Research Impact Ledger
Google Scholar only sees papers. Most of what I ship is software, benchmarks, datasets and writeups. This page is a self-hosted research-impact ledger: it tracks scholarly citations (OpenAlex + Crossref), repository usage (GitHub), artifact downloads (Hugging Face) and web mentions, and gives every serious project a persistent, citable identity.
Google Scholar has no public API, so it is the one source that is not queried automatically. My profile's "new citations" email alerts cover the scholarly side; anything worth recording is logged in the ledger by hand. Google Scholar profile →
Persistent identifier:
ORCID 0009-0005-4120-555X
— stamped into every CITATION.cff and Zenodo deposit below,
so the works attribute to me automatically.
Last updated November 21, 2026. Usage metrics refresh automatically; citations are detected by the daily ledger job.
† Web mentions are found automatically, with self-links excluded.
A daily Brave Search sweep looks for each artifact and drops results on my own
domains and accounts, so what remains is third-party coverage. Hand-logged
entries in _data/research_impact_mentions.yml are counted too.
A “—” means nothing has been found yet — not that nothing exists.
Citable artifacts
A real-phone Android agent benchmark with a 60-task public set and a 530-task dataset, for open-weight models and on-device SLMs.
All archived versions (2)
- initial archive 10.5281/zenodo.23196835
- v1.0.0 10.5281/zenodo.23197068
Zenodo mints a concept DOI (stable, above) plus one DOI per archive. Older unlabeled archives from the first mint sit alongside the labeled v1.0.0 deposit; cite the concept DOI or the preferred version.
Cite (BibTeX / CITATION.cff)
Machine-readable citation ships in the repo as CITATION.cff; the archived release is on Zenodo at doi.org/10.5281/zenodo.23196834.
@misc{ androidlife,
title = {AndroidLife: Real-Phone Android Agent Benchmark for Open-Weight Models and On-Device SLMs},
author = {Yuvraj Singh},
year = {2026},
url = {https://github.com/YuvrajSingh-mist/AndroidLife},
note = {Benchmark},
doi = {10.5281/zenodo.23196834}
}
Shards, replicates and reassembles ML checkpoints across a Raspberry Pi cluster over raw TCP, with SHA-256 integrity verification and mDNS discovery.
All archived versions (2)
- initial archive 10.5281/zenodo.23196819
- v1.0.0 10.5281/zenodo.23197058
Zenodo mints a concept DOI (stable, above) plus one DOI per archive. Older unlabeled archives from the first mint sit alongside the labeled v1.0.0 deposit; cite the concept DOI or the preferred version.
Cite (BibTeX / CITATION.cff)
Machine-readable citation ships in the repo as CITATION.cff; the archived release is on Zenodo at doi.org/10.5281/zenodo.23196818.
@misc{ smoltorrent,
title = {smoltorrent: A Distributed Storage System for ML Checkpoints using Raspberry Pis},
author = {Yuvraj Singh},
year = {2026},
url = {https://github.com/YuvrajSingh-mist/smoltorrent},
note = {Software},
doi = {10.5281/zenodo.23196818}
}
Distributed training and multi-node vLLM rollout library for Apple Silicon clusters, used for the GRPO summarization ablations.
All archived versions (2)
- initial archive 10.5281/zenodo.23196837
- v1.0.0 10.5281/zenodo.23197086
Zenodo mints a concept DOI (stable, above) plus one DOI per archive. Older unlabeled archives from the first mint sit alongside the labeled v1.0.0 deposit; cite the concept DOI or the preferred version.
Cite (BibTeX / CITATION.cff)
Machine-readable citation ships in the repo as CITATION.cff; the archived release is on Zenodo at doi.org/10.5281/zenodo.23196836.
@misc{ smolcluster,
title = {smolcluster: An Educational Distributed Training and Inference Library for Local Computing},
author = {Yuvraj Singh},
year = {2026},
url = {https://github.com/YuvrajSingh-mist/smolcluster},
note = {Software},
doi = {10.5281/zenodo.23196836}
}
A reproducible local LLM benchmark and leaderboard for edge boards: tok/s, tok/J, TTFT and power under locked power modes across Jetsons, Macs, Raspberry Pis, phones and tablets.
All archived versions (2)
- initial archive 10.5281/zenodo.23196839
- v1.0.0 10.5281/zenodo.23197098
Zenodo mints a concept DOI (stable, above) plus one DOI per archive. Older unlabeled archives from the first mint sit alongside the labeled v1.0.0 deposit; cite the concept DOI or the preferred version.
Cite (BibTeX / CITATION.cff)
Machine-readable citation ships in the repo as CITATION.cff; the archived release is on Zenodo at doi.org/10.5281/zenodo.23196838.
@misc{ smolperfleaderboard,
title = {smolperfleaderboard: On-Device LLM Leaderboard},
author = {Yuvraj Singh},
year = {2026},
url = {https://github.com/YuvrajSingh-mist/smolperfleaderboard},
note = {Benchmark},
doi = {10.5281/zenodo.23196838}
}
SmolHub
softwareTiny from-scratch language models deployed on Hugging Face Spaces: the Smol* and Story* playground family that powers SmolZoo.
Cite (BibTeX / CITATION.cff)
Machine-readable citation ships in the repo as CITATION.cff.
@misc{ smolhub,
title = {SmolHub},
author = {Yuvraj Singh},
year = {2025},
url = {https://github.com/YuvrajSingh-mist/SmolHub},
note = {Software}
}
Paper Replications and Experiments
softwareA large collection of from-scratch PyTorch replications of classic and state-of-the-art AI/ML papers, written to be read alongside the originals, plus small language models trained end to end (SmolHub) and my own post-training experiments.
Cite (BibTeX / CITATION.cff)
Machine-readable citation ships in the repo as CITATION.cff; the archived release is on Zenodo at doi.org/10.5281/zenodo.23198171.
@misc{ paper-replications,
title = {Paper Replications and Experiments},
author = {Yuvraj Singh},
year = {2026},
url = {https://github.com/YuvrajSingh-mist/Paper-Replications-and-Experiments},
note = {Software},
doi = {10.5281/zenodo.23198171}
}
Readable single-file PyTorch implementations of deep reinforcement learning algorithms (PPO, SAC, TD3, DDPG, DQN, A2C) with Gymnasium and Weights & Biases support.
Cite (BibTeX / CITATION.cff)
Machine-readable citation ships in the repo as CITATION.cff; the archived release is on Zenodo at doi.org/10.5281/zenodo.23198154.
@misc{ neatrl,
title = {NeatRL: Readable Single-File Reinforcement Learning Implementations in PyTorch},
author = {Yuvraj Singh},
year = {2026},
url = {https://github.com/YuvrajSingh-mist/NeatRL},
note = {Software},
doi = {10.5281/zenodo.23198154}
}
A benchmark for whether LLMs and VLMs can generate and edit ASCII diagrams in a terminal, using an LLM-as-a-judge evaluation framework.
Cite (BibTeX / CITATION.cff)
Machine-readable citation ships in the repo as CITATION.cff; the archived release is on Zenodo at doi.org/10.5281/zenodo.23198058.
@misc{ asciitermdraw-benchmark,
title = {ASCIITermDraw-Bench: Benchmarking ASCII Diagram Generation and Editing},
author = {Yuvraj Singh},
year = {2026},
url = {https://github.com/YuvrajSingh-mist/ASCIITermDraw-Benchmark},
note = {Benchmark},
doi = {10.5281/zenodo.23198058}
}
Tutorials & articles
These are writeups with no standalone code artifact, so they are tracked as web artifacts (search/mentions rather than DOI citations).
- Mac Minis Thunderbolt Cluster Setup Guide
- Clustering 4 Raspberry Pis 4B
- Clustering 3 Jetson Orin Nano Super
How this ledger works
Every artifact below flows through the same pipeline:
GitHub → CITATION.cff → Zenodo → DOI → OpenAlex /
Crossref → this ledger. Each repo is tagged (v1.0.0)
and archived on Zenodo, which mints a concept DOI (stable, always
resolves to the latest version) plus a version DOI per tagged release.
A scheduled job queries each detector, writes a cache, and alerts me when
something new appears.
Detectors
- OpenAlex — citing works for each DOI
- Crossref — citation counts for each DOI
- GitHub — stars, forks, contributors, releases, traffic
- Hugging Face — dataset / model downloads and likes
- Web mentions — Brave Search, self-links excluded
Output
- A cache at
_data/research_impact_cache.json - This page, rebuilt on every deploy
- Email / Telegram / Discord alerts on new citations
The ledger itself lives at _data/research_impact.yml and the
detectors at scripts/research_impact/, both in the site
repository. They are kept private, so there is nothing to link to; the
collected metrics are published on this page.