Enterprise AI Bootcamp

Enterprise AI Engineering Bootcamp

Five working demonstrations, built on real public datasets and permissively licensed components. Every step listed under each demonstration is a live screen, not a slide — the numbers are recomputed when you click, and every number can be drilled to the request that produced it.

Before the room fills: every badge below should read up. If one reads down, see the runbook — each demonstration has a documented fallback, and Demo 2 runs with the network disconnected as the designated substitute for any session blocked by infrastructure.
Demo 1 Day 3 checking…

Industrial Service Knowledge Copilot

Retrieval-augmented assistant over technical manuals

Open →
8 demonstrated steps
  1. Naive retrieval answers confidently and incorrectly — right procedure, wrong variant
  2. Baseline measured on an expert question set: context recall, faithfulness, citation accuracy
  3. Trace inspection proves the bottleneck is retrieval, not the model
  4. Live rebuild: layout-aware parsing, parent-child chunking, product-family and revision metadata
  5. Hybrid search plus cross-encoder reranking
  6. Identical re-run, before-and-after delta on one chart
  7. One question, three roles, three answers — one correctly refused
  8. Observability: retrieval quality, groundedness, latency, token cost, no-answer rate
Demo 2 Days 1 and 4 checking…

Predictive Maintenance and Anomaly Triage

Early failure detection from real air-compressor telemetry

Open →
10 demonstrated steps
  1. A model reporting 98% accuracy, presented before any caveat
  2. The confusion matrix: near-zero recall on the failure class
  3. Metric redefined with the room — recall at a fixed false-alarm budget, and lead time
  4. Measurement leakage exposed, and the score correction once features are causal
  5. Operating-regime segmentation and rolling features against the baseline
  6. A formal hypothesis test: null and alternative, p-value and effect size
  7. Lead-time distribution as the business-facing chart
  8. False-alarm rate converted into alerts per engineer per day
  9. Drift detection on a later window, and the retraining trigger
  10. Full experiment lineage across every configuration shown
Demo 3 Day 4 checking…

Computer Vision for Industrial Quality Inspection

Defect, wear and corrosion detection — and why the answer is rarely a better model

Open →
9 demonstrated steps
  1. A baseline detector with a mediocre score, and the room's instinctive answers recorded
  2. Error slicing by defect class, object size, contrast and lighting
  3. Labelling disagreement between two independent annotators on borderline images
  4. A corrected defect taxonomy, and the effect of re-labelling a subset
  5. Threshold tuning against the real cost of a false rejection versus a false acceptance
  6. An anomaly model trained on good samples only, catching a defect type absent from training
  7. An active-learning queue ranked by uncertainty, against labelling at random
  8. Edge export with measured latency and the accuracy cost of quantisation
  9. The licensing finding: a copyleft network-use obligation, and the permissive alternatives
Demo 4 Days 3 and 5 checking…

Agentic AI for Field Service Decision Support

Multi-agent diagnosis with tool integration, approval gates, Skills and guardrails

Open →
9 demonstrated steps
  1. A live diagnostic task resolved across three integrated tools
  2. The tool contracts themselves: typed schemas, descriptions, declared boundaries
  3. A vague tool description causing a wrong call — then corrected and re-run
  4. A denied approval, with the system escalating rather than improvising
  5. A prompt-injection attack planted in a service ticket, and the guardrail that defeats it
  6. Compounding-error arithmetic from a real multi-run reliability experiment
  7. Two agents replaced by deterministic code, and the reliability improvement
  8. A fully specified reusable Skill, with invocation and success rates
  9. The same flow as explicit code with declared state and failure handling
Demo 5 Day 5 checking…

AI-Powered Engineering Research Assistant

From published research to an executed experiment with a measured result

Open →
6 demonstrated steps
  1. Why research papers defeat naive ingestion — columns, equations, figures, tables
  2. A synthesised, citation-backed answer across the corpus
  3. The constraint-conditioned step: techniques correctly rejected as inapplicable
  4. A structured hypothesis card: claim, expected effect, data, statistical test, effort
  5. That card executed live through Demo 1's evaluation harness
  6. The outcome recorded in a research backlog and decision log — including a failure

Reading the suite

One shared evaluation harness

Demo 5 executes its hypothesis card through the harness Demo 1 is measured by — the same code, not two that agree. Closing that loop is the point of Day 5.

One shared trace store

Retrieval, reranking, tool calls and generation all emit spans using OpenTelemetry GenAI attribute names, so what you see here maps directly onto a real collector.

Licence discipline throughout

Every component is MIT, BSD or Apache-2.0. Three widely recommended packages were rejected on licence grounds, and that audit is itself a slide in Demo 3.