Ship Log • Specimens • Calibration • Instrument • Ground Control
I build systems that meet cameras, electrodes, users, and imperfect inputs as they actually are — AI engineering undergrad @ Amrita Vishwa Vidyapeetham.
gh pr list --author @me --state=all
I contribute upstream where a small, precise change removes a real developer paper-cut — then add the regression test or documentation that keeps it fixed.
32 pull requests merged to date — 27 upstream across six big-org labs (tabled below) + 5 more in peer and community repos, out of 119 filed overall.
accepted upstream
| lab | what shipped | pr | state |
|---|---|---|---|
 Microsoft · PyRIT |
Garak exploitation scenario — Jinja template injection + SQLi echo, 1.3k-line security feature | #2576 | 🟣 merged |
 NVIDIA · Curator |
treat DocumentSplitter separator as a literal string |
#2376 | 🟣 merged |
| NVIDIA · Megatron-Bridge |
in-process HashStore for temporary_distributed_context rendezvous |
#5989 | 🟣 merged |
| NVIDIA · Curator |
fix split_large_files RecursionError on single rows larger than target size |
#2377 | 🟣 merged |
| NVIDIA · DataDesigner |
tolerate literal braces in prompt validation (valid Jinja text) | #923 | 🟣 merged |
| NVIDIA · Speech |
SubsamplingReductionModule pooling computes lengths for a single MaxPool1d pass |
#16226 | 🟣 merged |
| NVIDIA · Gym |
normalize native LiteLLM responses before NeMoGymResponse validation | #3413 | 🟣 merged |
| NVIDIA · Gym |
distinguish Bird-SQL result mismatches and timeouts from execution failures in the evaluator | #3468 | 🟣 merged |
| NVIDIA · Megatron-Bridge |
allow Energon native sequence packing with MTP in the data pipeline | #5990 | 🟣 merged |
| Microsoft · agent-framework |
handle Literal annotations in is_instance_of |
#8351 | 🟣 merged |
| Microsoft · agent-framework |
preserve instruction order in prepend_instructions_to_messages dedup |
#8353 | 🟣 merged |
| Microsoft · agent-framework |
accept int timeouts in the A2A agent | #8516 | 🟣 merged |
| Microsoft · agent-framework |
preserve dict subclasses in workflow checkpoints | #8519 | 🟣 merged |
| Microsoft · agent-framework |
case-insensitive User-Agent detection prevents duplicate headers | #8521 | 🟣 merged |
| Microsoft · agent-framework |
skip empty and whitespace-only instructions in prepend |
#8524 | 🟣 merged |
| Microsoft · agent-framework |
allow text events when response_format is a json schema dictionary |
#8604 | 🟣 merged |
| Microsoft · agent-framework |
match empty expected tool call arguments when actual call arguments are None |
#8609 | 🟣 merged |
| Microsoft · agent-framework |
narrow pytest.raises blocks to wrap only get_response in error tests |
#8611 | 🟣 merged |
| Microsoft · agent-framework |
ensure add_usage_details returns copies and skips booleans |
#8613 | 🟣 merged |
| Microsoft · agent-framework |
fix positional invocations in detect_media_type_from_base64 docstrings |
#8614 | 🟣 merged |
| Microsoft · PyRIT |
restore keyboard focus to launch button on preview dismissal | #2686 | 🟣 merged |
 Meta · astryx |
elevation prop for ToggleButton — survived Meta's full design-system review loop |
#6037 | 🟣 merged |
| Meta · astryx |
guard TextInput.onEnter against Japanese-IME conversion commits |
#6083 | 🟣 merged |
| Meta · astryx |
scope the 16px input font floor to iOS only | #6085 | 🟣 merged |
 Google DeepMind · mujoco |
corrected misleading terminology in the physics-engine reference docs | #3550 | 🟣 merged |
 LangChain · langchain |
removed stale Args / Raises docstrings from core public APIs |
#40211 | 🟣 merged |
 Mistral AI · mistral-common |
fixed an error in the experimental usage guide | #306 | 🟣 merged |
in review — 83 signals awaiting a verdict
| lab | what i shipped | pr | state |
|---|---|---|---|
| Meta · astryx |
fix(Markdown): reject all data: schemes in parser isSafeUrl (#6323) | #6355 | 🟢 open |
| Microsoft · agent-framework |
break get_latest timestamp ties with the checkpoint lineage chain | #8514 | 🟢 open |
 OpenAI · openai-agents-js |
fix(sandbox): trim sandbox metadata and workspace paths in linear time | #1916 | 🟢 open |
 Anthropic · claude-code-action |
use job url origin for branch link so GitHub Enterprise Server works | #1844 | 🟢 open |
 Weaviate · weaviate |
fix(groupBy): correct minDistance and maxDistance in multi-shard merge and hybrid search (#13108) | #13131 | 🟢 open |
 deepset · haystack |
resolve cross-batch write-write conflicts by LLM call order in tool scheduling | #12628 | 🟢 open |
| Microsoft · markitdown |
feat(ipynb): render code cell outputs (streams, errors, text results) | #2534 | 🟢 open |
| NVIDIA · Speech |
fix(asr): repair ConvSubsampling forward paths missed by the MaskedConvSequential refactor | #16225 | 🟢 open |
 vLLM · vllm |
Install numactl in ROCm images so --numa-bind is not a silent no-op | #55433 | 🟢 open |
rest of the board — 50 more across astryx, NeMo (Curator/Gym/Speech/Run/Bridge), markitdown, semantic-kernel, claude-code-action, autogen, ollama, weaviate, comfyui, mujoco, datasets, presidio, aws-cdk, lm-eval-harness
| lab | what i shipped | pr | state |
|---|---|---|---|
| facebook/astryx |
docs(input fields): align statusVariant documentation and test coverage | #6428 | 🟢 open |
| facebook/astryx |
fix(Tokenizer): suppress Create when the typed label matches a selected token | #6367 | 🟢 open |
| facebook/astryx |
fix(TextInput): dispatch changeAction and update optimistic value on clear | #6425 | 🟢 open |
 data-privacy-stack/presidio |
fix(analyzer): keep Stanza multi-word tokens as surface tokens to preserve text and entity offsets | #2253 | 🟢 open |
| facebook/astryx |
fix(Markdown): preserve list structure for mixed task and bullet lists (#6330) | #6424 | 🟢 open |
| NVIDIA-NeMo/Gym |
fix(cli): pass config dict via process environment instead of command line (#3383) | #3472 | 🟢 open |
| facebook/astryx |
fix(FileInput): dispatch changeAction and update optimistic value on clear and selection | #6427 | 🟢 open |
| facebook/astryx |
fix(ChatComposerInput): anchor programmatic carets into trailing text nodes for IME composition (#6411) | #6423 | 🟢 open |
| facebook/astryx |
fix(Calendar): compute ISO week numbers without local DST offset leaking in | #6365 | 🟢 open |
| facebook/astryx |
render plain Link anchors inline so an ancestor Text clamp can truncate them | #6038 | 🟢 open |
| facebook/astryx |
floor SegmentedControlItem to a 44px touch target on coarse pointers | #6084 | 🟢 open |
| microsoft/markitdown |
fix(rss): emit item and entry links as heading anchors | #2525 | 🟢 open |
| microsoft/markitdown |
keep whitespace inside fenced code blocks when normalizing results | #2529 | 🟢 open |
| microsoft/markitdown |
fix(xlsx): render float values faithfully instead of 6-digit scientific notation | #2533 | 🟢 open |
| microsoft/markitdown |
fix(docx): clamp Word Heading 7-9 to markdown h6 | #2531 | 🟢 open |
| microsoft/semantic-kernel |
keep quoted named-arg values as single tokens in CodeTokenizer | #14461 | 🟢 open |
| microsoft/markitdown |
fix(pptx): render missing chart data points as empty cells | #2527 | 🟢 open |
| microsoft/semantic-kernel |
apply input_variables defaults when rendering prompt templates | #14459 | 🟢 open |
| anthropics/claude-code-action |
render actual text of failing MCP tool results in step summary | #1842 | 🟢 open |
| anthropics/claude-code-action |
keep "=" in PR link query values and encode path spaces | #1840 | 🟢 open |
| NVIDIA-NeMo/Speech |
fix(speechlm2): preserve relocated EOS token during early interruption augmentation | #16268 | 🟢 open |
 Comfy-Org/ComfyUI |
Repair mask editor painted-masked uploads so the original image survives under the mask | #16141 | 🟢 open |
 aws/aws-cdk |
docs(core): drop stale sep reference from ArnComponents.arnFormat default docs |
#38775 | 🟢 open |
| NVIDIA-NeMo/DataDesigner |
fix(compiler): support columns_added and columns_removed in processor config (#394) | #943 | 🟢 open |
| NVIDIA-NeMo/Run |
fix(cli): resolve string and future annotations in cli parser (#374) | #613 | 🟢 open |
| microsoft/autogen |
fix(autogen-ext): fix __unknown_models set difference and outdated docstrings (#8243) | #8246 | 🟢 open |
| anthropics/claude-code-action |
fix(mcp): support symbolic links with mode 120000 and target path content | #1837 | 🟢 open |
| NVIDIA-NeMo/Megatron-Bridge |
fix(ckpt): pass requested revision to snapshot_download in SafeTensorsStateSource | #6109 | 🟢 open |
| NVIDIA-NeMo/Megatron-Bridge |
fix(tokenizer): respect offline mode and local_files_only in _resolve_hf_tokenizer_revision | #6111 | 🟢 open |
| microsoft/semantic-kernel |
use field description over pydantic constraint objects in schema builder | #14444 | 🟢 open |
| microsoft/semantic-kernel |
fix kernel_function decorator mutating shared Annotated metadata dicts | #14442 | 🟢 open |
| microsoft/markitdown |
fix(xlsx): keep integer rendering when a blank cell promotes a column to float | #2485 | 🟢 open |
| microsoft/markitdown |
fix(html): keep head metadata out of markdown without dropping stray content | #2483 | 🟢 open |
| NVIDIA-NeMo/Run |
fix(cli): remove stray "Configuring global options" debug print | #611 | 🟢 open |
| NVIDIA-NeMo/Run |
fix(cli): forward entrypoint skip_confirmation to the generated command | #609 | 🟢 open |
| NVIDIA-NeMo/Run |
fix(cli): handle unparameterized list/dict annotations in value parser | #607 | 🟢 open |
| NVIDIA-NeMo/Run |
fix(cli): only treat delimited run/executor/plugins prefixes as overwrites | #605 | 🟢 open |
 EleutherAI/lm-evaluation-harness |
Fix typos in task documentation (16 files) | #4102 | 🟢 open |
| weaviate/weaviate |
add rq-4 to ALLOWED_COMPRESSION_TYPES valid entries | #12950 | 🟢 open |
| weaviate/weaviate |
Fix object_count metric Help text (copy-pasted from async_operations_running) | #12951 | 🟢 open |
 huggingface/datasets |
Fix typo in guide template | #8567 | 🟢 open |
| microsoft/autogen |
fix typos across documentation (16 files) | #8196 | 🟢 open |
 ollama/ollama |
log: report the source of the loaded context length | #18249 | 🟢 open |
| ollama/ollama |
normalize escaped pattern literals in tool/format schemas passed to llama-server | #18248 | 🟢 open |
| microsoft/autogen |
fix(ext): reject stale hunks in TextCanvas.apply_patch with context validation | #8195 | 🟢 open |
| google-deepmind/mujoco |
Fix inconsistent axis labels in solimp documentation figures | #3548 | 🟢 open |
| anthropics/claude-code-action |
skip malformed buffered inline comment lines instead of failing the post step | #1797 | 🟢 open |
| anthropics/claude-code-action |
only extract the user request from word-boundary trigger occurrences | #1799 | 🟢 open |
ls ~/lab/specimens
| case | friction | approach | proof | stack |
|---|---|---|---|---|
| SPX-01 · MOTION markerless motion capture |
useful mocap lives in laboratories — special suits, special cameras, special rooms | two ordinary camera views → MediaPipe landmarks → stereo DLT triangulation → filtering → joint kinematics | offline-first pipeline, live 3D inspection, real mocap export formats | Python FastAPI React Three.js OpenCV |
| SPX-02 · DECISIONS self-play reinforcement learning |
"the agent learned something" isn't evidence until tested against a known-optimal opponent | Double/Dueling DQN + NoisyNet + prioritized replay, pressured by self-play and a minimax baseline | ships both the learned policy and the sanity check — not just the score | Python PyTorch RL |
| SPX-03 · BIOSIGNALS EEG artifact recovery |
real EEG is eye movement, muscle noise, electrode drift — and somewhere in there, the signal | SVD separates low-rank artifact structure from 14-channel Emotiv EPOC X recordings | denoising as signal recovery, with the assumptions visible | MATLAB SVD signal processing |
| SPX-04 · ON-DEVICE VISION projX OMR grading |
document evaluation usually becomes "upload everything and wait" | capture, bubble detection, key comparison and export all stay on the phone | a complete grading workflow with zero server round trips | Flutter Dart OpenCV |
| SPX-05 · SYSTEMS POSIX file manager |
abstractions hide what the machine actually does | built directly on raw POSIX system calls — every operation mapped to its syscall boundary | know what's underneath your abstractions | C POSIX system calls |
cat calibration.md
Build it from scratch at least once — the magic disappears, the understanding stays.
If you can't reproduce it, you don't understand it yet.
Real data over clean examples — messy input teaches what tutorials hide.
Know what's underneath your abstractions. They all leak eventually.
./bench --readings
ping manohar
Email · 
LinkedIn · 
Repositories
a self-updating instrument · telemetry regenerated nightly from public GitHub data · visual telemetry designed for high-assurance biosignals and spatial systems