Machine learning model comparison & evaluation
Client: Applied Machine Learning Cases
Challenge: A series of applied cases focused on understanding how different machine-learning approaches perform across varied prediction problems, with particular attention to model selection, validation, class imbalance, error trade-offs and when additional model complexity is justified.
Activities title: Key Activities
Activities: Data preparation & feature scaling, preparing structured datasets for modelling and testing the impact of preprocessing choices;
Classification modelling, applying logistic regression and other classification approaches and interpreting predictive performance;
Imbalanced-data handling, assessing class imbalance and its effect on model behaviour and evaluation metrics;
Tree-based & ensemble modelling, working with decision trees, random forests and XGBoost and comparing their performance across cases;
Neural-network modelling, developing and evaluating multi-layer neural networks and exploring transfer-learning approaches;
Model validation & comparison, using train/test validation, confusion matrices and comparative performance metrics to assess alternative models;
Threshold & error-cost analysis, adjusting classification thresholds and examining trade-offs between different error types in operational contexts;
Embeddings & text-generation workflows, exploring representation and generation approaches for unstructured text;
Technical-to-business translation, connecting model performance to practical implications such as cost, ROI, payback and build-versus-vendor decisions.
Expertise: Data & Analytics · AI & Automation
Automated lead generation
Client: Digital Transformation Practice
Challenge: Prospect sourcing and qualification depended on time-intensive manual research, making it difficult to scale outreach without either increasing effort proportionally or lowering targeting quality. The solution needed to automate the upstream workflow while keeping human control over who ultimately entered outreach.
Activities title: Key Activities
Activities: End-to-end workflow design, mapping the sourcing, enrichment, qualification and review process into a repeatable automated pipeline;
Automation architecture & integration, building an n8n/Python workflow connecting contact sourcing, external APIs and LLM-based qualification;
Targeting & scoring methodology, translating ideal-customer criteria into structured qualification logic for automated assessment;
LLM-based lead qualification, evaluating ~5K contacts against predefined fit criteria and achieving an 85–95% pass rate among automatically selected leads;
Human-in-the-loop governance, introducing a mandatory review step before any qualified contact could enter outreach;
Reusable acquisition infrastructure, designing the workflow as a repeatable GTM system rather than a one-off prospecting exercise;
Efficiency & quality optimisation, reducing manual research burden while preserving the qualification threshold required for high-quality outreach.
Expertise: AI & Automation · Process & Workflow Design · Technology & Systems · Data & Analytics
Challenges: Operational Bottlenecks
Cross-tool knowledge platform from research to MVP
Client: Product Teams
Challenge: Product knowledge was fragmented across tools such as Slack, Jira and Notion, making it difficult for teams to retrieve context quickly and creating uncertainty around whether the problem was significant enough to justify building a dedicated product.
Activities title: Key Activities
Activities: Market & problem discovery, testing the fragmented-knowledge problem across four candidate customer segments through 60+ interviews;
User research & JTBD mapping, translating interviews and observed product-manager workflows into personas, friction points and prioritised user needs;
Segment selection & positioning, identifying product teams as the initial target segment and defining the product proposition around a single source of truth across fragmented knowledge;
Product strategy & roadmap, converting validated needs into product vision, user journeys, feature priorities and backlog;
Design-partner acquisition & staged validation, securing a ~€23M UK/NL design partner and testing the proposition progressively through narrative, prototype and MVP stages;
MVP solution design & delivery, defining data flows, API connectors, permissions and sync logic and shipping a functional RAG/API-powered MVP integrating Slack, Jira and Notion;
Evidence-led iteration, using observed behaviour, user feedback and AI-output evaluation to reprioritise features when real usage contradicted earlier interview assumptions.
Expertise: Strategy · Solution Design · AI & Automation · Technology & Systems
Challenges: Unclear Technological Choices · Operational Bottlenecks
Supplier sustainability analytics & reporting system
Client: Large healthcare provider
Challenge: Supplier sustainability reporting relied on ~200K rows of duplicated and poorly structured procurement data in Excel, making the analysis difficult to use and impossible to repeat efficiently across future reporting cycles. The replacement also needed to remain lightweight and maintainable within the organisation’s existing technology and capabilities.
Activities title: Key activities
Activities: Current-state data diagnosis, identifying duplication, structural issues and limitations of the existing Excel-based reporting process;
Solution strategy & technology choice, matching solution sophistication to internal capabilities and selecting a lightweight Python, n8n and LLM/API architecture over a heavier analytics platform;
Data restructuring & supplier-level modelling, transforming fragmented procurement records into a consistent analytical dataset;
Supplier assessment & prioritisation methodology, defining auditable screening, classification and prioritisation logic; AI-assisted data enrichment, combining external sources, APIs, LLM-supported research and human checks to supplement incomplete supplier information;
Full-dataset analysis & visualisation, identifying reporting patterns and priority supplier groups for procurement action;
Repeatable reporting system, turning the methodology and analysis into a reusable workflow for subsequent sustainability reporting cycles
Expertise: Strategy · Solution Design · Technology & Systems Challenges: Unclear Technological Choices · Operational Bottlenecks
Human x AI sustainability review workflow
Client: Sustainability Consultancy
Challenge: Sustainability reviews depended on experts repeatedly interpreting the same regulatory requirements, creating a bottleneck as review volume increased. The challenge was to reduce repetitive work without automating ambiguous cases that still required expert judgement.
Activities title: Key Activities
Activities: Current-state workflow diagnosis, identifying repeated expert interpretation as the main bottleneck in the regulatory review process;
Decision-process decomposition, separating repeatable cases from ambiguous cases requiring expert judgement;
Human–AI workflow design, codifying repeatable decisions into explicit rules while introducing human-review gates for uncertain interpretations;
AI solution design, defining an LLM-enabled copilot and interpretation library to support reviewers with reusable regulatory knowledge;
Governance & control design, defining where automation could act independently and where expert validation remained mandatory;
Pilot testing & performance measurement, validating the redesigned workflow in practice and demonstrating ~25% lower review time before the underlying business line was later deprioritised following legislative change.
Expertise: Process & Workflow Design · AI & Automation · Solution Design · Governance
Challenges: Operational Bottlenecks