Abhimanyu Prasad
ResearchOpen sourceProjectsAbout
Portrait of Abhimanyu Prasad

Abhimanyu Prasad

I work on language tools for Bhojpuri and on making AI systems and their files checkable.

India·Hello·नमस्ते·प्रणाम·पाँव लागीं

GitHubORCIDOpenReviewHugging Facecrates.io

abhimanyu@notebook: ~/now

$ now

Now building

Building Bhojpuri sentiment data from where LLMs disagree

PURVA paper · Under review at TMLR

last commit · 8 days ago · "feat: add claim-check script"

@abhiprd2000★ 4

1,400

contributions, past year

@abhiprd2000

2,100+

downloads, all models and datasets

nlp-sentiment-model

# synced 15 Nov 2026 · daily

“

भोजपुरी बोले वाला करोड़ों बा, मशीन अबहीं ना बूझे।Tens of millions speak Bhojpuri. Machines don't understand it yet.

जवन भाषा में हम सोचेनी, ओकरा के मशीन तक पहुँचावत बानी।I'm bringing the language I think in to machines.

in Bhojpuri

How it got here

  1. 2021
    First laptop
  2. 2023
    Sainik School Tilaiya (AISSEE state rank 2)
  3. 2023-24
    Cyber Olympiad international ranks
  4. Late 2023
    ChatGPT failed in Bhojpuri, so I started learning NLP
  5. 2024
    Learned Rust
  6. 2025
    Started Axiom offline

2026 →

  1. First public repos
  2. Design guidance from Prof. Subodh Kumar
  3. GitPython CVE
  4. Stanza issue
  5. NeurIPS workshop papers accepted
  6. Stanza v1.15.0
  7. IEEE AIoT paper accepted
  8. PURVA paper to TMLR

Ledger

  • Sole author, workshop paper, Grounded User Simulation workshop, NeurIPS 2026●AcceptedNeurIPS page↗
  • Sole author, workshop paper, ATTRIB 2026, NeurIPS 2026●AcceptedNeurIPS page↗
  • Credited in Stanza v1.15.0 release notes (1 Oct 2026)Release v1.15.0↗
  • Finder credit, GitPython, CVE-2026-67325 (High)●Finder creditAdvisory↗
  • 13 merged PRs in 5 outside open-source projects●MergedGitHub search↗

Featured work

Bhojpuri in Stanford's Stanza library

Stanza had no Bhojpuri models. I opened the issue and wrote two of the data scripts that v1.15.0 uses.

Issue #1654 openedPR #1655 mergedTreebank PR #2 mergedPR #1675 mergedStanza v1.15.0 released

जवन भाषा में हम सोचेनी, ओकरा के मशीन तक पहुँचावत बानी।I'm bringing the language I think in to machines.

PURVA

Bhojpuri sentiment resources built from AI-model disagreement.

11 Bhojpuri websitesCollect, clean, deduplicateCheck language ID by handFive models label every sentenceCombine votes: majority vote and Dawid-SkeneHuman check on a held-out test setTrain classifiers, release everything
From the PURVA repository README.

Axiom OS and model-integrity-bench

A bare-metal Rust kernel that checks a file's integrity every time it is read, not just once at boot.

1 µs100 µs10 ms2.8 µs146 µsDTLN0.36 MiB2.6 µs6.1 msMobileNetV213.3 MiB2.8 µs45 msResNet-5097.8 MiB
Yellow: one checked 4 KiB block. Dark: hashing the whole file. Log scale, laptop (Intel i5-5200U).

How the pieces fit

Language

  1. Bhojpuri/Maithili sentiment models (Mar 2026)
  2. →PURVA
  3. →Stanza Bhojpuri
  4. →Tokenizer cost study
  5. →PURVA paper (TMLR)

Systems you can trust

  1. Axiom OS
  2. →model-integrity-bench
  3. →IEEE AIoT paper+Patent application
  4. →Zombie IoT article

Also on this line: GitPython CVE, Kaggle path-traversal fix

Testing AI agents

  1. UserSim paper
  2. →ATTRIB paper

Life outside code

  • Running most mornings.
  • Call of Duty: Mobile, which I play a lot and play well.
  • I'm in my school's painting section.
  • NCC at Sainik School: 'A' certificate, and the CATC camp at Koderma.
  • Cross-country at Sainik School, and casual volleyball there too.
  • I'm the family's go-to tech person, fixing phones and laptops for relatives.
  • I grew up between Magahi, Khortha, Bhojpuri, Hindi and English.
  • I taught informal CS sessions to classmates.
Check my work