Building Real Time Tools to Help Yourself: A TensorFlow.js Posture Guard

That is the small idea behind this post. The future is not waiting for someone to make the tool you need. The future is building the tool yourself, in an afternoon, against a real problem in your own life.

The pain was specific. The fix had to be specific too. A general "stand up more" reminder app does not see how I am sitting. It just nags on a timer. What I wanted was a tool that watches my posture in the browser, draws the keypoints back at me, and tells me when I am drifting forward or hunching.

That is exactly what JC Posture Guard does. The webcam stays in the browser. Nothing is uploaded. Inference runs locally on every frame. When my posture slips, the indicator changes color and I get a real-time nudge to sit back up.

The whole thing took less time than a meeting. That is the part worth paying attention to.

I want to clear something up before going further, because the words get blurred in marketing copy and on social media. TensorFlow.js is a machine learning library. It is not an AI assistant. The two are related, but they are not the same.

Artificial intelligence is the broad goal — machines that perform tasks we associate with human intelligence: reasoning, planning, generating language, making judgements, holding a conversation. Machine learning is one technique for getting there. It is a subset of AI focused on systems that learn statistical patterns from data and then use those patterns to make predictions.

Posture detection sits squarely in the ML half of that picture. The model has been trained on millions of labeled images of human bodies. It does not "understand" your spine. It maps pixels to coordinates of joints. That is pattern recognition, not cognition.

Machine learning is about prediction over patterns. You give a model a lot of examples, it learns the statistical regularities in those examples, and at inference time it produces an output that fits the pattern. Image classifiers, speech-to-text, recommendation engines, fraud scoring, pose estimation. This is all machine learning.

A trained ML model is a frozen function. It is fast, deterministic given the same input, and it does not reason about its output. If you show it something far outside the training distribution, it will still confidently produce a guess. That is a feature in some contexts and a serious risk in others.

When most people say "AI" today, they mean large language models like the ones behind chat assistants. Those systems are also built using machine learning under the hood. They are trained on huge amounts of text, but the surface behavior feels different. They generate language, follow instructions, hold context across a conversation, and reason about ambiguous prompts.

That is the experience that has reshaped expectations. AI in the modern sense is conversational, generative, and flexible. ML in the older sense is narrow, fast, and focused. A single trained model doing one job very well.

The distinction matters because it changes what you ship and how it runs. Posture detection in the browser is narrow ML — a single pre-trained pose model running locally, 30+ frames per second, no API calls, no token costs, no latency from a server round trip.

If I had built this with a chat AI behind it, every frame would mean a network request, real money per call, and a feedback loop too slow to actually catch posture in real time. Choosing the right tool comes down to honest engineering: what does the problem actually need?

JC Posture Guard uses a pose estimation model that takes a video frame and returns the pixel coordinates of seventeen keypoints such as eyes, ears, shoulders, elbows, wrists, hips, knees, ankles. From those coordinates you can derive everything else: shoulder angle, head-forward position, slouch detection, even symmetry between left and right.

The math is light. Once the keypoints come back from the model, posture rules are just geometry. Is the head past the shoulders? Are the shoulders below their baseline? Is the back curved beyond a threshold? That is not AI deciding for you. That is plain trigonometry on top of a fast ML output.

The win is that the whole pipeline runs in your browser tab. No server, no upload, no privacy compromise. The webcam frame never leaves the page.

The most interesting tools coming next will combine both. ML handles the perception: vision, audio, sensor streams, pose, gesture, intent classification. AI handles the conversation around that perception:explaining, suggesting, summarizing, planning.

A few realistic combinations:

The pattern is the same in each case: fast, narrow ML at the edge, flexible AI on top of the result. That is the architecture worth learning.

The barrier to building has dropped. A pre-trained pose model is a few imports away. A chat model is an API call away. A static site host gives you a public URL in minutes. The skill that matters now is noticing a friction point in your own day and turning it into a working tool before the day ends.

You do not need a startup. You do not need a roadmap. You need to notice the small thing that is wasting your attention or hurting your body and decide that you are allowed to fix it. JC Posture Guard exists because I got tired of slouching at my desk. That is the entire origin story.

If you want to see it run, open JC Posture Guard in a browser, give the page permission to use your camera, and sit at your desk the way you usually do. The keypoints will draw onto the live feed. Adjust your shoulders, lower your chin, sit back — the indicator will respond.

Nothing about your video leaves the page. The whole thing is TensorFlow.js running locally.

The source is on GitHub at jessicacousins/tensorflow_posture.

Machine learning is a focused tool. Artificial intelligence is a flexible reasoning partner. They are not interchangeable, and pretending they are is how people end up overpaying for a feature they could ship for free.

When something in your day starts costing you attention or comfort, stop tolerating it. Build something. Ship it before dinner. Use ML where ML belongs and AI where AI belongs. The future is not handed to you. It is the one you build for yourself, in real time, against your own life.