The model you need, built on the hardware you’re shipping

If off-the-shelf models don’t meet your requirements for accuracy, latency, or memory, or don’t fit your use case, we develop a custom, production-ready model tailored to your product, deployment environment, and hardware platform.

What we build

Six ways embedded AI engagements take shape. Most projects combine two or three.

New Model Development

Architecture designed from requirements to binary. Custom topologies when no existing model fits the use case, the data, or the silicon.

Custom Dataset Training

Retrain proven architectures on your proprietary data. Higher accuracy in the conditions you actually
deploy into.

Model Optimization

Quantization, pruning, layer fusion, and operator mapping to hit your hardware's compute and memory budget without losing accuracy.

Dataset Creation

Collection protocols, labeling pipelines, augmentation strategy, structured datasets ready for training.

Model Porting

Move existing models across NPUs, MCUs, and silicon vendors. Re-validate accuracy and latency on the new target.

MLOps Pipeline

Continuous retraining, OTA model deployment, drift monitoring, and field telemetry integration.

Selected work across silicon platforms and use cases. Numbers are from production deployments and active engagements.

On-device face recognition across four silicon platforms

Built a complete face recognition pipeline – detection, alignment, feature extraction, recognition and deployed it across STM32N6, Alif Ensemble, Infineon PSoC Edge E84, and Synaptics Astra. Production performance on STM32N6: ~70ms end-to-end authentication, 96%+ accuracy, 1.7MB SRAM and 2.6MB Flash. Now offered as a licensable model: NovaEyeD.

On-device Contextual Language Model for venue and product Q&A

Trained and deployed a Contextual Language Model running entirely on an embedded edge processor — no cloud round-trip, no API calls. Three cooperating models on-chip in 146.8 MB total, fine-tuned to answer location- and product-aware questions about a specific physical environment. Use cases include retail product-to-aisle mapping, appliance user manuals turned into conversational Q&A, and venue navigation with walking-time estimates.

Wake-word-triggered ingredient recognition in 420 KB

Wake-word detection triggers a camera-based object identification model (FastestDet) that recognizes food ingredients in real time, then a recipe generator suggests dishes from what it sees. Two cooperating models in 420 KB combined, running on an MCU-class edge processor — the entire pipeline runs hands-free and entirely on-device.

Silicon depth

Active ongoing programs with the silicon vendors shaping the edge AI category.

How We Engage

Selected work across silicon platforms and use cases. Numbers are from production deployments and active engagements.

Fixed-Scope Project

Defined deliverables, timeline, and cost. SOW-based. Most common starting point for engagements with a clear use case, dataset, and silicon target. Typical duration: 4–16 weeks depending on complexity.

Ongoing Partner Program

Long-term engagement for silicon vendors and OEMs needing continuous model development, SDK maintenance, and roadmap alignment across an evolving product portfolio.

NDA available before any technical discussion.

From scoping call to production

Step 01 — Scoping call

A focused 30-minute conversation to understand your use case and constraints — no obligations, no sales pitch. Just a clear-eyed look at whether we're the right fit.
⏱ 30 minutes ✦ No charge

Step 02 — Proposal

Within 5 business days you'll receive a detailed proposal: scope, deliverables, timeline, and cost. Everything documented so there are no surprises.
5-day turnaround ✦ Scope & cost

Step 03 — Kick-off

Once aligned, we get everything in motion: data access, hardware provisioning, and introductions across both teams so work can begin immediately.
✦ Data access ✦ Hardware setup ✦ Team intros

Step 04 — Development

We build iteratively with clear milestones and performance reporting at every stage. Weekly reviews keep you informed and in control throughout the build.
✦ Iterative milestones ✦ Perf reporting ⏱ Weekly reviews

Step 05 — Delivery & handover

The finished system delivered in full — compiled binary, weights, complete documentation, and dedicated integration support. You're ready to go live.
✦ Compiled binary ✦ Weights 📄 Documentation ✦ Integration support

Why teams work with ModelNova

Building edge AI internally often means long development cycles, fragmented tooling, silicon optimization challenges, and deployment risk. ModelNova reduces that complexity with a complete engineering engagement designed to accelerate deployment.

Faster time to market

Skip 12+ months of model development and validation by starting from a production-validated baseline.

Built right the first time

The same team begin Fusion Studio and the ModelNova Zoo builds and validates your model.

Scale without hiring

Gain access to our parent company, embedUR’s bespoke embedded engineering services backed by 450+ experts

Maybe you don't need Custom Model

If your use case overlaps with something we’ve already built, you’ll get to production faster with a licensable model or our development platform.

NovaEyeD

Production-ready face recognition for edge silicon.

Fusion Studio

End-to-end edge AI model development and optimization platform.

ModelNova Zoo

Curated catalog of pre-trained, edge-optimized models.

Ready to scope your model?

30-minute call. No charge. NDA available on request.