Economics & Transfer
Factor 12 economic leverage through AI-assisted development — and what the LaSoFly pattern means for other domains.
1. The Economic Lever: Factor 12
The LaSoFly platform — around 30,000 lines of code, six sub-applications, mathematically verified consensus algorithm and TrustScore extension — was implemented by a single person in 37 active development days. A classical estimation along McConnell, ISBSG benchmarks and COCOMO II would have predicted around 534 person-days at roughly 375,000 €. The real effort was around 30,000 € — AI subscription and infrastructure included.
Classical estimation
534 person-days
~ 375,000 €
McConnell / ISBSG / COCOMO II benchmarks at a senior daily rate of 700 €.
AI-assisted, real
37 person-days
~ 30,000 €
Empirically measured at a senior daily rate of 800 €.
The ratio between the two figures is about 12 for cost and 14 for time — not a single data point: the TrustScore extension alone was implemented in 2 person-days against a pre-fixed internal plan of 18 to 22 days. Two independent measurements with consistent magnitude.
What that changes economically: At classical development costs of 375,000 € the platform would be viable only for industrial clients and large public grants. At a real 30,000 € subscriptions starting at 175 € per year and a free educational license for schools become feasible. For a German municipality of 200,000 inhabitants LaSoFly saves around 42,000 € in personnel cost per audit cycle — the platform pays for itself in a matter of weeks.
2. Three Principles for AI Projects
A more detailed treatment lives on the Quality Assurance page. The short form for everyone who wants to read on:
- A solution is only as good as the parts it is built from.
- The best AI model is worthless if deployed in the wrong context.
- Garbage in, garbage out — an AI predicts coherences, not truth.
3. Transfer to Other Domains
The pattern of the project — deterministic solution meets structural limit, foreign AI model fails at the domain gap, own curated dataset plus complementary information source resolves the problem — is not restricted to urban trees. Six examples from very different worlds show the same shape:
Solar potential cadastre on rooftops
From orthophotos alone one cannot tell whether a bright surface is a flat industrial roof, a pitched roof or a concrete firewall. Only the combination with the nDSM (roof slope from height model) and solar irradiance simulation gives a robust answer.
Surface sealing in residential areas for stormwater
Sealed surfaces are recognized well on RGB by classifiers — but asphalt, dark gravel and slate roofs look similar. Height information reliably separates driveable paths (~0 m) from buildings (>3 m).
Storm damage detection for insurers and forestry
Before/after aerial image comparisons suffer from the problem that any seasonal change (leaf fall, growth) can be misclassified as damage. An nDSM comparison shows structural changes — fallen trees, missing roofs — independent of vegetation state.
Radio field planning with real terrain attenuation
Classical radio planning tools work with terrain models and building layers but treat vegetation as a flat damping surcharge at best. An nDSM with classified vegetation (tree vs. shrub vs. bush) sharpens the propagation forecast in exactly the transition zones where LoRaWAN sensors fail today.
Quality inspection in industrial vision
A camera over the production line detects defects — but only those defect types contained in the training dataset. The LaSoFly mechanism of weighted consensus with multiple inspectors plus gold-tile system transfers one-to-one. The platform architecture would not need to be rethought.
Detection of archaeological structures in LiDAR
Faintly preserved ramparts, medieval terraces or vanished paths disappear in RGB under vegetation but emerge as fine height patterns in the high-resolution DGM. An AI trained on such patterns reveals finds that even trained eyes miss — provided the training data is curated by archaeologists, not scraped from image databases.
What unites these examples is not the domain, but the pattern: a problem that looks classically solvable meets a structural limit, a generic off-the-shelf AI model fails, and only the combination of an own curated dataset, weighted consensus and a complementary information source delivers reliable results.
4. Scaling and the radio bracket
The methodology is not hessian-specific. Digital terrain models and orthophotos are available nationwide at comparable resolution (BKG aggregates DGM1 from all 16 federal states at ±30 cm vertical accuracy); across the EU the INSPIRE directive harmonizes the data landscape. More than 26 EU countries provide documented open-data sources or established research licenses. A transfer to further German municipalities is a configuration question, not a development task.
Back to the starting point with the radio waves: the same trees that pose challenges for LoRaWAN sensors in the forest fire project are, in the urban climate, the tool with which heat islands can be dampened. Both projects show the same mechanism from two directions — a physical object that resists simple modeling. The solution is the same in both cases: the right data, at the right place, in the right architecture.
Read more about LaSoFly's consensus mechanics.
To Quality Assurance