graph)

Imbad0202/academic-research-skills ⭐ 57.6k 🔥 +34 today — Claude Code academic-paper skill suite

Submitted a GitHub Copilot compatibility concept (#457). The maintainer liked the pointer-file approach and rebuilt it with corrections as #465 (merged), crediting the idea through Co-authored-by.

  • Note: GitHub's contributor graph is built from the git author field only and omits co-author trailers. This credit is real and verifiable in the merged commit, though it will not show as a contributor avatar.
  • A small regex-escaping fix is open now (#956). Two earlier utility-mode PRs were closed as architectural mismatches with the repo's thin-trigger pattern.

Under review

Repo Stars PR What it fixes
udgamlabs/laghu 1 #2 RESULTS.md: note the ctx32k contamination-check bug predates this run
udgamlabs/lilavati 0 #3 README: note the try/except sandbox fix affects published GSM8K numbers
sunlabuiuc/PyHealth 1.7k #1515 Fix leading-gap visits poisoning later forward-filled padding
opencv/opencv 91.6k #30884 Fix segfault in masked float32 minMaxIdx/norm (IPP buffer over-read)
NVIDIA/auto-ontology 13.8k #537 Cap _short_table_block's columns like _shown_columns does
sebp/scikit-survival 1.3k #668 Fix IndexError in predict_proba with return_conf_int=True past last time
timeseriesAI/tsai 6.1k #990 Fix TSNormalize(by_sample=True) corrupting variables excluded from a listy by_var
CamDavidsonPilon/lifelines 2.6k #1721 Fix swapped lower/upper confidence interval columns in BreslowFlemingHarringtonFitter
fooof-tools/fooof 454 #413 Fix get_group(None, ...) producing wrong-shaped/empty event_time_results
neuropsychology/NeuroKit 2.4k #1257 Fix elgendi (2010) R-peak detector dropping a QRS block still open at the end of the signal
neurodsp-tools/neurodsp 361 #371 Fix off-by-one sample shift in burst-duration filtering
pactools/pactools 109 #42 Reject bandpass centers at or above Nyquist, fix auto grid endpoint
google-deepmind/alphagenome 2.2k #62 Fix Interval.resize(None) raising TypeError instead of a no-op
MRtrix3/mrtrix3 359 #3462 mrmath: fix median operation not ignoring infinite values
scverse/scanpy 2.6k #4508 fix: Ingest with an explicit PCA use_rep/use falls back to raw X
neuralinkcorp/datarepo 208 #87 Accept ISO date/datetime strings in ParquetTable filters
NVIDIA/Model-Optimizer 5.6k #3291 fix(quantization): pad MXFP4QTensor blocks like its sibling QTensor types
nilearn/nilearn 1.5k #6726 [FIX] multi-masker high_variance_confounds merge with list/tuple per-image confounds
Project-MONAI/MONAI 8.8k #9268 fix(metrics): SurfaceDistanceMetric leaks inf for ignore_index channel
wheresryan22/anatomy 28.1k #150 print.mjs: also look in the Windows Playwright cache for Chrome
google-deepmind/neural_testbed 192 #9 Fix correlated PRNG keys in GaussianSampleKL.kl_estimate
google-deepmind/alphagenome_research 930 #30 Mask opposite-strand tracks in GeneIntervalScorer.finalize_interval
ANTsX/ANTsPy 893 #957 Preserve channels_first through resample_image, smooth_image, and indexing
Lightning-AI/pytorch-lightning 31.4k #22145 Fix step-interval scheduler frequency under gradient accumulation
Project-MONAI/MONAILabel 913 #1939 Set display color on the highdicom DICOM-SEG export path
udgamlabs/udgam-english-125m 0 #1 Re-check committed shard hashes on resume in repaired packer
udgamlabs/shruti 13 #2 onnx: don't drop the trailing partial VAD window
scverse/anndata 786 #2737 Fix Reindexer._apply_to_dask_array densifying a sparse zero-axis input
TorchIO-project/torchio 2.5k #1549 Fix EnsureShapeMultiple computing target shape from the wrong image
MRtrix3/mrtrix3 359 #3457 Fix get_iqr() reading the wrong index for the lower quartile
RikiyaT/lm_understand-logs 1.3k #10 Fix acc_lenient always null for non-sharded eval runs in merge.py
NVIDIA-BioNeMo/cuik-molmaker 38 #8 Fix self-loop zero-fill writing to the wrong location and size
NVIDIA-BioNeMo/RNAPro 102 #7 Fix get_prot_nuc_frame checking only the first frame atom
NVIDIA-BioNeMo/megalodon 26 #2 Fix double-indexing of batch in discrete_fm interpolate_edges
NVIDIA-BioNeMo/warpKAN 4 #1 Fix order-3/4 backward B-spline kernels excluding left grid boundary
NVIDIA-BioNeMo/bionemo-agent-toolkit 509 #61 Fix ColabFold MSA script dropping selenomethionine (MSE) residues
neuralinkcorp/datarepo 208 #86 Fix contains/includes/includes any/includes all filters on ClickHouseTable
udgamlabs/udgam-evals 0 #1 Make g_units level 3 use the sampled unit pair, not a hardcoded one
NVIDIA/NemoClaw 23.0k #13925 Enforce Hermes's 64k Ollama context-window floor on inference set, no…
NVIDIA/DeepLearningExamples 14.9k #1470 Fix BERT pretraining periodic loss logging to average across all ranks
NVIDIA/Megatron-LM 18.4k #8649 Fix AbsorbedMLA to respect explicit config.softmax_scale override
NVIDIA/TensorRT 13.5k #4900 Resolve negative axis in ScatterElements plugin before indexing Dims
Imbad0202/academic-research-skills 57.6k #1020 fix(socratic): scope three unscoped rules to goal-oriented mode (#991)
pydicom/pydicom 2.2k #2403 Clamp LUT index before narrowing the index dtype in apply_color_lut()
TorchIO-project/torchio 2.5k #1541 Fix Crop/CropOrPad silently leaving a fully-cropped axis uncropped
NVIDIA/Model-Optimizer 5.6k #2867 Fix ONNX AWQ-Lite get_weight_scale mixing rows from different blocks
MRtrix3/mrtrix3 359 #3454 algo/histogram: fix Data::bin() silently dropping the maximum value
NVIDIA/Deep-Learning-Accelerator-SW 241 #42 Cast scale to float in save_scales to keep it JSON-serializable
Proximal-Labs/frontier-swe 238 #88 dart-style-haskell: fix instruction.md reference to nonexistent test dir
Proximal-Labs/frontier-swe 238 #87 Mount persisted trial-state volume at sandbox startup
TorchIO-project/torchio 2.5k #1527 Fix include/exclude being silently ignored in 9 remaining transforms
NVIDIA/Deep-Learning-Accelerator-SW 241 #40 Fix bias-less Conv handling in fuse_convs_horizontally
NVIDIA/Deep-Learning-Accelerator-SW 241 #38 Replace removed np.product with np.prod
NVIDIA/Model-Optimizer 5.6k #2567 fix(onnx): keep the converted weight's type in step with its zero point
TorchIO-project/torchio 2.5k #1518 Compute normalization statistics per sample
NVIDIA/Model-Optimizer 5.6k #2553 fix(autocast): stream calibration batches instead of materializing all of them
scverse/anndata 786 #2517 test(backed): add regression for filename=None after read_h5ad

Organizations

Owners of the repos above where a PR has merged, tagged rather than named.

21 organizations, 30 repos.

Imbad0202 InsightSoftwareConsortium MRtrix3 MakazhanAlpamys NVIDIA NeuroTechX Project-MONAI SpikeInterface alphaXiv galaxyproject google jordan-gibbs keras-team neuralinkcorp nilearn nipy pydicom scverse udgamlabs unstablebuild wheresryan22

Contact


224.5k combined stars across the 30 repos above (sum of each repo's current count, not a dedup of my own contribution).

This page refreshes automatically every day around 13:00 UTC (1:00 PM UTC) via GitHub Actions. Last run: 2026-11-18 20:57 UTC.

Pinned Loading

  1. Imbad0202/academic-research-skills Imbad0202/academic-research-skills Public

    Academic Research Skills for Claude Code: research → write → review → revise → finalize

    Python 57.7k 4.4k

  2. InsightSoftwareConsortium/ITK InsightSoftwareConsortium/ITK Public

    Insight Toolkit (ITK) -- Official Repository. ITK builds on a proven, spatially-oriented architecture for processing, segmentation, and registration of scientific images in two, three, or more dime…

    C++ 1.7k 760

  3. neuralinkcorp/datarepo neuralinkcorp/datarepo Public

    Python 208 48

  4. alphaXiv/OpenResearch alphaXiv/OpenResearch Public

    Turn your coding agents into research agents

    Rust 8.2k 565

  5. MakazhanAlpamys/Soup MakazhanAlpamys/Soup Public

    Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.

    Python 10.7k 1.8k