Product case study · Luarc · 9 October 2026

A solar design in seconds, for Goan roofs

Aaditya Lobo. What the problem really is, the decisions I made, and a working tool built in a day to test them.

Who is presenting

I look for the problem behind the problem, then build the fix

• Head of Automation, YAAS MediaTools I built run hiring, reporting and performance scoring there every day
• Twelve projects of my own, liveAmong them Kai, an AI phone receptionist taking real calls, and Fin, a video production studio
• Before thatStrategy and business development at Avyott, tying what customers need to what gets built
• How I workFrom the first question to launch, building with AI coding agents and checking the result myself

The brief

Hours of manual roof design, down to seconds

Detect the roof. Place the panels. Produce the 3D model, materials, wiring, drawings, yield and a sales proposal. Match Arka360, Aurora and Reslink, on Indian and Goan roofs.

I treated it as two questions. What would I decide, and why? And is the idea real enough to run? So there is a document, and there is a tool you can use today.

The problem, taken apart

The physics was never slow. The tracing is.

• Get the roof into the computerTraced by hand, heights typed inMost of the time
• Place the panelsSemi-automatic, then nudgedMinutes
• Wire themStrings drawn by handMinutes to tens of minutes
• Predict the energyAutomaticSeconds
• Drawings, materials, proposalExported, then assembledThe rest

A year of generation for twenty panels simulates in under 20 milliseconds. So the product is a way to stop tracing, and to never ask a person for what a rule can do.

The constraint that shapes everything

There is no height data for Goa

• AuroraAutomatic roofs need laser height data. Coverage: the US and Europe
• Arka360The India workflow is manual tracing with typed heights
• Google Solar APIIn India since April 2026: about 1 m and 5° of error, and it cannot see a water tank
• Google building mapFree heights for every building, good to about 1.5 m and 4 m of detail. Canopy uses it
• Free satellite picture0.58 m a pixel over Goa. A tank is two pixels

So matching Aurora cannot mean copying it. Reading an Indian roof automatically is an open problem, and the one place to be better, not just cheaper.

Against Arka360, Aurora and Reslink

Ahead on the first step. Behind on everything proven.

• Getting the roof inTraced by hand, or needs laser data India does not haveRead from the picture, with a confidence on each value
• Strings and single-line diagramBy hand, or on the top paid planAutomatic, from the inverter's limits
• Money for a Goan homeGeneric or US financeGoa's slabs, surcharge and both subsidies
• Price₹46,000 to ₹1,00,000 a year (Arka360), $159 to $259 a user a month (Aurora), $2,000 to $10,000 a year (Reslink)Not for sale; free data, no cost per design
• Track recordIn daily commercial useOne day old, never checked against a measured roof

Also behind: no accounts, team or support; drawings unreviewed; five roof forms; a small catalogue; imagery not licensed for a product.

The idea

A satellite photo is a sundial

Decision

Spend accuracy where it pays

How far the yearly output moves when the roof is read wrong. Measured with the tool's own simulation.

So pitch takes the usual value for the material, and the effort goes into direction, obstructions and shade.

First principles

Five ideas from nature held up. Three did not.

• Sand poured on a plateSettles into the hip roof of that plateKept the roof grammar
• A tree's lowest leafKept only while it earns its keepKept how many panels
• Veins and blood vesselsSized to balance loss against materialKept cable sizing
• How eyes read a sceneGuess, render, compareKept the roof detector
• A sunflower's seed headGolden-angle spacing packs evenlyKept sampling the sky for shade
• Fibonacci trees of panelsSaid to catch more lightDropped measured voltage, not power
• Slime-mould networksFind short routesDropped no gain at this size

The tool

How I know the numbers are right

Checked against a public model, and against Tersus's own figures

The money model was given Tersus's published assumptions and returned its published results. The tool then replaces those assumptions with each roof's own yield.

What is not right yet

A day buys a working spine, not a verified product

• Detection is unmeasuredNo roof has been compared with a tape measure. The error ranges are my estimates
• Height is good to about a floorShadow, Google's map and the floor-height rule. Pitch is the usual value for the material. Both are marked and one tap to fix
• Grey slabsColour cannot separate a slab from a road. Google's map gives a coarse outline that still needs a drag
• Drawings are first-passNo engineer has reviewed them. Structure and wind are not checked
• Prices are placeholdersQuantities follow rules. The Prices tab takes your own rates or price list

What I would do next

Find out how wrong it is, on roofs Tersus already built

• Two weeksRun it on twenty finished homes. Compare panels, kWp and yearly units with what was installed and what the app has metered
• Six weeksThe phone capture (location, tilt-sensor pitch, sighted height) tried on real sites, then tree height. Google's height model as a second opinion. An engineer's review of the drawings
• LaterEach measured roof sharpens the defaults for its village. Structure checks. Societies and factories

To close

Thank you

The tool, the case study and every number in this deck come from the same code, and the code says where each number came from. I would like to run it on one of your finished roofs.