Four hours, one loop
Hour one builds the machine. Hour two plugs it into the network. Hour three teaches it to read real Ericsson counters — verbatim formulas, animated tickers. Hour four closes the loop on real degradations, including the course’s flagship: drop-rate optimization with AI.
M1The Learning Machine
What AI actually is — every family opened up with telecom numbers: a regression computed by hand, gradient descent rolling downhill, trees and forests, k-means, PCA, isolation forests, neural networks with a glowing backprop blame-wave, attention, LLMs, reinforcement learning — and the discipline: splits, leakage, drift.
Play Module 1 →M2AI Meets the Network
Where intelligence plugs in: the PM/CM/FM data estate and its quality tax, SON, NWDAF (TS 23.288), TR 37.817 and the Rel-18 air interface, O-RAN’s two RICs — rApps & xApps over E2/A1/O1 — the Ericsson stack (EIAP, ENM, MicroSleep Tx, Performance Diagnostics), shadow mode, MLOps, SHAP and guardrails.
Play Module 2 →M3The KPI Engine — Ericsson, Verbatim
Real counters, real formulas, animated formula-tickers: DRB Accessibility as a product of three doors (→ 97.0%), Retainability with its sacred active qualifier (75 / 14,340 → 0.52%), the five-child drop-cause tree, mobility, throughput, HARQ BLER — then features, and five detectors racing the same six-day degradation: day 6 · 5 · 3 · 2.
Play Module 3 →M4Closing the Loop
The five verbs on real cases: the Bandra cell’s SHAP ledger, PIM-vs-external forensics (r = 0.81), a verified recovery 1.81% → 0.33%; the corridor’s MRO surgery inside guardrails; sleeping cells, RL energy saving, load balancing, fleet interference forensics, capacity — plus the Python pipeline, rApp deployment, drift day-two, and your Monday-morning plan.
Play Module 4 →M5Deep Learning Unpacked
The neural network, bolt by bolt — every number by hand. Build a neuron from arithmetic, run the forward pass by hand, compute backpropagation on paper, then climb through optimizers, initialization, regularization, the two loss curves, and the CNN and LSTM. A black box no more.
Play Module 5 →M6Generative AI, LLMs & Agents
Modern AI, opened all the way up — grounded in the NOC. Tokenization, embeddings, the attention mechanism worked by hand (query·key·value), the transformer block, next-token pretraining, alignment, sampling, RAG, and AI agents that use tools — the complete arc from a neuron to an agent.
Play Module 6 →Everything feeds the loop
Models from M1, placed by M2, fed by M3, closing in M4 — the same five verbs your network will run every night.
Then stop watching — and drive it
The Optimization Lab is a working simulator: a live market of 12 cells writing real Ericsson counters every ROP. Break one — PIM, a mobility fault, a sleeping cell — advance the clock, and watch the exact counters climb, the KPI formulas compute the damage, four detectors race, a classifier name the cause with SHAP evidence, and your fix bring the drop rate home against a control cell. Everything is computed live from a hidden fault-physics model; nothing is faked.
LABThe Optimization Lab
Inject any of six faults into any cell and run the full loop by hand: the counter warehouse, a live KPI engine with formula reveal, the detector race, the drop-cause tree + SHAP ledger + PIM-vs-external test, and act & verify against a matched control cell. Every number is generated from the same hidden physics your action changes.
Open the lab →And see it at work — six real case studies
The Case Files are fully-worked, engineering-grade studies of AI/ML optimization in production — each with the exact data sources, the model and why that model, the method and guardrails, honest before/after numbers, the catch that nearly broke it, and the transferable lesson. Fleet-scale PIM detection, RL energy saving, sleeping-cell autoencoders, MRO bandits, capacity forecasting, interference fingerprinting — plus ten more in brief.
CSCase Files — real AI/ML optimization
Six detailed case studies with before/after charts, real Ericsson counters, model rationale, method, and honest results — filterable by domain. The applied companion to the course and the lab.
Open the case files →