Pantera
A hybrid edge/cloud AI platform covering the full cycle — prevention, detection, response, impact.
It combines those cameras with satellite feeds and fire risk modeling to catch wildfires early, taking on one of Brazil's largest sources of CO₂ emissions. Detection accuracy sits at 95% — that's the number people ask for. The number that decides whether it was worth building is minutes-to-alert, and what happens in the hour after.
Some of that ground is the Pantanal and indigenous territories, where losing a few hours of detection time is the whole story. The rest is pulp/paper, agriculture, and conservation — clients who measure us in hectares that didn't burn.
I own the architecture end to end — edge inference, the hybrid on-prem/cloud platform, the ML behind detection and risk. Pantera itself is built by engineers, ML specialists and environmental scientists, and most of what it knows about fire, I learned from them.
Approach
Three things I've come to believe, the expensive way.
The model is the easy part. What breaks in the field is dust, lightning, heat, a lens nobody cleaned, 4G that drops for six hours. Ninety-nine percent of what's written about ML is about the model. Almost none of it is about what happens when inference runs on a pole in the middle of nowhere — which is where the system either works or doesn't.
Accuracy is a vanity metric on its own. The honest questions are what a false negative costs when it's a real fire, and what a false positive costs when you wake a brigade at 3 a.m. Those two numbers are not symmetric, and they're not the same for every client. I build to the operator's cost function, not to a leaderboard.
Research and operations are two different sports. I contribute to an academic fire simulator and I help run a system people depend on tonight. Standing on both sides is rare, and most of what I think comes out of the gap between them.
Elsewhere
Wildfire is where I learned this — unforgiving in every direction at once — and the discipline travels further than the domain does.
Agents running our own operation. Eight of them cover sales, customer success, support and marketing at 1.5°C, orchestrated on Hermes, an open-source framework I deployed on our own servers. An agent that quietly does the wrong thing is far worse than one that fails loudly, so most of the work isn't the prompt — it's the guardrails, the handoff back to a human, and being honest about what should never be automated at all.
ForeFire, since 2022. An open-source wildfire simulation engine in C++ built by CNRS at Université de Corse. Mostly plumbing: Dockerizing it, setting up CI/CD, and writing docs people can actually follow — the unglamorous work that decides whether research code ever leaves the lab.
Now
Two directions I'm paying attention to.
Climate and public technology. Fire, deforestation, monitoring — systems that governments, NGOs and communities actually operate, rather than pilots that end with the press release. This is where policy, community and engineering have to meet, and where I think I'm most useful.
Agentic systems outside of AI companies. The interesting problem isn't the models, it's putting them inside an operation where being wrong has a cost — which is most of the economy, and almost none of the current discourse.
Tech alone doesn't solve the climate crisis. It needs policy, community, and people who show up for the parts nobody claps for — that's the part I care about most.
None of this is work anyone does alone — the useful part almost always shows up when people who see the problem from different sides end up in the same room. If you're working on something in either direction, write to me: to trade notes, to collaborate, or to build something together.
Updated August 2026
Off the clock
I run — marathon PR sub-3:23 — cycle, play sax in a carnival bloco, and spend as much time as I can with family and friends, traveling when it fits. None of it is a metaphor for the work. It's just the rest of the life.