The time since I’ve last posted has been quite busy. I’ve completely recovered from my previous eye issues and have been able to start traveling again. In fact, I’m writing this post from a hotel room.
In addition to my consulting work at Cercana, I took on a role as the CTO of Photometrics AI, a startup dedicated to applying machine learning to the optimization of adaptive street lighting. This work has the potential to provide significant reductions in energy consumption and environmental impact along with improvements in safety. In addition to reduction of light pollution that can interfere with things like bird migrations, optimized lighting can lead to lower use of fossil fuels. Additionally, there is such a thing as too much light, and using AI/ML to adjust lighting can help optimize performance to reduce glare which can increase safety where automobile and pedestrian foot traffic interact.
I won’t go into a lot of detail because more information will be coming from Photometrics in the near future. Those who know me well know that I spent many years working in critical infrastructure protection. It was doing that work that showed me that the dichotomies between national security, economic, and environmental concerns are mostly false ones. There is often a large nexus between those issues where harmonized approaches serve multiple outcomes. The work of Photometrics is but one example of that.
That’s not the only AI/ML work I’ve been doing lately. I’ve been working an early-stage project as part of a small team to use LLMs to provide targeted analytics. If that sounds vague, it’s because I intend it to at this point. This work has our team working to deeply tune RAG-based workflows and inject authoritative data using MCP interfaces. It’s exposing interesting differences/challenges/opportunities between large commercially available LLMs, such as ChatGPT and Claude, and smaller, less-finely tuned LLMs hosted locally, such as those available via Ollama. In the long run, I think the latter will be more useful for our needs.
The truly interesting thing about this work is the intersection between prompt engineering and traditional software development. Prompt engineering provides the opportunity for deep subject matter expertise to directly contribute to system development without traditional programming. Traditional programming can help guide workflows, call out to more appropriate tool chains, and standardize outputs. There is still a lot of debate around the veracity of AI and LLM outputs, but if this technology area leads to a tighter coupling between subject matter experts and system developers, that will probably be a good thing in the long run.
All of this work has gotten much more deeply involved with Python. I am doing as much Python now as I did .Net work in the early 2000s. It had already taken over the geospatial space, but it has emerged as the primary language for working with AI/ML workflows. That is probably a big factor in driving the rapidly accelerating work with geospatial AI. I think the fact that the geospatial sector had already generally committed to Python leaves it in good stead in terms of deeply integrating with AI/ML. At this point that seems equally true for both proprietary and open-source geospatial.
At the nuts-and-bolts level, I’ve become more of a fan of conda for package management. I’m doing a lot of work that uses GDAL bindings and headless PyQGIS. Each of those requires bindings to specific binaries. In the latter case, it requires QGIS to be present. Conda has made those issues transparent, especially on Linux, in ways that virtual environments and uv don’t.
When I started Cercana over two years ago, I had dabbled with AI and ML lightly over the preceding several years. It has since taken over most of my current work. Especially with regard to generative AI, there is still a lot of churn to understand its capability and appropriate use while it is simultaneously maturing rapidly, but its overall potential to accelerate delivery and orchestration of results remains exciting to me.
