Why Heimdall, and why now

Charlie sent me a video on Facebook, some bloke who'd built an AI-driven laser that zaps bugs out of the air. His note said a mutual mate and I had talked about exactly this kind of thing years ago and reckoned it'd be cool to build. I thought, why do something someone's already done, and asked him back: "what if you could do something like that for foxes?"

I'd just been reading about Shinkei↗, a company building robots that use computer vision to perform ikejime, an old Japanese method of killing a fish instantly and without stress. Their whole pitch is that it's more humane and that it adds value for everyone in the chain instead of extracting it. Foxes aren't fish, but the shape of the idea stuck: see the animal clearly, act precisely, don't make it suffer.

So we tried to design the fox version of that. An autonomous rig that spots a fox, confirms it, and deals with it. And we couldn't make it work, on two counts, one practical and one that mattered more.

The practical one: Shinkei's computer vision is genuinely clever, and it works because a fish on a deck is a controlled problem. A fox at the far end of a paddock, at night, through dust, rain, mist and a lens the local spiders treat as prime real estate, is not. We're not going to pretend we can match that kind of precision in the field any time soon. And even if we could, you can't reliably stalk and kill a moving fox without putting it through exactly the distress we were trying to avoid.

The one that mattered more was ethical. The welfare case for a machine that decides, on its own, to pursue and kill an intelligent animal in a paddock didn't hold up. We kept circling back to the same conclusion: this shouldn't be automated, and it shouldn't be a robot pulling the trigger.

The breakthrough was subtraction. Don't build the killing machine at all.

Instead, send the farmer a notification and a live position of the fox. A person who already knows proper, lawful shooting practice makes the call and does the rest, humanely and accountably. Taking ourselves out of the kill didn't weaken the idea. It was the idea. Heimdall is a watchman, not a weapon. An AI-powered shepherd. We landed there because the alternative didn't sit right with us, not because it made a nice slogan.

Two things make this the moment.

The first is bigger than us. Agriculture is slowly walking away from blunt, indiscriminate chemistry. In weeds, that looks like laser weeders replacing blanket glyphosate spraying: targeted, plant-by-plant, instead of soaking a paddock. We think the same shift is coming for pest control. 1080 is the blanket spray of the animal world. It works, and it's genuinely effective, but it carries real risk to anything else that finds the bait. Working dogs are among the most susceptible animals to it, which is why its use is tightly permit-controlled. PestSmart / NSW DPI↗ Heimdall is aiming to be the animal kingdom's answer to the laser weeder: the same move from indiscriminate to targeted, with a human making the final decision. We're not claiming we've matched what the laser-weeder companies have built. We're saying that's the direction, and it's where we're pointed.

A leading startup accelerator recently named "AI for low-pesticide agriculture" as a frontier worth building in. We'd already started down this road, so it was good to see the wider world arrive at the same place. There are encouraging shifts happening already. PAPP, for instance, is a real step forward on 1080: it acts faster, the RSPCA↗ regards it as more humane, and it has an antidote, so a dosed working dog can sometimes be saved (if a vet gets to it within half an hour). But it still isn't a clean answer. It can't be aerially baited, partly because it also kills goannas, and it runs materially dearer than 1080, at launch several times the price per bait (ABC, 2016)↗. A kinder poison is still a poison, and a dearer one at that. We think the bigger prize is needing to reach for one less often.

The second reason is simpler, and more of a lifelong thing: the urge to leave some small corner of the world a little better than we found it, if we can. And the tooling has finally caught up. A two-person team can now build something like this, fast, because AI does a huge amount of the heavy lifting on the programming side. We direct it and we check its work to make sure it does what we intended. That means a company like ours can exist without handing the design to people who've never set foot on a farm or in a workshop and don't understand the ground they're building for. Charlie and I stand in the paddock. We believe in "go and see". We'd rather build it ourselves, from there.

I won't dress this up. Here's the real status:

• The first node cluster is built. The MVP is assembled and in bench testing, and we're running field tests on the power budget right now.
• A first full deployment at a farm near Nairne, South Australia is scheduled for early January 2027.
• We're pre-training an initial fox-recognition model on open-source and Creative Commons imagery, purely as a head start so the field system isn't learning from zero. It is not validated on real-world footage yet.
• There are no field results. Yet. The trials exist to find out what actually works, what breaks, and what's accurate enough to be useful as we iterate.

If you want the measured answers, identification distance, fox-versus-dog confusion, how well the nodes hand a track along a fence line, those don't exist yet, and I'm not going to pretend they do. That's what our trials are for.

Every grazier we talk to knows 1080 works. We're never going to pretend it doesn't. But our problem is that it's indiscriminate. Anything that eats the bait is at risk, and often you never even find out what you lost. A poisoned working dog and a dog that got into snail pellets can look the same, and nobody spends money on a necropsy for a dead farm dog. AWI 1080 first-aid guide for dogs↗

My cofounder, Charlie, put it in a way that was simultaneously blunt and very accurate. He told me he'd "rather be set on fire than ingest 1080", because at least the fire would be quicker. I haven't been able to shake that since.

So we're honest about where we fit. For now, Heimdall is an auxiliary: targeted evidence that works alongside whatever control a farmer already uses. The aim, if the field evidence earns it, is to eventually become a replacement, a reason to reach for the poison less often. That's a goal, not a promise, and it's gated on proof we don't have yet.

Between us, Charlie brings twenty years of farming alongside his family, plus a career in film location logistics, reading a landscape and coordinating people across it. I bring workshop floors, mine-spec vehicle fit-outs and a Lean background: sourcing a tight parts list and building rugged kit that survives the real world.

What we're both allergic to is the sealed-box model that's crept into farm machinery. Gear you've paid for in full but aren't allowed to fix yourself, despite generations of us fixing things with duct tape and fencing wire. Products that phone home for permission and quietly die when a vendor decides it's end-of-life. You know the type of business we're talking about. We want the opposite: hardware you own, can repair, and can keep using on your own terms. Farmers as colleagues, not customers to be locked in.

And we don't romanticise the fox. A disembowelled lamb is worse than a dead fox, and I'd never argue otherwise. But I refuse to accept that a slow, indiscriminate poison is the only tool on the shelf. We think technology can give farmers a third option: act earlier, act precisely, and give even a pest a cleaner end than it would otherwise get.

Start with the fox, because that's where we start. Around $17.5 million is lost by sheep farmers every year directly to foxes, with fox-related costs for the agricultural sector ballooning to between $198 and $228 million annually. Ag Victoria; SSAA↗ And if you look at the other marauders in the sector, furry or otherwise, that total could sit north of $350 million a year.^{Ag Victoria↗, ABC News↗} These are government and industry guesstimates, not audited ledgers, but the order of magnitude is the point.

A fox is where the idea starts, not where it ends. The same nodes that notice a fox slipping through a fence line can notice other things that move across a property at odd hours. Different problems, same blind spot: by the time you know, it's already happened. The principle doesn't change across any of them. Heimdall tells you what moved, and where, and when, and leaves the judgement to you. Never who someone is, or what they were up to. Just the movement, early enough to be useful.

We can't undo the fox being here. That ship sailed a long time ago. But if we keep going down this road, we think we can get pretty close to a better deal: for the farmer, for the stock, and even for the animal we're there to stop.

Leave the place a bit better than we found it. That's the whole idea.