What does Helix Edge do?

Helix Edge builds ruggedized industrial edge-AI hardware and the software platform to manage it. Our compute modules run computer vision and other neural network workloads directly on factory floors, retrofittable into existing PLC cabinets, with a managed cloud (Helix Studio) for model deployment and fleet management.

Why did you pivot from drones to industrial edge-AI?

Two reasons: drone-delivery regulatory unlocks in Europe kept getting delayed, and a single conversation with a BMW plant operator surfaced a much larger, more immediate market for the compute platform we'd built for the drones. The pivot took our existing onboard compute design from a drone payload to an industrial enclosure.

How long did the pivot from drones to edge-AI hardware take?

Roughly four months from the first customer conversation in May 2024 to the announcement of the new company in July, plus another twelve months to fully wind down drone work and ship a mature v1 product. The clean separation of drone IP from the underlying compute platform was what made the pivot possible at all.

What's the hardest part of running a hardware startup in Europe?

Supply chain volatility and the long iteration cadence. Each board revision costs €40-80K and 8-12 weeks of tooling and certification. You can't plan a hardware company the way you plan a software company — every decision is more expensive and slower to undo.

How did you handle losing investors during the pivot?

We lost one of two seed leads — a UK fund that had specifically bet on autonomous delivery. They voted against the pivot at an investor meeting (we won 4-1), and the relationship didn't fully recover. They didn't participate in the Series A. Some investors back theses; some back founders. Knowing which kind of investor you have matters more than people admit.

Dmitri Werner & Helix Edge: From Drones to Edge AI

IChapter 1 of 5

The Problem

## The Wrong Problem, Solved Well

By June 2024 my co-founder Hannes and I had spent twenty months and 1.3 million euros of seed money building autonomous last-mile delivery drones. We had three working prototypes and a proof-of-concept route across an industrial park in Augsburg where a drone could pick up a package, fly 1.2 kilometers, land on a target pad, and return without human intervention.

What didn't work was anyone wanting to buy it.

The drone-delivery thesis rested on a regulatory unlock that was supposed to happen at the EU level in 2023, then 2024, then "soon." Without that unlock, commercial drone delivery in Europe was a one-off pilot business. Fine for a research project. Not a product company. I knew this in February 2024. I didn't admit it to myself until April. I didn't admit it to Hannes until June.

## The Conversation in the BMW Cafeteria

The pivot started with a single conversation. In May 2024 I was at a BMW supplier event in Munich. I sat across from a woman named Petra who ran factory automation at a BMW transmission plant. I told her, without much filter, that I had spent two years building autonomous drones nobody wanted to buy.

She laughed and said: "I would pay you a lot of money for the brain of that drone."

She explained: every BMW transmission line had vision-based defect detection running on industrial PCs at the edge. The PCs were Intel-based, six to ten years old, and strained under modern inference workloads. Adding a new defect-detection model required either upgrading the entire PC fleet or shipping inference to a central server.

What Petra actually needed was a small, ruggedized, deterministic compute device, industrial-grade, that could run a CNN model with sub-50ms inference latency, sit on the factory floor for ten years without firmware drift, and integrate with the legacy Siemens PLCs that ran her line. Nobody made it. Nvidia's Jetson was too generic. Chinese alternatives weren't trusted by BMW procurement. Intel's industrial line was discontinued. Petra had been asking her vendors for two years and nobody had built it.

The Journey

## Building Helix Edge

The new company was Helix Edge. The first product was the HX-1, a ruggedized industrial AI compute module built on a custom ARM design with an NPU pipeline we'd originally specced for the drone. Form factor: a DIN-rail mounting box.

The first pilot with BMW closed in August 2024. Three units in one plant, retrofitted into an existing defect-detection line. €38,000. Small, but real. The deployment took six weeks. We learned more in those six weeks than in two years of drone work. The lessons were unglamorous. Connector standards wrong for industrial cabinets (we'd specced M12 D-coded; the plant used M12 X-coded). Our enclosure didn't fit the standard German DIN-rail spacing, off by 1.8mm. Our firmware update mechanism required network access, conflicting with BMW's air-gap policy.

Every issue was solvable. None would have shown up in a lab. All would have killed us if we'd shipped a hundred units before finding them.

## The First Year as a Hardware Company

By the end of 2024 we had eleven units deployed across three customers: BMW, a Bosch automotive subsidiary, and a Frankfurt specialty-chemicals manufacturer. €420,000 in revenue. €1.8M loss. We raised a €4.2M Series A in January 2025 from a German industrial VC and an Austrian family office.

The pace of hardware iteration was the hardest thing to internalize. In software you ship daily. In hardware you ship every six months at best, and every revision costs €40-80K in tooling and certification. We made the mistake of designing v2 in parallel with shipping v1. The v2 launch slipped from June 2025 to October. (We have, by now, a small ritual: every time someone proposes parallel hardware iterations, the team makes them re-read the v2 slip post-mortem.)

## The Software Stack Mattered More Than We Thought

Customers didn't buy hardware. They bought "a way to run our CV models on the factory floor without the headaches we have today." We hired a head of software in February 2025. He insisted we needed a managed cloud component, Helix Studio, handling model deployment, monitoring, and fleet updates. We launched Studio in May 2025.

Revenue from Studio in the first six months was about 60% of total revenue. The hardware was the wedge. The software was the business.

The Struggles

## The Drone Investors Were Furious

When we announced the pivot in July 2024, one of our two seed leads, a UK fund that had been excited about autonomous delivery specifically, felt blindsided. They pushed for an investor vote and lost 4-1. The partner stopped responding to our updates for four months. He didn't participate in the Series A. He sold his position in a secondary in early 2026 at a small loss.

He was right that the pivot exceeded what he'd bought into. I was right that the original strategy was a dead end. Both things can be true. There is no clean version of a pivot where every original believer stays with you.

## The Hardware Supply Chain in 2024-2025

Our NPU vendor, a Taiwanese fabless company, pushed back our delivery by 14 weeks in October 2024. The email arrived at 2 AM Munich time. I didn't sleep the rest of that night.

We held the BMW contract together by buying NPU chips from a grey-market broker in Shenzhen at 3.2x list price. €74,000 cost overrun. We absorbed it rather than pass it to BMW. The grey market is fine, until it isn't. In February 2025 we received forty chips that turned out to be re-marked older silicon. Hannes caught it on incoming QC. If we'd shipped those into BMW units, we would have failed in the field within months. After that we paid for independent IC verification on every incoming batch. €280 per reel.

## The Personal Toll

I have WPW syndrome, a heart condition that can trigger atrial fibrillation under stress. I had one episode in October 2024 that landed me in a Munich hospital for three nights. The doctor said: if I didn't change something about my stress level, the next episode would be worse.

I started running again, 30 minutes a day, non-negotiable. Hannes started joining me on weekends. The Helix Studio architecture was sketched on a napkin after a 12K run along the Isar.

I also had to renegotiate with my wife. She'd married an engineer who worked normal hours. She was now living with a founder who worked 80-hour weeks. We did couples therapy for nine months from late 2024. It probably saved my marriage. I now stop working at 7 PM most weeknights. The first time I held to that rule, I worried something would catch fire. Nothing did.

The Lessons

## What I'd Tell My Pre-Pivot Self

A few things I've learned, in the order I learned them.

§1. The thing you build is not always the thing customers buy

We built a drone. The customer bought the brain. Founders are often emotionally attached to a *form factor* when the actual value lives in a *capability*. Strip away the form. What underlying capability have you built? Who would pay for it in a different wrapper?

§2. Hardware iteration cadence is a strategic constraint

In software you ship a fix in an hour. In hardware you ship every six to twelve weeks at best, and every revision costs real money. You cannot plan a hardware company the way you plan software. The cost of being wrong about a board layout is six months and 80,000 euros, not an afternoon. Ship v1 fully, learn from the field, then design v2.

§3. The software around the hardware is the business

We thought we were a hardware company. Customer behavior told us we were a software company that ships hardware. Helix Studio is what customers renew on. The hardware is the entry point. The recurring software revenue is the moat.

§4. Pivots split your investor base, and that's okay

The UK fund partner who pushed back wasn't wrong. He was honoring the thesis he'd bought into. Some investors stay through pivots. Some don't. The ones who stay are usually the ones who invested in founders rather than the thesis. Identify which kind you have before you need them.

§5. The customer relationship is the most durable asset

Petra's cafeteria comment was the most valuable piece of information I received in two years of building a startup. It was a conversation with someone willing to be honest with a stranger she didn't owe anything to. Be in the room. Sit at tables you don't think will be productive. Talk to people who aren't on your target list. The signal you need is rarely where you're looking.

§6. Your body is the bottleneck

The hospital trip in October 2024 was a wake-up call I didn't want. I'd been treating my body as a renewable resource. It is not. The best decisions I've made for Helix Edge over 18 months were the ones I made on rested days. The worst were the ones I made in 11 PM panic windows. The math is not subtle.

§7. The pivot is not the failure

I used to think the drone-to-edge-AI pivot was evidence of my failure to read the market. It was the most important strategic move I've made. The failure would have been continuing to build drones in 2025. A clean pivot is a sign of operator quality, not founder weakness.

Dmitri Werner Almost Built Autonomous Drones. He Built Edge AI Hardware Instead.

The Problem

The Journey

The Struggles

The Breakthrough

The Lessons

§1. The thing you build is not always the thing customers buy

§2. Hardware iteration cadence is a strategic constraint

§3. The software around the hardware is the business

§4. Pivots split your investor base, and that's okay

§5. The customer relationship is the most durable asset

§6. Your body is the bottleneck

§7. The pivot is not the failure

Frequently asked

Dmitri Werner

Responses

Join the conversation