Your 30-day NVIDIA Jetson AI plan
From a fresh Jetson Orin NX to four working edge AI models, an on-device LLM advisor and a web demo — one day at a time, with a free course for each day and NVIDIA's ground rules built in.
- 30
- days, about 6 h each
- 4
- model types: image, sensor, audio, detection
- 15
- free NVIDIA DLI and partner courses
- $0
- course cost on the free path
Daily rhythm
90 min · Learn
- The free course listed on the day card
- Take three notes in your own words
4 h · Build
- Open the step page and run the notebook
- Only tick "Done when" after you see it pass
30 min · Log
- Write the day in BUILDLOG.md
- Numbers, what blocked you, top 3 for tomorrow
NVIDIA guidelines for the 30 days
Eight rules that keep the Jetson stable and your numbers trustworthy. They come from NVIDIA's JetPack, TensorRT and Jetson AI Lab guidance as applied to your Orin NX 16GB.
Stay on JetPack 6.2.1
Don't upgrade JetPack, L4T or CUDA during the 21-day notebook phase. Every wheel and engine on these pages is built for 6.2.1 / CUDA 12.6.
Install PyTorch from the Jetson index
Use the jp6/cu126 wheels and the constraints file from Step 02. A PyPI wheel is CPU-only and silently breaks the GPU.
Train in MAXN SUPER
Run sudo nvpmodel -m 0 before training and keep the fan working. Watch temperature and power in jtop.
Budget the 16 GB
CPU and GPU share memory. One big model next to training at most; keep 16 GB swap on and stop llama-server when you train.
TensorRT engines are per device
Build .engine files on the Jetson that will run them, with the same TensorRT version. Ship ONNX, rebuild engines on target.
Validate every export
Check ONNX parity, int8 accuracy drop and latency before you trust an exported model. Keep golden vectors.
Pick models that fit the edge
On 16 GB use Nemotron 3 Nano 4B (Q4) and Cosmos Reason 2 2B; leave 30B-class models for AGX Orin or a cloud GPU.
Respect licences and consent
Self-collected data, consent for faces, research-only for X-ray. Check each NVIDIA model's licence before shipping it.
Day by day
Tick a day when its check passes. Progress is saved in this browser only.
Week 1 · Jetson, CUDA and your first model
26 Oct – 1 Nov · Phase A- 1Mon
Set up the Jetson
Day 1 · Mon 26 Oct 2026 · Step 01Free course (90 min): Getting Started with AI on Jetson Nano
- Done when: jtop is healthy and your JetPack / CUDA versions are in BUILDLOG
- 2Tue
CUDA, Python and JupyterLab · NB00, NB01
Day 2 · Tue 27 Oct 2026 · Step 02Free course (90 min): An Even Easier Introduction to CUDA
- Done when: torch.cuda.is_available() is True and the GPU beats the CPU on a matmul
- 3Wed
ML basics · NB02
Day 3 · Wed 28 Oct 2026 · Step 03Free course (90 min): Building A Brain in 10 Minutes! · learnpytorch.io (Zero to Mastery, free book)
- Done when: You can explain overfitting from your own loss plot
- 4Thu
Image dataset + first transfer-learning run · NB10, NB11
Day 4 · Thu 29 Oct 2026 · Step 04Free course (90 min): learnpytorch.io (Zero to Mastery, free book)
- Done when: metrics.json saved with a test accuracy
- 5Fri
Augmentation, confusion matrix, seeds · NB11
Day 5 · Fri 30 Oct 2026 · Step 04Free course (90 min): learnpytorch.io (Zero to Mastery, free book)
- Done when: Test accuracy ≥ 85% and Restart & Run All reproduces it within ±1%
- 6Sat
Export + validate: ONNX, TensorRT, TFLite int8 · NB12
Day 6 · Sat 31 Oct 2026 · Step 05Free course (90 min): TensorRT quick start guide (NVIDIA docs)
- Done when: int8 drop < 3 points; TensorRT latency recorded
- 7Sun
Week 1 review
Day 7 · Sun 1 Nov 2026 · Step 05Course slot: catch up, clean notebooks, update BUILDLOG.md
- Done when: Image path works from photo to .tflite + .engine, pushed to git
Week 2 · Sensor and audio pipelines
2 – 8 Nov · Phase A- 8Mon
Sensor data with phyphox · NB20
Day 8 · Mon 2 Nov 2026 · Step 06Free course (90 min): learnpytorch.io (Zero to Mastery, free book)
- Done when: Clear per-class plots; no recording in both train and test
- 9Tue
1D-CNN on raw windows · NB21
Day 9 · Tue 3 Nov 2026 · Step 06Free course (90 min): learnpytorch.io (Zero to Mastery, free book)
- Done when: Test accuracy and confusion matrix recorded
- 10Wed
Tiny Transformer vs 1D-CNN · NB21
Day 10 · Wed 4 Nov 2026 · Step 06Free course (90 min): Generative AI Explained
- Done when: Comparison table; official sensor model chosen
- 11Thu
Audio data + MFCC from scratch · NB30, NB31
Day 11 · Thu 5 Nov 2026 · Step 07Free course (90 min): The Sound of AI (YouTube)
- Done when: MFCC shape (50, 13) per second; golden vectors saved
- 12Fri
Audio training + int8 · NB32
Day 12 · Fri 6 Nov 2026 · Step 07Free course (90 min): The Sound of AI (YouTube)
- Done when: Test accuracy ≥ 80% on your clips; int8 drop < 3 points
- 13Sat
pipeline.yaml + NB90 v0
Day 13 · Sat 7 Nov 2026 · Step 08Course slot: catch up, clean notebooks, update BUILDLOG.md
- Done when: image, sensor and audio configs run from a clean kernel
- 14Sun
Week 2 review
Day 14 · Sun 8 Nov 2026 · Step 08Course slot: catch up, clean notebooks, update BUILDLOG.md
- Done when: Three of the four model types work
Week 3 · Detection, profiling, NVIDIA models, gate
9 – 15 Nov · Phase A- 15Mon
Detection data in Label Studio · NB40
Day 15 · Mon 9 Nov 2026 · Step 09Free course (90 min): Building Video AI Applications at the Edge on Jetson Nano
- Done when: ≥ 200 labelled images; ≥ 100 boxes per class
- 16Tue
Centroid detector · NB41
Day 16 · Tue 10 Nov 2026 · Step 09Free course (90 min): Building Video AI Applications at the Edge on Jetson Nano
- Done when: F1 ≥ 0.75 per class; overlays look right
- 17Wed
Detection export + profiler · NB42, NB50
Day 17 · Wed 11 Nov 2026 · Step 10Free course (90 min): TensorRT quick start guide (NVIDIA docs)
- Done when: int8 F1 drop < 0.05; reports/profile.md saved
- 18Thu
Device simulator, no API · NB51
Day 18 · Thu 12 Nov 2026 · Step 10Course slot: catch up, clean notebooks, update BUILDLOG.md
- Done when: v2 beats v1; a corrupted file is rejected
- 19Fri
Nemotron 3 Nano 4B advisor + RAG · NB60, NB61
Day 19 · Fri 13 Nov 2026 · Step 11Free course (90 min): Augmenting LLMs using Retrieval Augmented Generation · Building RAG Agents for LLMs
- Done when: Advice uses only real numbers; RAG hit@4 ≥ 8/10
- 20Sat
NVIDIA vision helpers: Cosmos Reason 2, NanoOWL · NB62
Day 20 · Sat 14 Nov 2026 · Step 12Free course (90 min): Jetson AI Lab tutorials
- Done when: Time-saved table and an auto-labelling decision
- 21Sun
ADOPTION GATE · NB90 final
Day 21 · Sun 15 Nov 2026 · Step 13Course slot: catch up, clean notebooks, update BUILDLOG.md
- Done when: All 4 configs run end to end from a clean kernel
Week 4 · Notebooks → package → API → UI
16 – 22 Nov · Phase B- 22Mon
sutra_core → ml_core package
Day 22 · Mon 16 Nov 2026 · Step 14Free course (90 min): FastAPI official tutorial
- Done when: CLI reproduces the NB90 numbers; pytest passes
- 23Tue
FastAPI skeleton
Day 23 · Tue 17 Nov 2026 · Step 15Free course (90 min): FastAPI official tutorial
- Done when: All endpoints work with curl on the Jetson
- 24Wed
Job worker + SSE progress
Day 24 · Wed 18 Nov 2026 · Step 15Free course (90 min): FastAPI official tutorial
- Done when: Live progress with curl -N; cancel works
- 25Thu
Next.js minimal slice
Day 25 · Thu 19 Nov 2026 · Step 16Free course (90 min): Next.js Learn (official course)
- Done when: One template works end to end in the browser
- 26Fri
Export menu, profiler table, advisor text
Day 26 · Fri 20 Nov 2026 · Step 16Free course (90 min): Next.js Learn (official course)
- Done when: All 4 templates downloadable from the UI
- 27Sat
End-to-end demo
Day 27 · Sat 21 Nov 2026 · Step 16Course slot: catch up, clean notebooks, update BUILDLOG.md
- Done when: 5-minute demo video recorded
- 28Sun
Review and bug fixes
Day 28 · Sun 22 Nov 2026 · Step 16Course slot: catch up, clean notebooks, update BUILDLOG.md
- Done when: No open blocker in BUILDLOG
Days 29–30 · Write it down, plan Month 2
23 – 24 Nov · Phase B- 29Mon
Write it down: README, PIPELINE.md, API reference
Day 29 · Mon 23 Nov 2026 · Step 16Free course (90 min): NVIDIA Agentic AI Learning Path
- Done when: Docs let someone else rerun your numbers
- 30Tue
Plan Month 2 + show the demo
Day 30 · Tue 24 Nov 2026 · Step 16Free course (90 min): Hugging Face MCP Course
- Done when: Feedback from 1–2 people and a Month 2 kickoff list
Month 1 targets
| Metric | Target |
|---|---|
| Image (T5) test accuracy | ≥ 85% |
| Audio (T1) test accuracy | ≥ 80% |
| Sensor (T4/T6) test accuracy | ≥ 85% |
| Detection (T2) centroid F1 | ≥ 0.75 per class |
| int8 vs float32 drop | < 3 points (detection < 0.05 F1) |
| Web slice | Upload → train → download for all 4 templates |
The full schedule with notes is in SutraAI_30_Day_Plan.md.