We design noise-robust, ultra-compact AI architectures that deploy on $5 microcontrollers. From voice interfaces to multi-target tracking.
Each technology addresses a distinct sensing modality with a unified design principle: structural robustness without runtime overhead.
Noise-Conditioned State Space Model for always-on keyword spotting. Five noise-conditioning mechanisms achieve structural robustness without separate denoising — works on $5 MCUs with 7ms latency.
NC-Conv extends noise-conditioning to vision. A learned quality gate blends dynamic and static convolution paths — enabling robust lane detection through fog, rain, tunnels, and low-light on an $8 MCU.
COP resolves up to 2ρ(M−1) sources from M sensors via 4th-order cumulants, coupled with RFS multi-target tracking. Mamba-COP-RL adds a selective SSM temporal encoder with PPO-based adaptive track management — achieving GOSPA −4.5% and false tracks −8% over baseline.
First sub-100K parameter end-to-end spoken language understanding with few-shot intent addition. NC-OPAL two-stage incremental learning (prototype imprinting + LoRA + KD) adds new voice commands from just 20 examples — 23× smaller than SpeechCache.
5KB to 20KB INT8 models. Fits in on-chip SRAM — no external memory needed.
Deploys on $5 ARM Cortex-M MCUs. No cloud, no latency, no recurring cost.
Structural robustness baked into the architecture. No denoising module required.
Patent portfolio filed (KR + US). PCT international filing in progress.
Structurally noise-robust SSM via Spectral-Aware SSM (SA-SSM) with Δ-modulation and B-gating
Noise-conditioned SSM architecture with formal robustness analysis
2ρ(M−1) source resolution coupled with PHD multi-target tracking
First sub-100K SLU with NC-OPAL incremental learning — 3-way backbone comparison (NC-TCN / NC-TCN-Bi / NC-SSM) on Fluent Speech Commands
Quality-gated dynamic/static path blending for adverse-condition vision
5 independent claims + 20 dependent claims covering audio, vision, and sensor modalities
Covering NC-SSM, DualPCEN, selectivity modulation, and hardware accelerator specs
NanoAgentic AI is an IP-driven research company focused on building the smallest, most robust AI models for edge deployment. Our work spans audio intelligence, computer vision, and signal intelligence.
Every architecture we design follows a single principle: structural robustness without runtime overhead. Rather than bolting on denoising modules or augmentation hacks, we engineer noise immunity directly into the model's computation graph.
Our technologies are protected by a growing patent portfolio and validated through peer-reviewed publications at top-tier venues including IEEE, Interspeech, ACCV, CVPR, and ICCV.
We license our technologies for commercial deployment. Let's talk.
Get in Touch