Skip to content
View mohamed-bal's full-sized avatar
3 followers · 2 following

Organizations

@devupai-platform

Block or report mohamed-bal

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
Mohamed-bal/README.md
Mohamed Bal — Research, Engineering, Products

Mohamed Abdessalem Bal

AI Researcher · Full-Stack Engineer · Founder of DEVUP AI

I study how AI models work, and build products people can use.

Read my research on arXiv Connect on LinkedIn Email Mohamed Bal

Live icon from devupdz.com
devupdz.com · Personal portfolio ↗

Based in Algeria · Arabic / French / English



Research   /   Products   /   Toolkit   /   Certifications


01 / Research

I'm an independent AI researcher working on mechanistic interpretability, sparse autoencoders and causal validation. Alongside research, I build AI infrastructure, developer tools and digital products.

FEATURED PAPER · arXiv PREPRINT · JULY 2026

From Geometric Recovery to Causal Validation

A Reproducible Audit of Sparse Autoencoder Features, from Superposition Geometry to Causal Inertness

Does a geometrically recovered feature actually affect model behavior? My work audits SAE features through activation checks, ablation and steering, separating decoder alignment from causal evidence.

Mohamed Abdessalem Bal · Mechanistic interpretability · Reproducible evaluation

Read paper ↗   ·   Audit toolkit   ·   Experiments & reproduction

02 / Products I build

From API architecture to the interface: I work across the full product, with local payments, multilingual experiences and useful developer workflows in mind.

DEVUP AI official logo

DEVUP AI ↗

FOUNDER & LEAD ARCHITECT

Global AI access. Local DZD billing.

A unified, OpenAI-compatible gateway to 200+ AI models for developers building AI-powered applications.

  • One integration across model providers.
  • CIB, Edahabia and BaridiMob payments.
  • Model discovery and developer documentation.

Platform   ·   Documentation

 DOXDZ official logo

DOXDZ ↗

CHIEF TECHNOLOGY OFFICER

Administrative knowledge, made practical.

A digital administrative platform with 10,000+ documents and templates, built for the Algerian context.

  • Find documents and administrative references.
  • Customize templates with AI assistance.
  • Edit and export documents to Word or PDF.

Explore platform
DEVELOPER TOOLING

DEVUP Code ↗

An AI coding assistant for VS Code, Cursor and Google Antigravity, connected to the DEVUP ecosystem.

Understand context. Work with selections, files and workspace context.
Review changes. Inspect suggestions in a diff before accepting them.
Move work forward. Explain code, request fixes and generate tests.

Product   ·   VS Code Marketplace   ·   Open VSX

03 / Expertise & engineering toolkit

From model training and causal research to coordinated agents, full-stack applications and production infrastructure.

Languages & foundations

Python JavaScript TypeScript C++ Rust Go Java PHP Bash SQL HTML CSS

Web & API engineering

React Next.js Node.js Express FastAPI Django GraphQL Tailwind CSS

AI & model training

PyTorch TensorFlow scikit-learn NumPy Pandas Hugging Face Jupyter OpenAI API Ollama ONNX

Agents & workflow orchestration

LangChain LangGraph n8n CrewAI AutoGen MCP Temporal Airflow

Data, servers & delivery

PostgreSQL Supabase MongoDB Redis MySQL Prisma Docker Kubernetes Linux NGINX Git GitHub Actions Google Cloud Vercel Kafka Cloudflare

Research → Train → Connect agents → Deploy
Connecting models, tools, memory and specialized agents into coordinated workflows.

Applied AI: LLM integrations · RAG · Embeddings · Semantic search · Streaming · Structured outputs

How I approach engineering
  • Design clear API contracts and documentation alongside the implementation.
  • Make changes easy to review, validate and recover.
  • Treat language, payments and local constraints as product requirements.
  • Distinguish a promising model result from evidence that survives intervention.

04 / Certifications

PreviewCredentialIssued by
Train A Small Language Model certificate Train A Small Language Model
b0a4e29d-094e-4753		Google DeepMind
Deploy Multi-Agent Architectures certificate Deploy Multi-Agent Architectures
9a2829d0-54f8-4ca4		Google Cloud
Microsoft Applied Skills GitHub Copilot certificate Applied Skills: GitHub Copilot
E7D56442EAB03851		Microsoft
Software Engineer certificate Software Engineer
B90EF4E15AAA		HackerRank
Frontend Developer React certificate Frontend Developer (React)
BOBO266AF426		HackerRank

Click a preview to open the full certificate.



From understanding models to shipping products.

Open to conversations about interpretability, AI infrastructure and developer tools.

Portfolio ↗   ·   Let's build something useful ↗



Popular repositories Loading

  1. sae-causal-audit sae-causal-audit Public

    Official implementation for the paper: "From Geometric Recovery to Causal Validation: A Reproducible Audit of Sparse Autoencoder Features"

    Python 4

  2. tt-linear-pytorch tt-linear-pytorch Public

    Production-ready PyTorch drop-in replacement for nn.Linear using Tensor-Train (TT-SVD) decomposition to slash LLM VRAM footprint.

    Python 3

  3. devupai-node devupai-node Public

    Official Node.js SDK for DEVUP AI — Algeria's AI inference gateway. OpenAI-compatible, 170+ models, DZD billing.

    TypeScript 2

  4. superposition-to-monosemanticity superposition-to-monosemanticity Public

    Reproducing "Toy Models of Superposition" & L1 vs. TopK SAEs with 77% causal gap findings.

    Python 2

  5. Mohamed-bal Mohamed-bal Public

    1

  6. cineops cineops Public

    Evidence-grounded AI incident response for media production pipelines — Google ADK + Gemini + Grafana MCP.

    Python