Skip to content
View VARUN3WARE's full-sized avatar
🐧
Like a Penguin since 2023
🐧
Like a Penguin since 2023
17 followers · 10 following

Achievements

Achievement: Pull Sharkx2Achievement: YOLOAchievement: QuickdrawAchievement: Pair Extraordinaire

Achievements

Achievement: Pull Sharkx2Achievement: YOLOAchievement: QuickdrawAchievement: Pair Extraordinaire

Highlights

  • Pro

Block or report VARUN3WARE

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. Markdown is supported. 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
VARUN3WARE/README.md

Varun Rao: AI/ML engineer, agent reliability, robotics data infra

AI/ML engineer at IIT Bhilai (B.Tech, Data Science & AI) building agent reliability tooling, robotics data infrastructure, and ML systems that survive outside notebooks.

varunrao.is-a.dev · LinkedIn · Medium · X · Email

Merged upstream PRs HFlow ARGUS NVIDIA nvcf PyPI

Now

BUILDING     Human Slop: an anti-AI social platform built on typing forensics
CONTRIBUTING HFlow (Hebbian Robotics) · ARGUS (agent silent-failure detection) · NVIDIA nvcf
SHIPPED      pytorch-dml on PyPI · 19 merged upstream PRs · 70+ articles on Medium
OPEN TO      ML engineering and research roles where prototypes become products

Open source

19 merged upstream PRs: HFlow 12, ARGUS 4, kornia 2, NVIDIA nvcf 1; 2 open in shap

HFlow · Hebbian Robotics

SDK for robotics teams to verify the quality of the data they train models on · 12 merged PRs

My work there centers on delivery integrity: making sure a dataset snapshot or LeRobot import can prove it arrived intact, and refusing malformed or hostile input with a clear exit code instead of a traceback.

Area What shipped
Verification hflow verify lerobot-import and a shared VerificationReport contract for the whole hflow verify family (#454)
Snapshot integrity Per-table and per-asset sha256 receipts plus a content_id in format.json (#401, closes #397)
LeRobot import Imports that publish straight into S3/GCS/Azure data roots (#377) and resume at episode boundaries after a mid-batch failure (#390)
Catalog UI hflow catalog ui on bucket-backed catalogs (#366), with live rebinding when Parquet lands mid-session (#538)
Hardening Path-escape refusal, SQL statement gating, and malformed-receipt handling (#515, #564, #576, #672)
Full HFlow contribution log (12 merged)
  • #672 · merged · latest · closes #671. hflow verify snapshot now treats an integrity key that is present but not a JSON object as a malformed receipt (exit 2), instead of misreporting it as a legacy pre-#401 snapshot.
  • #639 · merged · closes #638. hflow curate and hflow stale report DuckDB binder and catalog errors cleanly with exit 2 instead of crashing, while the hosted server keeps its separate 400 mapping.
  • #576 · merged · closes #575. Refused null or wrong-type integrity.tables / integrity.assets containers with a clear error instead of an AttributeError traceback.
  • #564 · merged · closes #563. Applied the single-SELECT SQL gate to hflow curate --dry-run so it matches the manifest-write path and the hosted server, refusing DESCRIBE / SHOW / PIVOT shapes.
  • #554 · merged · closes #553. Removed stale "verifier not shipped" wording from the snapshot docs and exporter docstring.
  • #538 · merged · closes #537. The catalog UI rebinds empty in-memory tables when their Parquet appears mid-session, so tables like ingest_failures become queryable without a restart.
  • #515 · merged · closes #469. Snapshot verification refuses receipt paths that are absolute, contain .., or use drive letters before reading them, restoring the "joined only onto the handed directory" guarantee for tenant-supplied snapshots.
  • #454 · merged. Added hflow verify lerobot-import to check prepared-manifest deliveries against their receipts without re-converting, with exit codes 0 clean / 1 damaged / 2 unreadable / 3 unverifiable.
  • #401 · merged · closes #397. Recorded path / size_bytes / sha256 for every snapshot table and copied asset in format.json, plus a content_id so a deleted member is detectable later.
  • #390 · merged · closes #303. Multi-episode LeRobot imports resume at episode boundaries, reusing finished episodes only after checking their import identity (source commit, camera keys, converter version).
  • #377 · merged · closes #304. hflow import lerobot publishes directly into s3://, gs://, and az:// data roots, writing the manifest only after every episode succeeds.
  • #366 · merged · closes #305. hflow catalog ui opens existing S3/GCS/Azure catalogs through a local mirror, without ever writing to the bucket.

ARGUS · ArgusLabs

Catches silent failures in AI agents before users do · 4 merged PRs

  • #118 · merged · closes #90. Stopped writing ARGUS_RUN_ID on every finished run, which let the last of many batched runs overwrite the pointer so argus check graded the wrong run.
  • #77 · merged · closes #73. Added argus check --strict warn_as_fail so CI can fail on warning-level tool failures such as HTTP 429s without changing the default behavior.
  • #68 · merged · closes #56. Made ARGUS run its own pytest --argus plugin in CI on every PR, so a plugin regression can't hide behind a green pipeline.
  • #66 · merged · closes #53. pytest --argus now watches LangGraph ainvoke, stream, astream, batch, and abatch, not just invoke, so silent failures on those paths are caught.

NVIDIA nvcf · NVIDIA

Platform for deploying and routing GPU-accelerated inference, streaming, and batch workloads · 1 merged PR

  • #1836 · merged. Added a build-only Docker Buildx CI job that validates the OpenBao migrations image for both linux/amd64 and linux/arm64 without publishing anything.

Other contributions

  • kornia · 2 merged. SEO meta descriptions across 46 documentation modules (#3129) and a docstring for _detach_tensor_to_cpu (#3131).
  • shap · 2 open. Fixed waterfall plot labels being cut off in saved figures (#4249) and LinearExplainer ignoring the link parameter for classifiers (#4252).

Selected work

Project What I built
Human Slop Anti-AI social platform that scores how human a post is from keystroke dynamics (WPM, bursts, pauses), with hardware-bound auth. Deployed on web and mobile.
pytorch-dml · PyPI PyTorch library for Deep Mutual Learning and knowledge distillation, with AMP, DDP, and ONNX export. 2–5% accuracy gain over independent training.
Hedgera Multi-agent trading system combining real-time streaming, RL forecasting, and LLM debate. Led an 8-person team; about 20% returns with 5–8% max drawdown in backtests and paper trading.
Paged-Attention Implementation of PagedAttention from the vLLM paper, managing the KV cache in pages like virtual memory.
ArthJAX GPU-accelerated agent-based macroeconomic simulator in JAX with households, banks, contagion, and shocks.
RAPIDADB GPU-native vector database in C++/CUDA and PyTorch for sub-millisecond similarity search.
BPlusSQL Disk-backed B+ tree storage engine in C++17 with an LRU buffer pool and write-ahead logging.
Kerala Ayurveda RAG Medical RAG system using corrective RAG and hybrid BM25 + vector retrieval. Under 10% hallucination rate on RAGAS.
evalforge Engine that scores and stress-tests ML models for reliability, robustness, and hidden failures before deployment.

Experience

  • AI Developer Intern, Kartavya Technology (remote, Jun–Aug 2024). Built multi-agent automation and cloud APIs on AWS and GCP; cut manual effort by 40% and infrastructure costs by 25%.
  • Team Lead, Autonomous Financial Intelligence Platform (IIT Bhilai, Oct–Dec 2025). Led the 8-person team behind Hedgera.
  • Coordinator, Data Science & AI Club, IIT Bhilai. Ran workshops and hackathons.

Achievements

  • Kaggle Expert. Silver medal in the MITSUI Commodity Prediction Challenge (rank 36 of 1,711); top 10% in GQ Volatility Forecasting (rank 34 of 386).
  • Amazon ML Challenge 2025. All-India rank 278.
  • Winner, Pixel Perfect Hackathon (IIT Bhilai). Improved the baseline by 23%.
  • Technical writer. 70+ ML articles on Medium with 800+ monthly readers.

Toolbox

LANGUAGES   Python · C++ · Rust · SQL · TypeScript · JavaScript · Bash
ML / AI     PyTorch · JAX · TensorFlow · scikit-learn · Hugging Face · XGBoost · OpenCV
AGENTS      LangChain · LangGraph · RAG · multi-agent systems · evals
DATA        DuckDB · Parquet · PostgreSQL · MongoDB · Neo4j · FAISS · ChromaDB
INFRA       CUDA · Docker · Kubernetes · AWS · GCP · FastAPI · GitHub Actions · MLflow

Contribution graph: Vegeta's Galick Gun clashes with Goku's Kamehameha

train → evaluate → break it on purpose → ship

Pinned Loading

  1. dml-py dml-py Public

    PyTorch library for Deep Mutual Learning, knowledge distillation, and collaborative neural network training with advanced features and research-backed strategies.

    Python 1 3

  2. Hedgera Hedgera Public

    Modular AI trading system using FinRL, multi-agent analysis, and real-time data pipelines.

    Python 1

  3. Kerala-Ayurveda-RAG Kerala-Ayurveda-RAG Public

    Safe, accurate Ayurveda content generation using CRAG RAG + 6 AI agents. Prevents hallucinations, checks contraindications. OpenAI + Streamlit UI included.

    Python

  4. Reciprocal-Contextual-Poetry-Generation Reciprocal-Contextual-Poetry-Generation Public

    This project develops a novel poetry generation system that integrates context-based learning with reciprocal human feedback, reinforcement learning, and personalization to model a dyadic human–AI …

    Jupyter Notebook 1 1

  5. RAPIDADB RAPIDADB Public

    High-performance GPU-native vector database built with C++/CUDA and PyTorch for sub-millisecond search and massive throughput.

    Python 1

  6. BPlusSQL BPlusSQL Public

    A high-performance, disk-backed B+ tree storage engine in C++17 featuring an LRU buffer pool, WAL crash recovery, a SQL layer, and a TCP server.

    C++