End skill instability. Build careers that last.
Tech skills have become a moving target. Developers in the Global South—and everywhere—invest years learning frameworks that become obsolete in 18 months. Bootcamps chase buzzwords. HR departments post "must-have" skills that fade before the job posting expires. The result: burnout, wasted potential, and broken careers.
AdaRSS is the intelligence layer that solves this. We're building an open-source AI system that reads the global job market and separates:
- Enduring foundations — principles that survive decades (data structures, systems thinking)
- Emergent necessities — skills gaining traction now (MLOps, transformer fine-tuning)
- Transient noise — framework syntax that changes yearly (React hooks version specifics, NextJS annoying updates)
From that classification, we prescribe stable, long-term learning pathways that won't go obsolete. We're enabling the Forever Talent ecosystem — careers built on skills that matter.
This is funded through Altruva Lab and deployed through LAEM Institute to serve emerging markets first.
5_ADARSS_Adaptive_Representation_Systems_For_Structured_Skilled_Data/
├── README.md ← You are here
├── notes/
│ ├── ideaguide_1.md # Initial vision
│ └── ideaguide_2.md # Full implementation guide
└── adarss/
├── README.md # Prototype-specific setup
├── requirements.txt # Python dependencies
├── .gitignore # Git exclusions
├── train.py # CLI training script
├── prototype.ipynb # Jupyter notebook (interactive)
└── data/
└── sample_annotated.csv # 50 hand-labeled skill samples (0=enduring, 1=emergent, 2=transient)
-
Clone/navigate to the project:
cd adarss/ -
Create a virtual environment:
python3 -m venv venv source venv/bin/activate # On Windows: venv\Scripts\activate
-
Install dependencies:
pip install -r requirements.txt
-
Run the notebook (interactive, recommended first time):
jupyter notebook prototype.ipynb
Then run all cells in order. The model will fine-tune for ~3 epochs and save to
./adarss-model/.Or run the CLI script (headless):
python train.py
- Loads 50 labeled tech skill / job title pairs
- Tokenizes with DistilBERT (lightweight, CPU-friendly)
- Fine-tunes a sequence classifier for 3 epochs
- Evaluates accuracy & confusion matrix
- Saves model & tokenizer
- Provides inference function to classify new skills
Output:
./adarss-model/— Fine-tuned DistilBERT + tokenizer./results/— Training logs and checkpoints./logs/— TensorBoard logs
Input: "Data Engineer: Apache Spark"
Output: Emergent (label: 1)
Reasoning: Currently dominant in big data, but competing frameworks emerging.
| Label | Category | Example | Rationale |
|---|---|---|---|
| 0 | Enduring | SQL, REST APIs, Systems Thinking | Survived 20+ years; core to all paradigms |
| 1 | Emergent | Docker, Kubernetes, Transformer LLMs | Gaining fast adoption; 3-7 year horizon |
| 2 | Transient | React hooks, Vue.js, Framework syntax | Tied to specific tools; lifespan ~12-18 months |
A developer who learns enduring skills builds a career that compounds. A developer chasing transient skills resets every 2 years. AdaRSS surfaces the difference—making it data-driven, not opinion-based.
adarss/data/sample_annotated.csv contains 50 curated samples covering:
- Job titles: Frontend Dev, Backend Dev, Data Engineer, ML Engineer, DevOps, Security, Mobile, QA, Systems Architect, etc.
- Skills: Languages (Python, SQL), frameworks (React, Vue, Next.js), tools (Docker, Kubernetes, Terraform), paradigms (REST APIs, microservices, reinforcement learning)
- Labels: Balanced mix of enduring (0), emergent (1), and transient (2)
- Justifications: Human-readable reasoning for each label
Next phase: Scale to 50,000+ real job postings from job boards + LinkedIn.
- Model: DistilBERT (lightweight, 67M params, open-source)
- Framework: Hugging Face Transformers + PyTorch
- Data: Pandas, NumPy
- Evaluation: scikit-learn (accuracy, confusion matrix)
- Notebook: Jupyter
- Environment: Python 3.10+
All open-source, all free to use.
- Hand-annotated 50-sample dataset
- DistilBERT fine-tuning pipeline
- Proof-of-concept inference
- GitHub repo & documentation
- 50,000+ real job postings
- Multi-label classification (skills can be multiple types)
- Web API for inference
- Dashboard for skill trend tracking
- Integration with LAEM Institute learning platform
- Personalized upskilling recommendations
- Global job market analytics
- Open data releases for research
This is an open-source project. We welcome:
- Data annotation — Label more job postings
- Model improvements — Try larger models, data augmentation, ensemble approaches
- Integration — Connect AdaRSS to learning platforms, career tracking tools
- Research — Publish findings on skill lifecycle & labor market dynamics
See CONTRIBUTING.md (coming soon) for guidelines.
- ATLAS AI — An open-source model that teaches AI's causal evolution interactively. GitHub
- Lineage — Structured, opinionated curriculum of AI's intellectual lineage. GitHub
- LAEM Institute — Workforce development ecosystem deploying Forever Talent across Africa. LinkedIn
- Concourx - A Social Network for professionals. Website
Abdulhakeem Muhammed
Altruva Lab
- GitHub: abdulhakeem-muhammed
- LinkedIn: abdulhakeem-muhammed-ibiyemi
MIT License — Open to all. No restrictions. Let's end skill instability together.
- My brother, for asking the question that sparked this
- Altruva Lab team for believing in the vision
- LAEM Institute for the deployment partnership
- Open-source community for Transformers, PyTorch, Hugging Face
Last updated: June 2026
Status: Early prototype, actively maintained
Next phase: Investment round for scale