ASCENT: ML Engineering from Foundations to Mastery

1,333 lecture slides. 10 modules. 320 hours. Zero to masters.

ASCENT is the open-source ML engineering programme from Terrene Open Academy. It takes working professionals from their first line of Python to production-grade ML systems with full governance. Every concept is taught at three depths simultaneously. Every equation is derived, not asserted.

	Foundation Ascent (M1-M5)	Summit Ascent (M6-M10)
Level	Zero Python to production ML	Advanced to masters
Hours	160h (40 lessons)	160h (40 lessons)
Outcome	Deploy governed ML models	Build aligned AI agent systems

Curriculum

#	Module	What You Master	Slides
1	Python & Data Fluency	Python from scratch, Polars, data profiling, visualization	85
2	Statistical Foundations	20+ distributions, MLE, Bayesian inference, hypothesis testing, bootstrap, information theory	131
3	Feature Engineering & Experiments	CUPED variance reduction, DiD, causal forests, Double ML, 9 encoding methods, Boruta, leakage detection	99
4	Supervised ML	Complete model zoo (linear through CatBoost), XGBoost 2nd-order Taylor, bias-variance decomposition, conformal prediction	83
5	ML Engineering & Production	SHAP axioms + TreeSHAP, LIME, ALE, fairness (impossibility theorem), workflows, DataFlow, model registry, ensembles	150
6	Unsupervised ML & Pattern Discovery	K-means through HDBSCAN, EM/GMM (full derivation), PCA-SVD connection, t-SNE, UMAP, LDA, NMF, BERTopic, anomaly detection	146
7	Deep Learning	Linear regression as NN, backpropagation (full chain rule), parallelized training (data/model/pipeline/tensor), CNN, ResNet, Adam derivation	100
8	NLP & Transformers	BPE tokenization, Word2Vec (negative sampling derivation), LSTM gates, self-attention (why divide by sqrt d_k), transformer architecture, BERT, GPT, Flash Attention	150
9	LLMs, AI Agents & RAG	LLM landscape Q1 2026, 7 RAG architectures, hybrid retrieval, RAGAS evaluation, ReAct/Reflexion agents, multi-agent A2A, MCP protocol, Nexus deployment	235
10	Alignment, RL & Governance	LoRA/QLoRA, DPO (5-step derivation from RLHF), GRPO, PPO (clipped objective + GAE), Bellman equations, EU AI Act, PACT D/T/R governance, full platform capstone	154
	Total		1,333

Three teaching layers

Every concept is presented at three depths:

Layer	Marker	Audience	Example (Bias-Variance)
Intuition	Foundations	Zero-background professionals	”Imagine throwing darts at a target. Bias is how far the center of your throws is from the bullseye. Variance is how spread out they are.”
Mathematics	Theory	Intermediate practitioners	E[(y-y_hat)^2] = Bias^2(y_hat) + Var(y_hat) + sigma^2, derived step by step
Research	Advanced	Masters+ / PhD holders	Double descent (Belkin et al., 2019): test error decreases past the interpolation threshold in over-parameterized models

A banker and a PhD sit in the same classroom. Both leave having learned something they did not know.

What ships with the programme

Component	Count	Details
Lecture decks	10	Reveal.js HTML, three-layer depth, KaTeX math, speaker notes
Slides	1,333	Every equation derived, every algorithm stepped through
Exercises	80	Solutions + local Python + Jupyter + Colab (three-format consistency)
Datasets	11	Singapore-context: HDB resale 15M, taxi 50K, credit 100K, experiment 500K
Quizzes	10	246 AI-resilient questions (context-specific, not recall)
SDK textbook	163 tutorials	83 Python + 80 Rust, basic to advanced

Delivery formats

Format	Location	Best for
Local Python	modules/ascent*/local/*.py	Full async, Nexus deployment
Jupyter	modules/ascent*/notebooks/*.ipynb	Interactive exploration
Google Colab	modules/ascent*/colab/*.ipynb	Zero-install, GPU access

Not vendor lock-in

ASCENT teaches industry-standard tools. The Kailash Python SDK (the Foundation’s open-source ML orchestration platform) provides governance and orchestration on top:

What You Learn	Industry Standard	What Kailash Adds
Data	Polars (Apache Arrow)	DataExplorer: automated profiling, 8 alert types
Classical ML	scikit-learn, XGBoost, LightGBM, CatBoost	TrainingPipeline: orchestrated training + model registry
Deep learning	PyTorch	OnnxBridge: portable ONNX export
NLP	BERTopic, sentence-transformers	ModelVisualizer: interactive Plotly analysis
LLM agents	OpenAI / Anthropic / Groq APIs	Kaizen Delegate: structured output with cost budgets
Governance	EU AI Act / Singapore AI Verify	PACT: D/T/R accountability with operating envelopes

If you move to a different stack, you keep the math, the scikit-learn, the PyTorch, and the architectural patterns.

Quick start

git clone https://github.com/terrene-foundation/ascent.git
cd ascent
uv venv && uv sync
cp .env.example .env  # API keys for M9-M10

# Your first exercise
uv run python modules/ascent01/local/ex_1.py

# View lecture deck
open decks/ascent01/deck.html

License

Apache 2.0 (code and exercises). CC BY 4.0 (lecture content). Use it, extend it, teach with it.

Source: github.com/terrene-foundation/ascent