Skip to content
View Axwolf13's full-sized avatar

Highlights

  • Pro

Block or report Axwolf13

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
Axwolf13/README.md

Akshay Ashok

ML engineer and applied AI researcher, currently doing my M.Sc. in Data Science & AI at Saarland University and working part-time in data science at ORBIS Value+ in Saarbrücken. I work across the full pipeline (dataset construction, model benchmarking, evaluation design) in NLP and computer vision, with a bias toward measurements you can reproduce and systems that know when to distrust themselves.

Saarbrücken, Germany · axwolf13.github.io · akshay57ax@gmail.com · LinkedIn

Publication

Real-time yoga pose estimation with a hybrid MoveNet architecture, first-author, Sādhanā (Indian Academy of Sciences, Springer, 2025). A Thunder-train, Lightning-infer split that classifies 11 asanas at 98.8% test accuracy (98.4% verified on the shipped quantized model; reproduce it with verify_model.py) while holding 30+ FPS on consumer hardware.

Paper (DOI) · Code · Live browser demo

Selected work

  • molca-reproduction: A reproduction of MolCA (EMNLP 2023), a language model that reads molecules both as text and as graphs. Reproduced on two machines (62.32 and 62.77 BLEU-2 against the paper's 62.0), then the experiment the paper never ran: give it one molecule's text and another molecule's graph. It describes the graph's molecule in about 90% of the cases where it clearly picks one. Write-up
  • arabic-doc-triage-research: Benchmarking Arabic OCR (Surya, PaddleOCR) and Arabic-to-German translation on an 8 GB laptop GPU. KITAB-Bench evaluation, confidence-based failure routing and the safety signals a triage pipeline needs before its output reaches a human decision. Write-up
  • dodi-analysis: DODI, a deterministic index scoring how hard a platform's Terms of Service works to hide that "Buy now" means "revocable licence". Ten platforms across 2015–2024, checked against ToS;DR's human grades and three LLM judges from two vendors. v2 weighs each contract against what the store actually promised, using archived purchase pages. Also runs as an MCP server, so AI agents can call the scorer as a tool. Write-up · LLM-judge write-up
  • dodi-web: the DODI scorer shipped as a public web service: FastAPI app, Dockerized, test suite, CI on every push, deployed on Render. Score any ToS live
  • p2p-process-mining: Process mining on the public BPI Challenge 2019 purchase-to-pay log (251,734 order items, 1.6M events): throughput, matching-rule compliance, after-the-fact purchase orders, payment blocks, manual work, variants and a process map. Write-up
  • Shadow-Correction-DaS-Project: Quantitative study of piracy as a market signal for the Data and Society seminar.

Writing

Long-form write-ups of benchmarks and reproductions live at axwolf13.github.io/writing, published when there's something real to show.

Open to a Master's thesis topic for summer 2027, research collaborations and conversations about evaluating models and agents.

Pinned Loading

  1. molca-reproduction molca-reproduction Public

    Reproduction of MolCA (EMNLP 2023) on two machines, plus the test the paper never ran: when SMILES and graph disagree, the model follows the graph.

    Python

  2. dodi-analysis dodi-analysis Public

    DODI: a deterministic Terms of Service deception scorer with an MCP server. A 2015-2024 study of 10 platforms with ToS;DR and LLM-judge validation, plus v2, which weighs each contract against what …

    Python

  3. p2p-process-mining p2p-process-mining Public

    Process mining on the BPI Challenge 2019 purchase-to-pay log in plain pandas: throughput, matching-rule compliance, after-the-fact POs, payment blocks, automation, variants and a process map.

    Python

  4. arabic-doc-triage-research arabic-doc-triage-research Public

    Benchmarking Arabic OCR and translation on 8 GB VRAM: Surya vs PaddleOCR on KITAB-Bench, Opus-MT vs NLLB, confidence-based failure routing.

    Jupyter Notebook

  5. pose-estimation-movenet pose-estimation-movenet Public

    ode and trained model behind my Sādhanā 2025 paper: 11-asana classification from MoveNet keypoints, 98.4% verified. Live browser demo linked.

    Jupyter Notebook

  6. dodi-web dodi-web Public

    The DODI Terms of Service scorer as a web service: FastAPI, Docker, tests and CI, deployed on Render.

    Python