What shipped, release by release

Added

• Camera("realsense") — any Intel RealSense (D4xx/L5xx) now drives the full SpatialAI stack: models, filters, and the streaming API run on real depth frames. pip install "spatialai[realsense]"
• Field of view read from stream intrinsics at runtime, keeping person/pet size classification physically correct on any camera
• spatialai dump CLI (0.4.1) — raw UART capture with frame-marker/checksum analysis for Spatial Home bring-up

Added

• spatialai.opn6001 — Spatial Vision (OPN6001/OPN6002) toolkit: MIPI RAW12 frame decode (720×320 raw → 240×320 depth/IR/status/BK planes, 0.25 mm LSB), with both vendor decode variants selectable pending test-pattern validation
• Vendor-recommended post-processing chain: IR filter, status-based confidence filter with over-exposure dilation, R=2 median filter
• 32-bit SPI transaction builders (burst, dummy-word rules) and the documented boot/ready register sequence
• 9 new tests including encode/decode round-trips (32 total)

Notes

• Full V4L2 capture backend needs a platform kernel driver (vendor RK3566 reference available) — planned once modules are on the bench

Added

• Spatial Home (DMOS5030A/5031A) UART backend — full Dragonfly protocol v1.4.4: framing, ISP init, ROI/binning, stream control, depth-frame parsing. Camera("spatial-home", port="/dev/ttyUSB0")
• Spatial Mini (DM0301) native I²C driver at address 0x41 — continuous ranging, interrupt-driven data-ready, millimeter results with confidence
• pip install "spatialai[hardware]" extra (smbus2 + pyserial); per-device extras remain
• 10 new protocol-level tests (23 total): golden-byte command encoding, split/corrupt frame parsing, fake-bus register sequences

Fixed

• DM0301 backend no longer assumes VL53L4CD protocol compatibility — the datasheet confirms pin-compatibility only (different I²C address and register map)

Shipped

• Fanless edge AI box — octa-core Arm CPU + NPU (up to 43.2 TOPS INT4), ingests up to 32 × 1080p30 camera streams
• Runs SpatialAI SDK v0.2 and both shipping models out of the box; NPU-accelerated marketplace models will arrive via OTA
• $599, available in the store now — card, Apple Pay & Google Pay

Added

• cam.load_model(name) — loaded models run on every stream frame (frame.models) and on demand via cam.infer()
• presence-core — debounced presence & approach detection with adaptive background tracking; works on every device including 1D sensors
• occupancy-home-v1 — room occupancy from depth CV: background estimation, depth-aware connected components (splits overlapping people), physical-height person/pet classification. 90% agreement with simulator ground truth
• Simulator now renders physically-scaled synthetic depth frames (FOV-correct person size), so depth models are testable without hardware
• Announced-but-unshipped models raise ModelNotAvailableYet with waitlist instructions

Changed

• Model Marketplace page now shows per-model status badges — shipping models are free and marked "Available now"
• Test suite grew to 13 tests

Added

• pip install spatialai — package registered and published on PyPI
• Camera / Sensor API: detect_people(), stream(), distance_m(), context-manager support
• Realistic simulator for all four products — the five-line quickstart runs on any machine with no hardware
• Real I²C backend for Spatial Mini (DM0301, VL53L4CD-compatible) via pip install "spatialai[dm0301]" on Raspberry Pi and Linux SBCs
• CLI: spatialai demo, scan, devices
• MIT license, pytest suite

Shipped

• Full product site: four ToF products with real supplier specs, store with cart & order checkout, developer docs, blog, support
• Cloudflare Workers deployment with staged 404 handling and auto-deploy from main
• AI-generated product imagery matched to actual hardware form factors