Work brought into harbor
A selection of engagements across the Nordics and beyond. Names withheld; outcomes real.
One warehouse for forty stores
Challenge:
Sales, stock and loyalty data lived in separate systems that never agreed, so head office planned on stale spreadsheets.
What we did:
We built a single streaming warehouse with automated quality tests and a modelled semantic layer feeding every report.
A grounded contract assistant
Challenge:
Lawyers spent hours searching 12,000 documents, and off-the-shelf AI hallucinated clauses that did not exist.
What we did:
We built a retrieval-grounded assistant that answers only from the document set and cites every source paragraph.
Catching drift before customers
Challenge:
A credit-risk model had silently decayed for months; nobody noticed until approvals looked wrong.
What we did:
We added drift detection, live performance dashboards, alerting and a controlled automated retraining loop.
Defect detection on the line
Challenge:
Manual visual inspection was inconsistent and could not keep pace with the production line at peak.
What we did:
We trained a vision model on real line images and deployed it to an edge device for millisecond inference.
Support triage that routes itself
Challenge:
A rising ticket volume buried the support team, and simple questions waited behind complex ones.
What we did:
We built a bounded agent that classifies, drafts replies and routes only the hard cases to humans, with full logging.
Reports that finally reconcile
Challenge:
Every department reported different numbers for the same KPI, so leadership meetings argued over data.
What we did:
We built a governed semantic layer with one agreed definition per metric, feeding self-service dashboards.
A data foundation before the AI
Challenge:
An insurer wanted AI but had no data catalogue, unclear ownership and looming GDPR questions.
What we did:
We delivered a maturity assessment, a governance framework, a metric dictionary and a prioritised roadmap.
AI wired into the CRM
Challenge:
Reps copied data between four systems by hand, losing hours a week and making transcription errors.
What we did:
We connected the systems with event-driven flows and added in-place AI drafting and enrichment.
Forecasting demand honestly
Challenge:
Demand forecasts were guesswork, leading to costly over- and under-supply across the grid region.
What we did:
We built a forecasting model with explicit uncertainty ranges the planners could actually reason about.
EU AI Act readiness
Challenge:
A public body used automated decision support but could not classify it or evidence oversight for auditors.
What we did:
We classified each system by risk, produced model cards, tested for bias and designed human oversight points.
Semantic search for an archive
Challenge:
A media archive of decades of content was only searchable by exact keywords, hiding most of its value.
What we did:
We built embeddings-based semantic search so editors find by meaning, with relevance they can trust.
From nightly batch to real-time
Challenge:
Network event data arrived a day late, so problems were understood only after customers were affected.
What we did:
We moved the pipeline to streaming with change data capture, surfacing events within seconds.
Bring your data into harbor.
A 30-minute discovery call, no obligation. We will tell you honestly whether we can help.
Book a discovery call →