Engineering production ML and data workflows.
Assessment and handoff engineering for the platform behind Greehill's urban tree intelligence. Delivered through WeSpeakAI.
greehill
- Client
- Greehill, urban tree intelligence for cities
- Delivered through
- WeSpeakAI, enterprise AI consultancy, Budapest
- Engagement
- Data engineering and architecture consulting, followed by implementation
- Technologies
- Python, Airflow, Kubernetes, PostgreSQL, Azure
Context
Greehill helps cities manage urban trees as infrastructure. LiDAR scans of streets are analysed with machine learning to produce continuously updated data for every tree: health, storm risk, clearance conflicts, climate impact. The company reports more than 70 data points per tree across 3 million trees and 130+ cities.
That output depends on a production, GPU-backed ML and data platform, with preprocessing produced in one repository and consumed by model workloads in another. Meniva's engagement, delivered through WeSpeakAI, had two phases: assess the platform with evidence, then engineer the handoff between those repositories.
Technical challenges
A working platform that had become hard to reason about.
Architecture knowledge spread thin
How the components fit together, where ownership sat and which dependencies carried risk was not written down in one place.
Reliability and capacity under GPU constraints
Failure modes, reproducibility gaps and capacity limits needed to be understood from production evidence, not assumptions.
An implicit handoff between repositories
Preprocessing output crossed a codebase boundary on implicit assumptions. A change on one side could break the other without anyone noticing early.
Meniva's contribution
From evidence to decisions, then to code.
Phase 1 · Assessment
An evidence-backed architecture, reliability and capacity assessment of the production platform. Production evidence was turned into a ranked set of engineering decisions, each with explicit acceptance criteria the team could verify.
Phase 2 · Implementation
Hands-on engineering on the cross-repository preprocessing handoff: reproducibility controls, and an explicit producer/consumer contract between the repositories across Python, Airflow, Kubernetes, PostgreSQL and Azure.
Verified deliverables
- Architecture, reliability and capacity assessment
- Prioritised engineering decisions with acceptance criteria
- Reproducibility controls on the preprocessing handoff
- Producer/consumer boundary between the repositories
When this pattern is useful
A production data or ML platform works, but architecture knowledge, reliability controls or cross-system ownership have become difficult to reason about.
Where the claim stops
Greehill is shown with permission; the engagement was delivered through WeSpeakAI. No confidential client code, data, commercial information, internal architecture detail or measured business outcome is reproduced or claimed. Company figures above are Greehill's own public statements. The visualization above is a conceptual, procedurally generated scene, not client data.
Next project
Contour →