The need
A small engineering company of three engineers was well known locally through word of mouth and almost unknown beyond that. Two things limited its growth: too few new leads, and engineers who handled every lead themselves. Anything that added to their day would defeat the purpose.
The goal was concrete: at least three commercial and industrial solar leads a month, with enough extra leads for the engineers to pick the best ones. Factories, warehouses, hotels and schools have map listings, often have websites, and have roofs visible from above, so a system can learn a lot about them before any contact.
What I owned
I designed and built the system: the product scope with the company, the scan pipeline, the calculation engine, the review workflow and the application around them.
The solution
Most lead generation waits for a client to answer an advert. This system works the other way around. It looks at a territory from above, finds the businesses in it, measures their roofs, checks for existing panels, reads what the companies say about themselves, and works out what solar could mean for each one. Engineers then check the result and send every message personally.
Three principles run through the design:
- AI reads and judges, rules calculate. AI never does the maths and never decides alone which building belongs to a company, so the same input always gives the same result.
- Show the evidence. Every fact about a company cites a saved source or it is dropped, and estimates are shown as ranges.
- Spending stays under control. Every scan has a budget cap and pauses itself when it reaches it.
What happens to each site
A site leaves the chain early when there is no reason to spend more on it.
- Rules Discovery Grid search of the area, plus industrial zones.
- Rules Locate and attribute Works out which buildings belong to the site from map outlines.
- Rules Size filter Archives roofs below the scan's minimum.
- Rules Satellite image Fetched at a fixed zoom with its capture date.
- AI Solar check Haiku looks for panels, Sonnet rechecks unclear images.
- AI Details and sector Places the business in a standard industry sector.
- AI Research Sonnet reads the website and builds a sourced profile.
- Rules Estimate Load, system size, cost, savings and a 25-year cash flow.
- Rules Score A score out of 100 places the lead in band A, B or C.
- Rules Review queue Bands A and B wait for an engineer.
Key modules
Company names are blurred in these screenshots.
Review queue
One lead at a time, with four checks. Checks the AI is confident about are ticked in advance, so attention goes to the uncertain ones. Every correction feeds an accuracy report.
Map
Every lead on a map, coloured by score, solar status, lead status or sector. A user can draw a shape and start a new scan of that area straight away.
Lead page
Image, roof, company, estimate, calls and history. Changing a value recalculates the three scenarios and the 25-year cash flow on the spot.
Scans
Set up in five steps, then follow live progress per step. Scans can be paused, resumed, retried and repeated later, processing only what has changed.
Documents
A teaser PDF prepared for each company in English or Arabic, with its own roof image and figures. The numbers cannot be edited there, and exporting freezes what was sent.
Pipeline
Cards for approved leads, moved through the sales stages until they are won or lost, with results feeding the reports.
How it was built
The calculations depend on many assumptions that change with the market, so an admin edits them in a shared draft, publishes a new version with a note, and chooses whether to recalculate open leads. Each lead records which version it used.
Outside services are always reached through a small connector, so they can be swapped or replayed from saved answers during testing without spending money. The financial model reflects local reality, where businesses rely on part-time grid power and diesel generators, so proposals speak in diesel litres and generator hours.
Results
- A full working application: scan, review, study, teaser, call and pipeline, with reports and role-based access.
- Three solar scenarios per lead, with low, expected and high ranges for size, price, savings, payback and CO₂.
- The whole system ran on real companies in a real territory.