Building a production-ready TinyML model normally requires a mixture of data engineering, signal processing, machine learning, and embedded software expertise — four disciplines that rarely live in one engineer. SensiML’s Analytics Studio has automated the hardest parts of that chain with AutoML for years, but a person still had to drive it, screen by screen, decision by decision.

That changes today. We’re introducing Agentic AI Assist mode: a SensiML Agent panel that sits alongside your project in Analytics Studio and does most of the work of getting from raw sensor recordings to deployable firmware. Agentic AI Assist is live now in public beta at app.sensiml.cloud, and we put together a 10-minute demo that compresses the entire workflow — data to a compiled, optimized model — into one sitting.

Watch the full demo above, or keep reading for a walkthrough of what happens at each step.

Not a chatbot telling you which buttons to press

It’s worth being precise about what “agentic” means here, because the term gets used loosely. The SensiML Agent can inspect your project data, create queries, generate synthetic training samples, configure preprocessing and model pipelines, launch training jobs, analyze the results, and prepare deployable embedded code. Every pipeline-altering step — building a query, kicking off training, compiling firmware — comes back as an explicit approval card (“this action needs your approval before it runs”) that you confirm or cancel, so the work stays visible and reviewable the whole way through. The agent is performing the work, not just describing which menu to click next.

That distinction is the whole point. AutoML already removed the guesswork from model architecture and hyperparameters. What it never removed was the orchestration around it — reviewing a dataset, deciding on a query and label scheme, kicking off feature extraction, checking whether the winning model actually fits your flash and RAM budget, and only then generating firmware. That orchestration is exactly the kind of multi-step, tool-using task agentic AI is good at, and it’s also the part of the workflow that has kept edge AI out of reach for engineers who don’t live and breathe ML pipelines every day.

What the demo shows

The video builds a project called Directional Keyword Spotting, with a goal stated right in the project description: “Create a low power, microphone input keyword spotting model that controls a toy robot’s movements from a user’s spoken commands.” The stated requirements: classify the spoken keywords “Up,” “Down,” “Left,” and “Right,” hit at least 90% validation accuracy, and deploy on the Silicon Labs xG24 Dev Kit.

The project starts with 1,183 imported one-second WAV captures — spoken directional keywords plus “unknown” and noise clips as negative examples, pulled from Google’s Open Speech Commands dataset — and zero queries, pipelines, or models. From there, the whole thing plays out as a conversation with the SensiML Agent, next to the normal project workspace:

1. “Analyze the data we have.” The agent reviews the 1,183 captures, summarizes the five classes and their counts, and flags that there’s no query, pipeline, or model yet — with next steps to get there.
2. “Create a query.” You can still adjust parameters by hand before it runs — here, that means keeping “Up,” “Down,” “Left,” and “Right” as the labels the model should target and excluding the catch-all “Unknown” class, narrowing the training set to 954 labeled segments across the four directional classes. The agent returns a histogram showing class balance once the query executes.
3. “Will this data reach 90%+ accuracy?” Asked to speculate before any training happens, the agent reasons through the specifics — sample volume (954 segments), class balance, signal quality, and speaker diversity (4 speakers, 2 genders, 3 recording sessions) — and concludes it’s “very likely.”

4. “Build the model.” The agent proposes a keyword-spotting pipeline and presents several model variants trading off size and latency against accuracy — from a ~13 KB / ~42 ms option up to a ~47 KB / ~300 ms one — with a ~25 KB, ~130 ms variant marked as its recommendation. You can override the pick; every option still needs your confirmation before the pipeline is created.

5. Training runs, trackable step-by-step, through preprocessing (query input, windowing, segment filtering), feature extraction (feature generation, quantization, transform), and model training (TensorFlow Lite for Microcontrollers, transfer learning, stratified validation split) — all visible in a live console log. The resulting model, KWS_Directional_v1_Fold_0, comes back at 92% accuracy (91.62% precisely), above the 90% target, with a confusion matrix, feature embeddings, and model summary available to explore further.
6. “Compile a model for the Silicon Labs xG24 platform.” The agent proposes a specific build — the trained model, compiled for the EFR32MG24B310F1536IM48 processor on the xG24 Dev Kit, as a static .lib with the Silicon Labs MVP AI accelerator enabled — and explains why: that accelerator offloads audio feature extraction and inference, and the static library links straight into your C/embedded application without modification.

7. Confirm, and click Build Knowledge Pack. A zip file lands in the downloads folder a few seconds later — the compiled Knowledge Pack, ready to pull into an embedded project, flash, and test.

Raw dataset to a compiled, optimized model for the xG24, in about ten minutes — worth watching in full to see how the agent handles the judgment calls along the way, not just the happy path.

More than this one demo

Agentic AI Assist is the newest way in, but it sits on top of the same Analytics Toolkit SensiML has always shipped. Beyond the agent workflow shown here, that includes a richly featured desktop application for production-grade sensor data management, generative synthetic data augmentation, and platform-specific profiling for memory use and execution time — plus model templates for predictive maintenance, anomaly detection, gesture recognition, audio classification, and more.

Who this is for

Agentic AI Assist mode doesn’t replace the control that SensiML’s existing tools give experienced ML engineers — Data Studio, Analytics Studio, and the Python SDK are all still there when you want to drive the pipeline by hand or automate it in your own CI. What it adds is a much shorter on-ramp:

• Firmware and product engineers who understand their sensor and their hardware constraints but don’t want to become feature-engineering experts to get a working model.
• Teams evaluating edge AI who want to go from a raw dataset to a running proof of concept in one sitting, before committing engineering time to a full pipeline build-out.
• Experienced SensiML users who want the agent to handle a first pass — query setup, model search, a baseline Knowledge Pack — and then take over for fine-tuning from there.

Try it in public beta

Agentic AI Assist mode is live now in public beta at app.sensiml.cloud. Sign in, point it at a labeled dataset, and describe what you’re trying to build — the fastest way to see what it can do is to watch it work through a real dataset in the full 10-minute demo above, then try the same thing on your own sensor data.

As with any public beta, we’re actively refining Agentic AI Assist based on what real pipelines throw at it, so if you hit a rough edge or a workflow it doesn’t yet handle well, we want to hear about it.