# SutraAI — End-to-End Product and Solutions Catalog (v4)

What SutraAI offers: **product modules**, **ML blocks**, the **4 reference model types**, **solutions by application and industry** (Edge Impulse parity + India-focused verticals), **hardware targets**, and the **6 reference templates** proven in notebooks in Month 1.

**Phase key:**
- **P1 (Month 1):** the model type works end-to-end in Jupyter notebooks on the Jetson (train → int8 export → profile → simulated device), plus a minimal web slice in Week 4.
- **P2 (Months 2–3):** full web platform (Studio, SOC API, public projects), chip SDKs, and the new blocks B3, B7, B4+.
- **P3 (Month 4+):** advanced blocks.

**Device key:** **MCU** = ESP32-S3 / Silicon Labs xG24 / Nordic / STM32 / Renesas RA · **Edge** = Jetson / Raspberry Pi 5 / Linux gateway · **Cloud** = SutraAI Cloud API · **Phone** = browser/mobile

> **v4 change:** the project is renamed **SutraAI**, and phases now follow the notebook-first plan. Audio (**B6**) and centroid detection (**B4**) are still in **P1**. Full bounding-box detection is P2.

---

## 1. Product Modules

| Module | What it does | Edge Impulse equivalent | Phase |
|---|---|---|---|
| **ML core (`ml_core`)** | The pipeline engine: input → DSP → learning → int8 export → profile, proven in Jupyter notebooks first | Impulse engine | **P1** (notebooks, then a Python package) |
| **Sutra Studio** | No-code web app: Data → Pipeline → Train → Results → Deploy → Devices → Versions | Studio | P1 minimal slice (upload → train → download) · full app P2 |
| **Sutra Data** | Upload, import (folder, YOLO, COCO, VOC, CSV), SOC ingestion, phone-as-device; viewers for images, audio + spectrogram, sensor waveforms | Data acquisition | P1 zip upload · rest P2 |
| **Sutra Label** | Keyboard class labeling, **bounding-box labeling**, fix-from-results | Labeling | P1 external tool (Label Studio) · built-in P2 · AI-assisted P3 |
| **Pipeline Designer** | Input → Processing (image / raw / spectral / MFCC / MFE) → Learning block, with feature preview | Impulse design | P1 as `pipeline.yaml` · visual designer P2 |
| **Sutra Train** | Presets, live progress, CPU/GPU workers | Training | P1 |
| **Sutra Test** | Val/test × float32/int8 accuracy, confusion matrix, detection F1, live test (upload, webcam, mic) | Model testing / live classification | P1 (notebook + basic web results) |
| **On-Device Profiler** | Latency, peak RAM, and flash per target (estimated in P1, measured on real boards in P2) | On-device performance | P1 |
| **Sutra Deploy** | TFLite int8, C array, ONNX, ONNX Web, TensorRT script, Cloud API | Deployment | P1 (file exports) |
| **Sutra SOC API** | Device registry, ingestion, OTA model delivery (incl. DSP config), cloud inference, telemetry | Ingestion + Remote mgmt APIs | P2 (flow proven offline in P1 with the notebook device simulator) |
| **Sutra Fleet** | Devices, channels (dev/stable), OTA, telemetry, review queue | Devices | P2 |
| **Sutra Hub** | **Versions (`live` / `latest` / `vN`), public project pages, README, licence, views, clones, Clone** | Public projects | P2 |
| **Sutra Assistant** | Nemotron + RAG help bot, project-aware Training Advisor | — (differentiator) | P1 notebook prototype · service P2 |
| **Sutra Tuner** | Device-aware search for the best config that fits the target | EON Tuner | P2 |
| **Sutra SDKs** | Python SDK + `sutra` CLI; C SDKs for ESP-IDF / Silicon Labs / Renesas FSP / Zephyr with DSP in C | SDKs, CLI | P2 |
| **Sutra On-Prem** | Same stack inside the customer's network, offline-capable | Enterprise | P2 |
| **Sutra Vision Suite** | Box detection, visual anomaly, counting/tracking, OCR | Visual Inspection Suite | P2–P3 |

---

## 2. ML Blocks

| ID | Block | Processing | Learning | Runs on | Phase |
|---|---|---|---|---|---|
| **B1** | Image classification | Image (resize, gray/RGB) | Transfer learning (MobileNetV2/V3, EfficientNet) | MCU · Edge · Cloud · Phone | **P1** |
| **B2** | Sensor / time-series classification | Raw, spectral (FFT, RMS) | 1D-CNN, **tiny Transformer** (experimental on MCU) | MCU · Edge · Cloud · Phone | **P1** |
| **B4** | Object detection: centroid | Image | MobileNetV2 stride-8 + per-cell class head (location + count) | MCU · Edge · Cloud | **P1** |
| **B4+** | Object detection: bounding boxes | Image | Apache/BSD-licensed box detector | Edge · Cloud | P2 |
| **B6** | Audio classification / keyword spotting | **MFCC, MFE (log-mel)** | 1D/2D-CNN | MCU · Edge · Cloud | **P1** |
| **B3** | Sensor anomaly detection | Spectral | Autoencoder / Gaussian mixture | MCU · Edge · Cloud | P2 |
| **B7** | Regression | Raw / spectral | MLP / 1D-CNN | MCU · Edge · Cloud | P2 |
| **B5** | Visual anomaly detection | Image | PatchCore/PaDiM-style | Edge · Cloud | P3 |
| **B8** | OCR | Image | Detector + recognizer | Edge · Cloud | P3 |
| **B9** | Counting + tracking (video) | Image sequence | Detector + tracker | Edge · Cloud | P3 |
| **B10** | Pose / keypoints | Image | Lightweight pose model | Edge · Cloud · Phone | P3 |
| **B11** | GenAI data tools | — | VLM auto-labeling, synthetic data | Cloud | P3 |

---

## 3. The 4 Reference Model Types (From Edge Impulse Public Projects)

| Ref | Edge Impulse project | Type | Pipeline | Labels | Our template |
|---|---|---|---|---|---|
| R1 | [Bird sound classifier #16060](https://studio.edgeimpulse.com/public/16060/latest) | Audio, 16 kHz | MFCC → 1D-CNN (B6) | house_sparrow, rose_ringed_parakeet, noise | **T1** |
| R2 | [Driver drowsiness #510570](https://studio.edgeimpulse.com/public/510570/live) | Image | Object detection (B4) | eyes_open, eyes_closed, yawn, no_yawn | **T2** |
| R3 | [Pneumonia X-ray #348027](https://studio.edgeimpulse.com/public/348027/latest) | Image | Classifier (B1) | normal, pneumonia | **T3** (research/education only) |
| R4 | [Fall detection Transformer #208622](https://studio.edgeimpulse.com/public/208622/latest) | Accelerometer, 50 Hz | Raw → Transformer (B2) | ADL, FALL | **T4** |

---

## 4. Solutions by Application (Horizontal)

### 4.1 Computer Vision

| Solution | Data | Block | Device | Phase |
|---|---|---|---|---|
| Visual inspection: good vs defect | Camera | B1 → B5 | MCU · Edge | **P1** → P3 (anomaly) |
| Product/part recognition and sorting | Camera | B1 | MCU · Edge | **P1** |
| PPE / helmet compliance | Camera | B4 | Edge | **P1** |
| **Driver drowsiness (eyes closed, yawning)** | Cabin camera | B4 | Edge (RPi 5) | **P1** (T2) |
| Vehicle detection / type | Camera | B4 → B4+ | Edge · Cloud | **P1** → P2 |
| Static counting (tray, pallet, shelf items) | Camera | B4 | MCU · Edge | **P1** |
| Moving counting (conveyor, traffic) | Video | B4+ + B9 | Edge | P3 |
| OCR: meters, labels, number plates | Camera | B8 | Edge · Cloud | P3 |
| Anomaly without defect data | Camera | B5 | Edge | P3 |

### 4.2 Asset Tracking and Monitoring

| Solution | Data | Block | Device | Phase |
|---|---|---|---|---|
| Parcel handling events (drop, shock, tilt) | IMU | B2 | MCU | **P1** |
| Smart animal tracking (behaviour) | Collar IMU | B2 | MCU | **P1** |
| Remote environment monitoring (chainsaw, gunshot) | Audio | B6 | MCU | **P1** |
| Worker protection (PPE, man-down) | Camera / wearable IMU | B4 / B2 | Edge · MCU | **P1** |
| Cold-chain excursions | Temperature, humidity | B3 | MCU | P2 |
| Animal health anomaly | Body temperature, heart rate | B3 | MCU | P2 |

### 4.3 Human Interfaces

| Solution | Data | Block | Device | Phase |
|---|---|---|---|---|
| Gesture control | IMU | B2 | MCU | **P1** |
| Head-tilt / motion commands | Accelerometer + gyroscope | B2 | MCU | **P1** |
| Wake word / voice commands | Audio | B6 | MCU | **P1** |
| Acoustic events (baby cry, alarm, glass break) | Audio | B6 | MCU | **P1** |
| Exercise / rep classification | IMU → camera pose | B2 → B10 | MCU · Phone | **P1** → P3 |

### 4.4 Predictive Maintenance

| Solution | Data | Block | Device | Phase |
|---|---|---|---|---|
| Motor / pump / fan fault classification | Vibration | B2 | MCU | **P1** (T6) |
| Machine sound fault classification | Audio | B6 | MCU | **P1** |
| Water leak detection | Acoustic, pressure | B6 / B2 | MCU | **P1** |
| Assembly-line inspection | Camera | B1 / B4 | Edge | **P1** |
| Road inspection (potholes, cracks) | Camera | B1 / B4 | Edge · Phone | **P1** |
| Equipment anomaly (no fault data) | Vibration, current | B3 | MCU | P2 |
| Appliance failure (vibration anomaly) | Vibration, humidity, temperature | B3 | MCU | P2 |
| Crop disaster risk | Environmental sensors | B7 / B3 | MCU | P2 |

---

## 5. Solutions by Industry (Edge Impulse Parity)

### 5.1 Appliances

| Solution | Data | Block | Device | Phase |
|---|---|---|---|---|
| Leak detection | Humidity, pressure, acoustic | B2 / B6 | MCU | **P1** |
| Food recognition in fridge / oven | Camera | B1 / B4 | MCU · Edge | **P1** |
| Smart coffee machine (pod recognition) | Camera | B1 | MCU | **P1** |
| Washer load / water sensing | Humidity, load | B7 | MCU | P2 |
| Condenser failure prediction | Humidity, temperature | B3 | MCU | P2 |
| Appliance failure (vibration anomaly) | Vibration | B3 | MCU | P2 |

### 5.2 Buildings

| Solution | Data | Block | Device | Phase |
|---|---|---|---|---|
| Glass-break detection | Audio | B6 | MCU | **P1** |
| Occupancy / presence | PIR, mmWave radar | B2 | MCU | **P1** |
| Voice control for fixtures | Audio | B6 | MCU | **P1** |
| Smart security camera (person present, location) | Camera | B4 → B4+ | Edge | **P1** → P2 |
| HVAC predictive maintenance | Vibration, current | B2 → B3 | MCU | **P1** → P2 |
| Intelligent environmental control | Temperature, humidity, CO₂ | B7 | MCU | P2 |

### 5.3 Health (Wellness / Assistive / Research; Not Diagnostic)

| Solution | Data | Block | Device | Phase |
|---|---|---|---|---|
| **Fall / impact detection** | Wrist or belt IMU | B2 | MCU | **P1** (T4) |
| Baby monitor (cry detection) | Audio | B6 | MCU | **P1** |
| Heat-stress early warning | Skin temperature, heart rate | B2 → B3 | MCU | **P1** → P2 |
| **Chest X-ray classification demo** | X-ray images | B1 | Cloud · Edge | **P1** research/education (T3); clinical use needs regulatory approval |
| Cough / breathing sound research | Audio | B6 | MCU · Phone | **P1** research only |
| Sleep quality scoring | PPG, motion | B2 / B7 | MCU | P2 |

### 5.4 Industrial

| Solution | Data | Block | Device | Phase |
|---|---|---|---|---|
| Defect detection / QC | Camera | B1 → B5 | Edge · MCU | **P1** (T5) → P3 |
| Hard-hat / PPE detection | Camera | B4 | Edge | **P1** |
| Assembly verification (screws, lug nuts present) | Camera | B4 (count per part) | Edge | **P1** |
| Static inventory counting | Camera | B4 | Edge | **P1** |
| Predictive maintenance | Vibration, sound | B2 / B6 → B3 | MCU | **P1** → P2 |
| Automotive diagnostics (engine sound) | Audio | B6 | Edge | **P1** |
| Powerline / grid asset monitoring | Current, vibration | B3 | MCU | P2 |
| Moving inventory / logistics counting | Video | B9 | Edge | P3 |

### 5.5 Infrastructure / Smart Cities

| Solution | Data | Block | Device | Phase |
|---|---|---|---|---|
| Audio events (gunshot, crash, siren) | Audio | B6 | MCU · Edge | **P1** |
| Water utility leak detection | Acoustic, pressure | B6 → B3 | MCU | **P1** → P2 |
| Smart streetlights (presence, events) | Radar, audio | B2 / B6 | MCU | **P1** |
| Automated road inspection | Camera | B1 / B4 | Edge · Phone | **P1** |
| Grid power monitoring | Current, voltage | B3 | MCU | P2 |
| Traffic counting / classification | Video | B4+ + B9 | Edge | P3 |

### 5.6 Wearables

| Solution | Data | Block | Device | Phase |
|---|---|---|---|---|
| Wake word on earbuds/headsets | Audio | B6 | MCU | **P1** |
| Head-tilt volume control | IMU | B2 | MCU | **P1** |
| Gesture control | IMU | B2 | MCU | **P1** |
| Fitness coaching (exercise type, reps) | IMU | B2 | MCU | **P1** |
| Acoustic event alerts | Audio | B6 | MCU | **P1** |
| Performance analysis (HR, SpO₂) | PPG, IMU | B7 | MCU | P2 |

---

## 6. India-Focused Verticals

### 6.1 Agriculture

| Solution | Data | Block | Device | Phase |
|---|---|---|---|---|
| **Bird pest alert** (parakeets raiding orchards and fields) | Audio | B6 | MCU | **P1** (T1 base) |
| Crop leaf disease identification | Phone camera | B1 | Phone · Edge | **P1** |
| Fruit / vegetable / grain grading | Camera | B1 | Edge | **P1** |
| Pest counting on sticky traps | Camera | B4 | Edge · Phone | **P1** |
| Weed vs crop location | Camera | B4 | Edge | **P1** |
| Animal intrusion alert | Camera / audio / PIR | B4 / B6 / B2 | Edge · MCU | **P1** |
| Irrigation need prediction | Soil moisture, temperature, humidity | B7 | MCU | P2 |

### 6.2 Livestock and Dairy

| Solution | Data | Block | Device | Phase |
|---|---|---|---|---|
| Cattle activity and heat (estrus) detection | Collar IMU | B2 | MCU | **P1** |
| Rumination / feeding behaviour | Collar IMU / audio | B2 / B6 | MCU | **P1** |
| Animal health anomaly | Temperature, activity | B3 | MCU | P2 |
| Cattle identification (muzzle pattern) | Phone camera | Similarity matching | Phone · Cloud | P3 |

### 6.3 Manufacturing and Textiles (MSME)

| Solution | Data | Block | Device | Phase |
|---|---|---|---|---|
| Fabric / print defect detection | Line camera | B1 → B5 | Edge | **P1** → P3 |
| Machine idle / running / fault state | Vibration, current | B2 | MCU | **P1** |
| Packaging check (label present, seal OK) | Camera | B1 / B4 | Edge | **P1** |
| Gauge / meter reading | Camera | B8 | Edge · Cloud | P3 |

### 6.4 Retail and Logistics

| Solution | Data | Block | Device | Phase |
|---|---|---|---|---|
| Product recognition at counter | Camera | B1 | Edge · Phone | **P1** |
| Shelf gaps / out-of-stock | Camera | B4 | Edge | **P1** |
| Parcel rough-handling tag | IMU | B2 | MCU | **P1** |
| Footfall counting | Radar (presence) / video (tracking) | B2 / B9 | MCU · Edge | **P1** (radar) → P3 (video) |

### 6.5 Energy, Solar and Mobility

| Solution | Data | Block | Device | Phase |
|---|---|---|---|---|
| **Driver drowsiness alert** (fleets, buses, trucks) | Cabin camera | B4 | Edge (RPi 5 / Jetson) | **P1** (T2) |
| Solar panel soiling / crack detection | Drone or fixed camera | B1 | Edge · Cloud | **P1** |
| Transformer / inverter sound fault | Audio | B6 → B3 | MCU | **P1** → P2 |
| EV motor / battery anomaly | Vibration, temperature, current | B3 | MCU | P2 |

### 6.6 Education

| Solution | Data | Block | Device | Phase |
|---|---|---|---|---|
| College AI lab kit: public templates + phone-as-device + ESP32 | Image, audio, IMU | B1 / B2 / B4 / B6 | Phone · MCU | **P1** (notebooks) → P2 (phone-as-device + ESP32 SDK) |

---

## 7. Reference Templates (Proven in Notebooks in Month 1, Published as Public Projects in P2)

| # | Template | Block | Labels | Data plan | Recommended device |
|---|---|---|---|---|---|
| **T1** | **Bird Sound Classifier** (ref R1) | B6: MFCC → 1D-CNN | house_sparrow / rose_ringed_parakeet / noise | Own 16 kHz recordings + background noise; check licences on any public clips | MCU (ESP32-S3 / xG24) |
| **T2** | **Driver Drowsiness** (ref R2) | B4: centroid detection | eyes_open / eyes_closed / yawn / no_yawn | Webcam images, **collected with consent** | Edge (RPi 5 / Jetson) |
| **T3** | **Chest X-ray Classifier** (ref R3), research/education only | B1: transfer learning | normal / pneumonia | Public chest X-ray dataset (verify licence) | Cloud · Edge |
| **T4** | **Fall Detection** (ref R4) | B2: raw → tiny Transformer (1D-CNN fallback) | ADL / FALL | Phone accelerometer at 50 Hz; falls simulated on a mattress | MCU (wearable) |
| **T5** | **Product Defect QC** | B1: transfer learning | good / defect | 100+ phone photos per class | Edge → MCU |
| **T6** | **Machine Vibration Health** | B2: spectral → 1D-CNN | off / normal / imbalance / loose | IMU or phone on a fan or motor | MCU |

**Weekly additions after launch:** PPE helmet (B4), crop leaf disease (B1), pest counting (B4), keyword spotting (B6), gesture control (B2), cattle activity (B2).

Each template is a **public project** with a README, licence, pipeline, **our own measured metrics** (val/test, float32/int8), an on-device profile, sample data, and a Clone button.

---

## 8. Hardware / Target Matrix

| Target | P1 (files exported from notebooks) | P2+ |
|---|---|---|
| **ESP32-S3** (ESP32-P4 later) | TFLite int8 + C array + labels + DSP spec | ESP-IDF SDK (ESP-DSP MFCC), firmware builder, browser flashing, OTA |
| ESP32 / ESP32-C3 | Sensor/audio-light models, TFLite int8 | Same SDK (small models) |
| **Silicon Labs EFR32xG24** | TFLite int8 + C array + DSP spec | SLC component (with its ML accelerator), OTA |
| Generic Cortex-M4F / M7 (profiling) | Estimated latency/RAM/flash | Measured on real boards |
| Nordic nRF52 / nRF54 | TFLite int8 + C array | Zephyr module (CMSIS-DSP) |
| STM32 | TFLite int8 + C array | Zephyr / STM32 pack (P3) |
| **Renesas RA6 / RA8** | TFLite int8 + C array | FSP component (CMSIS-DSP); RZ/V DRP-AI (P3) |
| **Raspberry Pi 5** | ONNX + Python SDK (T2 target) | Linux agent (OTA + telemetry) |
| **NVIDIA Jetson** | ONNX + `build_engine.sh` + Python sample | Jetson agent, DeepStream config |
| Android / iOS | ONNX / TFLite | Mobile SDK, CoreML |
| **Browser / phone** | ONNX Web export (tested in a notebook) | Phone-as-device page, browser MFCC for audio |
| **Cloud API** | ONNX Runtime in notebooks; minimal FastAPI `/infer` in Week 4 | `POST /soc/v1/infer/{model}`, self-host Docker, autoscaling GPU |
| Neuromorphic (e.g., BrainChip Akida) | — | P3 evaluation |

---

## 9. How to Sell It

| Offer | For whom | Notes |
|---|---|---|
| **SutraAI Cloud** (SaaS) | Makers, startups, colleges | Free tier (limited train minutes and devices, public projects) → Pro (private projects, more devices, Cloud API calls) |
| **SutraAI On-Prem** | Factories, PSUs, defence, hospitals | Annual licence; offline mode; private by default |
| **Solution Kits** | MSMEs, fleets (drowsiness), agri / dairy cooperatives | Template + recommended hardware + 2–4 week proof-of-concept. This is the fastest early revenue. |
| **Edu Lab Kit** | Engineering colleges, ATL labs | Public templates + phone-as-device + ESP32 (P2) + course material |

---

## 10. Solution Page Template (Website)

```
# <Solution name>                         e.g., "Bird Pest Alert"
One-line promise                          "Hear parakeets before they strip your orchard."
Problem      — 2–3 lines in the customer's language
Sensors      — mic 16 kHz / camera / IMU 50 Hz
How it works — Collect (SOC API / phone) → Pipeline (MFCC → CNN) → Deploy (ESP32 / xG24 / RPi / Cloud)
Devices      — supported boards + latency/RAM/flash (estimated or measured badge)
Results      — our measured val/test accuracy (float32 and int8)
CTA          — [Open public project]  [Clone]  [Book a PoC]
```

---

## Sources

- Reference projects: [#16060](https://studio.edgeimpulse.com/public/16060/latest) · [#510570](https://studio.edgeimpulse.com/public/510570/live) · [#348027](https://studio.edgeimpulse.com/public/348027/latest) · [#208622](https://studio.edgeimpulse.com/public/208622/latest)
- Edge Impulse verticals: [Appliances](https://www.edgeimpulse.com/solutions/appliances) · [Buildings](https://www.edgeimpulse.com/solutions/buildings) · [Health](https://www.edgeimpulse.com/solutions/health) · [Industrial](https://www.edgeimpulse.com/industrial-new) · [Infrastructure](https://www.edgeimpulse.com/solutions/infrastructure) · [Wearables](https://www.edgeimpulse.com/solutions/wearables)
- Edge Impulse applications: [Computer Vision](https://www.edgeimpulse.com/computer-vision) · [Asset tracking](https://www.edgeimpulse.com/solutions/asset-tracking-monitoring) · [Human interfaces](https://www.edgeimpulse.com/solutions/human-interfaces) · [Predictive maintenance](https://www.edgeimpulse.com/solutions/predictive-maintenance)
- [Edge Impulse API reference](https://docs.edgeimpulse.com/reference) · [Data acquisition format](https://docs.edgeimpulse.com/reference/data-acquisition-format)