Drone manufacturing with AI — the prototype flew; unit 300 of the series is another story.

The problem

The drone does not fail on the line — it fails in the customer's first hour of flight.

Manufacturing drones/UAS has a peculiarity almost no other product shares: the volume curve climbs before the processes are mature. You go from runs of dozens to runs of hundreds in months, with the same team, the same building and the same notebook. And production mixes two natures that the plant manages with different logic:

1. The electronics — control board, motor controllers, power and signal wiring, cell pack. The defects are small-scale: a cold solder joint on a controller, a bridge between traces, a half-inserted connector, two different cell batches inside the same pack. A visual check at the end of the line does not catch them; the test bench catches them sometimes, and only when the failure is clear-cut.
2. The structure — composite arms and chassis, motor-propeller drivetrain, tightening torques. The defects are of another order: a delamination under the layup, a bubble, a torque below spec, a misalignment of millimeters that at 5,000 rpm turns into vibration.
3. The unit's file — which cell batch, which PCB serial, which motors, which bench result, which acceptance flight. It lives scattered across labels, spreadsheets, bench logs and the memory of the technician who built it.

When the RMA arrives — and in drones it arrives — the quality team faces a simple question with an expensive answer: how many more units have the same problem? Without genealogy by serial number, the only safe answer is “we review the whole month's production”. It is not that the plant is bad — it is that traceability stayed at prototype pace while volume multiplied by ten.

How it fits the IRIS system

iLEAN does not replace your test bench or your ERP — it seals the crack between the line, the bench and each drone's file.

The problem of a growing drone plant is not buying more equipment: it is that the electronics line, the structural assembly, the test bench, the acceptance flight and the ERP do not talk to each other in real time. iLEAN acts as putty between the systems you already have, without asking you to change the bench, the soldering station or the ERP.

Edge sees the soldering and assembly the quick check lets through. Connect captures the bench log and the acceptance flight, whether it arrives as a file, a photographed screen or a batch label. The agent chains cell → PCB → motor → drone and prepares the file by serial number — the responsible person signs.

The three iLEAN pieces applied to a drone/UAS line:

• iLEAN Edge — a terminal with computer vision (CNN) over the electronics station and the assembly table. It detects cold or bridged solder joints, shifted components, half-inserted connectors, wiring routed outside its guide and, on the structure, surface marks consistent with delamination or a bubble in the arm's layup. It flags the unit for the technician's review. It works without a network: if the building loses connectivity, Edge keeps inspecting and syncs when it recovers.
• iLEAN Connect — captures identifiers where they are already generated: the cell batch label, the PCB's laser mark, motor and controller serials, the arm's composite batch. And it captures the test results: the functional bench log file, acceptance flight telemetry, or the equipment's screen photographed when the bench is old. The sheet that today gets typed up the next day becomes usable data at second zero.
• iLEAN Agents — the agent chains cell batch ↔ PCB ↔ motor ↔ drone serial number and assembles the per-unit file that certification and after-sales ask for. It does not sign: it prepares the file and hands it to the person responsible for quality or airworthiness. The human signature is sacred; the three safety rings are there precisely to protect it.

See the full IRIS architecture →

Before and after

Drone line in prototype mode vs. line traced with iLEAN

Aspect	Visual check + spreadsheet	With iLEAN Edge + Connect + Agents
Solder defect in electronics	Caught only if the bench makes it fail outright	CNN flags it on the line, before assembly
Defect in a composite arm	Assembler's visual check, no record	Edge inspection with an image tied to the serial
Functional bench test	Written down by hand and typed up later	Log captured and tied to the serial number
Acceptance flight	Paper checklist, filed in a folder	Telemetry captured, validated by the technician
Genealogy cell → PCB → motor → drone	Manual reconstruction, if it exists at all	Chained at second zero, queryable
Scope of an RMA	“The whole month's production”	The serials actually affected by the batch
File for certification	Assembled when requested, in weeks	Prepared by the agent, signed by the responsible person

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your line. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Drone/UAS manufacturer scaling from short runs to mid-size series, with an electronics line, a structural assembly area, a functional test bench and an acceptance flight.
• Edge pilot over the electronics station plus test bench capture. First value expected within a few weeks: a soldering or insertion defect pattern surfaces that until now only showed up in the field.
• Expected reduction of rework in electronics and assembly ≥30% and of per-unit file preparation time ≥50%. Indicative payback between 4 and 9 months, depending on monthly volume, current RMA rate and the hours spent today rebuilding traceability.
• The hard lever is the RMAs avoided or scoped down: with genealogy by serial number, a review campaign goes from “the whole month” to “these 34 serials”.

And the quality manager's reasonable doubt

“What if the CNN invents a defect or, worse, lets a real one through on a control board?” — hallucination is a problem of free generation, not of anchored tasks. Flagging candidate defects on a board and comparing them against the golden-sample reference is an anchored task: the best models brought error below 1.5% ^[1]. And even so, the system does not decide — it proposes, flags, records. The signature of the electronics technician, the assembly lead or the acceptance pilot stays right where it belongs. The three safety rings are there precisely for this: the decision on a unit that is going to fly is not closed by an AI, it is closed by a person.

^[1] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.

Related: Avionics with AI · CFRP/GFRP composites with AI vision · Satellite components with AI