An ESP32 with an AMG8833 “Grid-EYE” 8×8 infrared sensor pushes frames over TLS to a VPS. The VPS keeps only the raw 8×8 temperature grid, renders every picture server-side, and streams it to any browser at up to 30 fps. A phone camera can overlay the heat map on the live view to hunt hot components; a vision AI names the component and finds a cheap price.
ESP32 + AMG8833
+-- TLS POST (128 B/frame) --> VPS :8090
+-- SSE (2 PNGs) --> any browser
+-- camera snap --> gallery + AI
2 · What is stored
Raw 8×8 frames only — last 30, in RAM (_buf). The thermal stream is never written to disk.
Nothing else of the live feed is persisted. All colour, stacking, stretch, contour and outline happen fresh each frame.
Gallery — the composited phone JPEGs, their metadata and AI result, under /opt/thermal/gallery/ (index.json + <id>.jpg), capped at 300.
3 · The processing pipeline
Applies to the Processed pane and, when astro is on, to the camera overlay:
Sigma-clipped rolling mean stack (last N frames) → denoise.
Bicubic upscale 8→256, minus a background estimate.
asinh stretch (the Stretch slider) with a guard so k=0 is linear.
Additive local contrast (CLAHE-style).
Ordered dither (8×8 Bayer) to kill 32-level banding, then quantise to 32 levels.
Contour lines drawn as 20%-black twins of each ramp colour (pre-composited, bandwidth-friendly).
Otsu outline of the hot shape, drawn in green ([0,255,120]), always on.
Pack to an indexed palette PNG (65-colour palette, Up filter, zlib 6).
The colour window, the Otsu threshold and the overlay all read the same normalised field, so they never disagree.
4 · Controls
Control
What it does
Default
Start astro processing
Turns the whole Processed pipeline on/off. Off = raw, only scaled.
off
Render — Sharp / Contour / Blocks
Upscale kernel + contour overlay. Sharp is the cleanest.
Sharp
🔥 Hotspot alarm
Any cell at/above this raises the red alarm banner + beep and marks snapped photos as hot.
40.0 °C
☀ Stretch
asinh strength, 0 (linear) → 8 (max). Higher = more faint detail, until it saturates.
2.0
🔬 Local contrast
Additive CLAHE gain, 0–3. Brings out small differences across the frame.
1.0
✂ Sigma clip
Stack outlier rejection strength (0.3–8σ). Low = smoother but can eat faint real detail; high = keeps more noise. 0.15 °C floor always applies.
3.0
▤ Stack
Frames averaged, 1–30. More = smoother but slower to react.
12
🌈 Colour range
AUTO fits min→max every frame; manual sets the low/high °C that map to blue/red. Never inverts (≥0.5 °C span).
AUTO 30–40
Every control is live-synced: change it on one phone and every other viewer updates instantly (badge + toast).
5 · Camera & thermal overlay
Tap 📷 Camera & log. The rear camera becomes the viewfinder; the heat map is warped onto it.
🎯 Align is the 5-second re-nudge: hold the sensor on something hot, drag the heat over it, tweak Size / Rotate / Mirror, then tap Align again. Saved in the phone's localStorage, so it survives reloads. Reset restores defaults.
When astro is on the overlay colours from the processed field; when off it uses the raw 8×8 through the colour window.
The hottest cell gets a box + its °C; red if it crosses the hotspot alarm.
6 · Gallery & access
The Hotspot log sits on the main page and is public — anyone with this link can see every photo and its AI result. Only capturing and deleting need the passcode. Tap any photo to open it full-size.
Action
Public (no passcode)
With passcode
Open the camera panel / live overlay
✅
✅
View the gallery / a full photo
✅
✅
Snap a new photo
❌ 401
✅
Delete a photo
❌ 401
✅
The passcode is checked server-side and sets an HTTP-only cookie (30 days). It is not in the page source. Old photos auto-delete once the gallery passes 300.
7 · AI component ID & price
On every snap the VPS runs a background job (key never touches the phone):
snap -> vision model : identify the component under the hotspot box
-> price model : live web search for the cheapest seller
-> stored : shown under the photo
Returns component type, any read part number, package, a one-line description, a confidence %, and the cheapest price + seller + link — Amazon.in or Flipkart only, in ₹ (INR).
If nothing is identified, it stops (no made-up prices).
Prices are an approximate estimate from live web search (marked ≈), restricted to a real Amazon.in / Flipkart product page (single unit, no bundles). Tap search↗ to check the exact part yourself.
Roughly $0.01 per snap. The card shows “identifying part…” until done (~10 s).