The developer who built the food recognition model for MyFitnessPal is asking a grounded question: what are people actually deploying in computer vision right now, and what still hurts? The post, from a practitioner with years of real-world experience, cuts through the hype around edge AI and API-based services to ask about trade-offs that matter. Are teams still running models on devices, or have they shifted to sending requests to cloud APIs? What drives that choice? More importantly, what problems keep blocking delivery or requiring awkward workarounds? This is the kind of honest, problem-oriented inquiry that moves the field forward. It reminds us that the most valuable insights often come from people who have built and shipped, not from those who pitch or predict. For anyone working with data tools, this mirrors the tension we see in other domains, for instance, the practical challenges of Optimize SLM: Batch Data Length, Not Individual Items, where optimization requires rethinking default approaches rather than just applying new technology.
Our take is straightforward: the real frontier in computer vision is not about building more powerful models. It is about making deployment less painful. The original poster has already shipped production systems, so the question is not whether edge models or APIs are superior in theory. It is about what actually works given latency constraints, cost pressures, privacy requirements, and team bandwidth. This is exactly the kind of practical trade-off analysis that spreadsheet users face when deciding whether to adopt AI-native tools versus sticking with traditional formulas. The same principle applies: the best solution depends on context, not on what is newest. The author's search for "problems I could build useful tooling around" is a model for how tool builders should think, start with real friction, not with a technology looking for an application. It echoes the systems-thinking approach seen in Explore the Forrester Function: Beyond Mathematics, a Tool for Machine Learning, where a mathematical function becomes a lens for understanding optimization challenges rather than just an abstract equation.
Here is the specific takeaway we would give to any reader considering a computer vision project: do not start with the model architecture. Start by mapping your deployment constraints, latency, data locality, cost per inference, and team expertise. The poster's experience with mobile-optimized models suggests that edge deployment is still very much alive, but only when the use case demands it. If your application can tolerate a few hundred milliseconds of network latency, an API might save you months of engineering. But if you need real-time inference on a device with no internet connection, edge models are non-negotiable. The pain points the poster is soliciting, blocked delivery, costly workarounds, are exactly the signals that should guide tooling decisions. For example, the challenge of batching data efficiently, as highlighted in the SLM optimization article, is a concrete friction point that could benefit from better tooling. The open question worth watching is whether the community will respond with shared solutions or if each team will keep reinventing the same workarounds.