Built to run, not to demo
Five practices that overlap by design. Most engagements start in one and pull in the others once the constraint becomes clear — security most often, because it is the one people leave until last.
Data engineering & analytics
A warehouse your analysts trust, fed by pipelines that fail loudly rather than quietly.
Deliverable: a documented warehouse, tested pipelines, and analysts who no longer reconcile by hand.
Discuss this workScope
- Ingestion from operational databases, SaaS APIs, event streams and the spreadsheets nobody will give up
- Dimensional and semantic modelling, so two teams asking the same question get the same number
- Warehouse and lakehouse builds on Snowflake, BigQuery, Databricks or plain object storage with an open table format
- Data quality tests, freshness monitoring and lineage — the parts that decide whether anyone trusts the output
- Dashboards and self-serve layers designed around the decisions they support
Applied AI
Models scoped against a measurable baseline, evaluated honestly, and deployed with monitoring.
We publish precision, recall and the failure cases. If the model does not beat the baseline, we say so and stop.
Discuss this workScope
- Forecasting and demand planning where seasonality, promotions and long tails all matter
- Classification, anomaly detection and risk scoring on operational data
- Document understanding — extraction, structuring and validation of contracts, invoices and technical archives
- Retrieval-augmented assistants over your own corpus, with citations so answers can be verified
- Evaluation harnesses, drift monitoring and retraining pipelines, because accuracy on launch day is not the number that matters
Cloud consulting
Foundations, migration and cost engineering — plus the operating model to run it afterwards.
Works across AWS, Azure and Google Cloud. We do not have a preferred vendor to place.
Discuss this workScope
- Landing zones with account structure, network topology, identity and guardrails set before the first workload
- Migration planning and execution, sequenced so each wave can be rolled back
- FinOps: cost attribution per service, rightsizing, commitment strategy and the guardrails that keep savings
- Performance and reliability engineering against SLOs that map to something a customer would notice
- Security baselines and compliance evidence generated by the platform rather than assembled by hand
Platform & governance
The connective tissue that keeps data and AI work operable after the consultants leave.
Our aim is that dependence on us becomes a choice you make, not a position we engineered.
Discuss this workScope
- MLOps and DataOps: reproducible builds, versioned datasets, promotion between environments
- Access governance, PII handling and retention aligned to GDPR, DPDP and sector rules
- Model and data catalogues so people can find what already exists before rebuilding it
- Cost and usage reporting for data and inference workloads
- Enablement sessions, runbooks and pairing with your engineers throughout delivery
AI security
An AI system fails in ways a web application does not, and a conventional penetration test will not find them. We treat the model as part of the attack surface, because attackers already do.
Engagements run either as a review of a system you have already built, or alongside delivery from the first design session — the second is considerably cheaper.
Discuss an AI security reviewScope
- Threat modelling for LLM and agent systems — trust boundaries, data flow, and where a model sits astride them
- Prompt injection assessment, direct and indirect, against your own prompts and retrieval corpus
- Retrieval access control review: permissions enforced at query time and inherited from the source system
- Model supply chain — provenance, pinning, and scanning of weights, adapters and serialised formats that execute on load
- Secrets and PII boundary review across prompts, logs, traces and stored outputs
- Agent authority design — least privilege per tool, and human approval on irreversible actions
- Adversarial evaluation suite, kept as a regression test rather than a one-off exercise
- Governance mapping to OWASP LLM Top 10, NIST AI RMF, MITRE ATLAS, ISO/IEC 42001 and EU AI Act obligations
Start with the feasibility work
Two weeks, fixed price. We measure the current baseline and tell you what a realistic improvement looks like — including the case where the answer is that you do not need a model at all.