I’m building PathWild.ai—an AI-powered platform for predicting wildlife activity patterns. But this isn’t just about the destination. This series will document everything I learn along the way, forcing me to understand AI/ML concepts deeply enough to explain them clearly. If you’re looking to build your own AI/ML project as a beginner, I hope this journey helps you too.
Why I’m Building PathWild
I’m currently a Director of Software Engineering at AWS and I’m fascinated by AI/ML. I’m soon transitioning into a new role focused on AI transformation, and I need hands-on AI/ML experience, fast. I also happen to be an elk hunter with a personal hunt planned for October 2026 in Wyoming.
PathWild serves both purposes: it’s a real commercial ML platform I can build and potentially monetize, and it’s my vehicle for learning AI/ML by doing rather than just reading about it.
The core problem PathWild solves? Predicting where wildlife will be active based on environmental conditions, historical patterns, and real-time data. Think of it as a weather forecast, but for elk movement patterns.
What I Hope to Get Out of This
For my career: Practical, hands-on AI/ML experience that I can immediately apply in my new role. Theory is valuable, but I learn best by building.
For this project: A working ML platform that can actually predict wildlife activity patterns with enough accuracy to be useful and ethical. Success means I can use it for my 2026 elk hunt and potentially help other hunters make better decisions.
For this blog series: By explaining what I’m learning, I’ll be forced to understand it at a deeper level. The Feynman technique in action—if I can’t explain it clearly, I don’t understand it well enough.
The Framework: Building ML Powered Applications
I’m generally following the approach outlined in Emmanuel Ameisen’s excellent book “Building Machine Learning Powered Applications.” The book presents a pragmatic four-part framework that focuses on building ML systems that actually work in production, not just in notebooks.
Here’s how I’m applying it to PathWild:
Part 1: Find the Right ML Approach
This is where most beginners get it wrong—they jump straight to models. Ameisen argues you need to start with fundamentals:
Define a clear product goal. For PathWild, that’s: predict the location and population size of elk for a given location and date range. Notice this is a product goal, not a technical goal. I’m not saying “build a regression model” or “achieve 95% accuracy.” I’m defining what users need.
Determine if ML is the right approach. This seems obvious, but it’s critical. Could I solve this with rules alone? With a database lookup? With traditional statistics? ML is powerful but complex—you should only use it when simpler approaches won’t work. For wildlife prediction, the interaction between environmental factors (temperature, pressure, wind, elevation) is non-linear and seasonal, which makes ML a good fit.
Build heuristics based on domain knowledge. Before writing ML code, encode what we already know:
- Elk move to higher elevations as temperatures rise in late summer
- They’re most active during dawn and dusk (crepuscular behavior)
- Wind direction affects their movement patterns for scent detection
- Barometric pressure changes often precede increased activity
These heuristics serve three purposes: they create a working baseline system, they give us features to test in ML models, and they provide a benchmark—if our ML model can’t beat well-crafted heuristics, it’s not ready.
Define the product shape by designing the inference API. This is the interface users will interact with. What inputs do they provide? What outputs do they get? How is uncertainty communicated? For PathWild, the API might look like:
Input: location (lat/lon), date range, weather forecast Output: predicted activity zones, confidence scores, explanation
The “explanation” is crucial. A prediction without context is just a number. Users need to understand why the model made its prediction.
Parts 2-4: The Path Forward
The subsequent parts of Ameisen’s framework will guide the rest of this journey:
Part 2: Build a Working Pipeline – Moving from prototype to reproducible data collection, feature engineering, and model training workflows.
Part 3: Iterate on Models – Experimenting with different approaches, evaluating performance, and understanding what works (and what doesn’t).
Part 4: Deploy and Monitor – Getting the model into production and ensuring it continues to perform well over time.
Each of these parts will be covered in depth through future blog posts, with real code examples from PathWild.
What’s Next
I’ll be documenting my progress through each phase of this framework. Early posts will focus on Part 1—building the inference prototype and scoring algorithm based on domain heuristics. Then we’ll move into building data pipelines, training models, and eventually deploying a production system.
I’m not following a rigid timeline. Some weeks I’ll make huge progress, other weeks I’ll hit dead ends. I’ll document all of it—the breakthroughs and the frustrations.
I’m not an AI/ML expert. I’m learning this alongside you. That means I’ll make mistakes, get things wrong, and have to backtrack. That’s the point. If you’re also trying to break into AI/ML, I hope seeing the messy reality of learning helps more than another polished tutorial.
Follow Along
I’m building PathWild in the open. Every struggle, every breakthrough, every “why isn’t this working?” moment will be documented here. If you’re trying to break into AI/ML, or if you just enjoy watching someone learn by doing, I’d love to have you follow along.
Next post: Building the first heuristic-based prediction
This is post 1 in a series documenting my journey building PathWild.ai. Follow along as I learn AI/ML by building a real wildlife prediction platform.
Recommended reading: “Building Machine Learning Powered Applications” by Emmanuel Ameisen
Project: PathWild.ai
Learning approach: 80% doing, 20% theory
Current focus: Part 1 – Finding the Right ML Approach