Skip to content
View usman7384's full-sized avatar
🎯
Focusing
🎯
Focusing

Block or report usman7384

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
usman7384/README.md

Muhammad Usman Altaf

Machine learning engineer trained as a software engineer. I build models that survive rigorous evaluation, and the backend systems that serve them.

MSc Data Science & AI, Université Côte d'Azur. Based in Sophia Antipolis, France.

Work

Tech Soft 3D — Geometric deep learning for B-Rep CAD shape retrieval on HOOPS AI. Built the evaluation pipeline the team validates every release against, and a reranker using invariant 3D shape descriptors that took Precision@1 from 72% to 94%, now in production.

INRIA / Université Côte d'Azur — Research on architecture and data efficiency in 124M-parameter language models. Controlled ablations across GQA, RoPE, RMSNorm, SwiGLU and Muon under fixed compute. Showed that backpropagating only the top 60% of tokens by excess loss beats isolated architectural tuning on few-shot MMLU and LAMBADA.

Mycophyto — RAG pipeline over thousands of mycorrhizal-fungi papers. Domain-aware chunking, Pinecone vector search, Mistral generation, orchestrated in LangChain, returning grounded and citable answers.

Carbonteq — Adversarial robustness for a clinical pneumonia classifier. Attacked with FGSM, BIM, PGD and C&W, then benchmarked 11 defence strategies across four families, raising robust accuracy from 35% to 74% with no loss in clean performance.

Oli Systems — Backend for a decentralised energy platform. RESTful APIs and integrations over the OCPP and OCPI standards covering charge-point communication, smart-charging sessions and roaming data exchange, plus a simulation modelling incentive schemes that shift EV charging toward grid-friendly periods.

Stack

Python, Java, C++, TypeScript, SQL. PyTorch, PyTorch Geometric, TensorFlow, MLflow. RAG pipelines, LLMs, LangChain, vector databases, fine-tuning. FastAPI, Spring Boot, Docker, CI/CD, AWS, GCP.

Elsewhere

Some of my work lives in private repositories or with the companies it was built for. If something here is relevant to what you are working on, get in touch and I am happy to talk through it in more detail.

LinkedIn · mu3931368@gmail.com

Open to machine learning engineer roles in France and across Europe.

Profile Views

Pinned Loading

  1. modernized-nanogpt modernized-nanogpt Public

    Forked from karpathy/nanoGPT

    A controlled, single-GPU benchmark of modern GPT architecture against the original GPT2.

    Python

  2. PointBot PointBot Public

    Interactive AI agent for 3D point cloud search, comparison, and embedding-based analysis.

    Jupyter Notebook

  3. step-to-uvnet-graph step-to-uvnet-graph Public

    Convert CAD STEP B-Rep models into UV-Net style face-adjacency graphs with sampled geometric features. A learning-focused exploration of CAD topology + GNN representations.

    Jupyter Notebook

  4. fahh-on-failure fahh-on-failure Public

    VS Code extension that plays the viral “fahh” sound when your code fails

    TypeScript

  5. forth forth Public

    Introducing a Forth interpreter, a powerful tool for executing Forth code efficiently. Harnessing the simplicity and performance of Forth, this interpreter provides a seamless experience for execut…

    TypeScript 2

  6. pythondnsclient pythondnsclient Public

    Dive into networking with our Python DNS client. Resolve domain names to IP addresses and vice versa seamlessly. Explore the codebase and integrate DNS functionality into your Python projects effor…

    Python