Framework Library
Reusable AI value-creation frameworks
The leverage model of the Group CTO role, made concrete. Each framework is solved once, reference-implemented in the AI & Automation Specialist's lab, then transferred to the PortCos where it applies. No PortCo should ever rebuild one of these from scratch.
AI-Assisted Design Generation
contrib · HeroGenerative + evaluative loop for design artefacts (energy model, layout, configuration).
Architecture
Generator (LLM / diffusion / parametric) → evaluator (physics engine / simulator) → Pareto ranking → user-selectable options.
Data requirements
Simulator access (e.g., Chenath), constraint specification, historical design dataset for priors.
Cost model
Simulator-bound — typically $0.01-1.00 per candidate; value is hours saved per design.
Failure modes
Simulator fidelity gap; over-fit to historical priors; over-automation removes assessor judgement.
2 deployments in portfolio
- Hero — Plan-to-model CV — auto-populate NatHERS from PDF/DWG floor plans
- Energy Inspection — Plan-to-model CV (shared with Hero) — BERSPro auto-populate from floor plans
Anomaly Detection in Time Series
contrib · Unified StreamingUnsupervised + semi-supervised anomaly surfacing with LLM triage of root-cause hypotheses.
Architecture
Metric stream → rolling stats + isolation forest → LLM hypothesis generator → ranked alert with evidence links.
Data requirements
High-cardinality metric history; labelled incidents for eval.
Cost model
Mostly streaming infra; LLM triage <$0.05/alert. ROI from reduced MTTR.
Failure modes
Alert fatigue; false negatives on novel failure modes; seasonality mis-modelled.
3 deployments in portfolio
- Scope Systems — Predictive-maintenance recommender on mine equipment telemetry
- Dbvisit — Replication-health anomaly detection + NL DBA assistant
- Unified Streaming — ML-driven per-title encoding-ladder optimisation (QoE + CDN cost)
Constrained Optimisation Agent
LLM-planner + OR-Tools solver for scheduling, allocation, pricing, loading problems.
Architecture
Problem capture (NL → structured) → OR-Tools / CP-SAT solver → LLM explanation + what-if generator.
Data requirements
Constraint taxonomy from domain SMEs; historical decisions for baseline comparison.
Cost model
Solver compute $10-100/run; LLM $0.05-0.20/explanation. Transformative if recurring decision.
Failure modes
Infeasible problems; mis-translated constraints; user distrust without explainability.
Conversational In-Product Copilot
contrib · ActivateDomain-grounded RAG copilot with tool-calling to existing product APIs.
Architecture
RAG index (per-tenant) + tool-registry → LLM (Claude Sonnet) with function-calling → audit trail.
Data requirements
Product docs, tenant data APIs, RBAC scope mapping.
Cost model
$0.10-0.50/session at scale; offset by support-ticket deflection.
Failure modes
Prompt injection via tenant data; tool-call errors; role-scope leakage; support cannibalisation if mis-designed.
3 deployments in portfolio
- Omnitronics — Voice-to-incident-log + cross-jurisdiction real-time translation
- Removify — Multi-platform removal orchestration agent — autonomous detect→submit→appeal
- Activate — Conversational IAM — NL access requests auto-routed via ServiceNow/Entra
Document Intelligence
LLM + OCR pipeline for classification, extraction, summarisation, auto-filing.
Architecture
Document ingest → OCR (AWS Textract / Azure AI Document Intelligence) → LLM classifier (Claude Haiku / GPT-4o-mini) → structured metadata → auto-routing rules.
Data requirements
≥10k labelled documents for eval harness; customer-tenant isolation; RAG index per tenant.
Cost model
$0.002-$0.01 / document at scale; amortised over retention value. Target <5% of per-tenant ARR.
Failure modes
Hallucinated metadata; cross-tenant leak; Copilot co-existence; OCR degradation on handwritten content.
1 deployment in portfolio
- MacroView — Ship MacroView Copilot-extension: records-mgmt + compliance filing on top of Copilot
Predictive Margin / Churn
Time-series ML forecasting project profitability and customer churn propensity.
Architecture
Event stream → feature store (DuckDB / ClickHouse) → gradient-boosted model (XGBoost) + LLM narrative layer for explainability.
Data requirements
12+ months of job/customer lifecycle events; outcome labels; feature stability over time.
Cost model
Training: <$500/month compute. Inference: negligible. LLM narrative: ~$0.01/prediction.
Failure modes
Label leakage; distribution shift at seasonal boundaries; customer over-reliance on forecast.
2 deployments in portfolio
- Streamtime — Job margin guardrail — predict negative-margin trajectory + draft client comms
- WCC Group — Skills-graph embedding for cross-border PES; career-path recommender