The bet.
Delivery-drone logistics is already solved — Amazon, Wing and Zipline have flown for years. The unsolved problem is human: getting people to actually want one overhead, and getting a town to allow it. A delivery is paid for twice — in money, by the merchant, and in permission, by the city and the people below. In Europe, permission is the scarce currency, and it isn't bought with payload. It's bought with what people live under: the sound, the size, the way it moves.
Accepted, not endured.
So MaxFly is built to be let through, not put up with. Eyes on a screen that look around and pay attention as it lands. A voice designed on purpose — a low rumble, not a shriek. Flight paths that curve, anticipate and slow down near people. It delivers by winch: the parcel comes down on a line while the drone stays up, visible, out of reach. Small on purpose, too — noise climbs steeply with weight, so every gram is a decibel.
A drone you accept — not a robot you endure.
the shape — v1 targets
- 1 kg payload, in a reusable container
- ~55 cm and 3–3.5 kg all-up — small enough to be welcome
- 5 km per flight, 10 km across a hub network
- a winch drop, so it never has to land on you
- eyes, a designed voice, organic motion
Those are the numbers V1 is being designed to hit — targets, not yet achievements. Here's what already exists.
Under the hood.
There's more software here than a drone this size has any right to carry — and most of it already runs, in simulation.
- An acoustic design engine, not a spreadsheet. A real physics pipeline — blade-element momentum theory with low-Reynolds polars from NeuralFoil, Gutin tonal and NASA-BPM broadband noise, then psychoacoustics (loudness, sharpness) through MoSQITo — wrapped in an NSGA-II search that ranks rotor designs by perceived annoyance, not raw decibels, at ~0.15 s per candidate. It's how "every gram is a decibel" becomes a number instead of a slogan.
- The whole network flies in one command. An order comes in, the backend assigns a drone in ~0.2 s, an agent flies the organic mission in real ArduPilot SITL, telemetry streams to an ops console, and the delivery closes with proof. A simulated 16-hour day — 6 drones, 2 docks — ran 500 of 500 deliveries and measured the economics of battery-swap versus charging. Same MAVLink bridge and versioned data contracts as the real aircraft, so it transfers without a rewrite.
- Built like an org, not a hobby. 600-plus tests, pydantic→JSON-Schema contracts, ArduPilot SITL flown in CI on every commit, a Playwright + computer-vision QA harness, a self-hosted deploy pipeline — and a small constellation of named surfaces behind the drone: Kepler, Polaris, Vega, Ceres, Tycho.
Where it stands.
V0 is flying. It's a validation platform, not the product yet — an 8-inch open-prop frame (Mark 4 V2) on a Pixhawk 6C Mini running ArduCopter, with long-range radio and GPS. It's been flown hard: flight #86 reached, three crashes, and a stubborn ~16 Hz battery resonance the whole build now revolves around — diagnosed from telemetry alone, tap tests and a limit-cycle proof included, down to the finding that the wobble was burning roughly 40% of hover power. The hardware eyes, the speaker and the ducted body come next; until they do, the expressive layer lives honestly in simulation.
It's all built in the open, flight after flight, crash autopsies included, at maxfly.tech. My project of the moment — and the one I'd bet on.