Skip to content
IEMI Labs
PumpOne

Our own circuit boards, in the lift station's control panel

Boards we design for the smart pump panel, with every chip soldered on them: the RV1126B processor and its memory, the station camera chip, the radar chip, the heat chip, LTE-M with a soldered eSIM, the station ADC and its isolated inputs, the e-paper and call light, the security chip, and on its own board the supply from the panel’s control circuit and a backup cell. They stand on the right side wall and inside the door of the same 24 × 20 × 8 in stainless panel, beside its own breakers and contactors; two latched leads join them. No plugs, ribbons or adapters inside.

Stage 1 of 6 done: shape and placementNot yet ready to manufactureUpdated 4 November 2026 · rev A-0

The board in 3D

Every part at its real size, in its real place. Turn it, zoom in, switch between the boards in the panel and laid flat, and click any part to see what it is. The chip bodies are simple blocks until each part's own 3D model is added.

Loading the 3D board…

Drag to turn · scroll to zoom · right-drag to move · click a partDrag to turn · tap a part

Where everything is in the panel

Station camera and radarright side wall

A bare Sony IMX415 camera chip under a 102° M12 lens, 15° down through a 20 mm window, and a TI IWRL6432AOP 60 GHz radar chip behind the PC window beside it, each on a tab: the wet-well hatch, both riser gauges, the gate, and a person at the hatch in the dark.

Processor and radiosright side wall, top

The RV1126B with its memory on the main board, a heat block to the stainless wall; LTE-M with a soldered eSIM and BLE for the operator’s phone, both through one IP67 puck antenna in the top wall, because the closed stainless box stops the signal.

Station inputsright side wall, bottom

Spring-clamp terminals beside the pump leads for both current clamps, the 4–20 mA level loop (read through an isolated amplifier), the high float and both pump run calls (through optocouplers), and a relay contact to the panel’s own alarm light. PumpOne never switches a pump.

Heat cameradoor board

A Melexis MLX90640 heat chip on the inside of the door, looking straight back at both breakers, both contactors and overloads and the terminal strip.

E-paper, light and tonedoor board

The 2.13 in e-paper in its door window (level, pumps, power, last call), 8 call-light LEDs behind the slot and a piezo behind the tone holes; a 3D Hall chip under a magnet logs every door opening.

Power and backup cellpower board

The panel’s 120 V control circuit into a class II 5 V supply, and one soldered 18650 cell with a thermistor that keeps the power-lost call and level reports going for about 19 h.

The panel boards laid flat, both sides, every part at its real size

Laid flat: the door board on the left; the main board with its two sensor tabs on the right and the power board below it. Tap the drawing to open it full size.

Design review: checked before we order

Every board-level number PumpOne's calls depend on, checked against its target with the source of each number, using the rules every one of our boards shares (products/_shared/board-rules.mjs). Fixed means a lesson changed this board; to prove means it passes on paper and is measured on rev A before boards are ordered.

No check fails18 pass · 5 fixed by the review · 10 to prove before we order · 0 failDesign review of 4 November 2026
AreaCheckOur boardTargetStatus
CameraStation camera: a 100 mm discharge gauge dial on the riser at 1.99 m
M12 lens, 102° across: about 37.6 px per degree across 3840 px. The needle reader needs about 40 px across the dial (est.); it reads only in daylight, glare or rain on the dial means no reading.
108.4 px≥ 40 pxpasses
CameraStation camera: frame rate the calls need
A standard mode of the IMX415.
5 fps at 3840 × 2160chip: 30 fps at 3840 × 2160passes
CameraStation camera: data rate vs its 4-lane link
88 % spare on the link.
0.41 Gbps (10-bit)≤ 80 % of 3.564 Gbpspasses
CameraStation camera: a person (0.5 m wide) at the station gate at 3.35 m
M12 lens, 102° across: about 37.6 px per degree across 3840 px. The person detector works from about 20 px across a person (est.); no faces are kept.
321.3 px≥ 20 pxpasses
CameraStation camera: frame rate the calls need
A standard mode of the IMX415.
5 fps at 3840 × 2160chip: 30 fps at 3840 × 2160passes
CameraStation camera: data rate vs its 4-lane link
88 % spare on the link.
0.41 Gbps (10-bit)≤ 80 % of 3.564 Gbpspasses
CameraPanel heat camera: one contactor terminal block (25 mm) on the back plate at 0.29 m
32 × 24 pixels across 75°: 0.4 px per degree. The call compares a part with the matching part of the other pump while it runs, averaged over many frames; the chip is turned so its 75° runs across the back plate and its 110° from the breakers down to the terminal strip.
2.1 px≥ 1 pxpasses
CameraImage processor load, all cameras
2 cameras into one image processor.
83 Mpx/s≤ 360 Mpx/s (RV1126B ISP)passes
CameraAI engine time per frame
One person model (int8) on the camera, woken by the radar; the needle reader runs on a still at each pump start. Level, currents and heat are read by the CPU. Measured on the dev kit.
3 TOPSthe gauge reader on each pump run and the person check within 5 s of the radar seeing someoneto prove on the dev kit
PowerSupply
About 3.2 W watching (RV1126B 1.5 W, camera 0.3 W, radar 0.1 W, heat chip 0.1 W, ADC and inputs 0.1 W, LTE-M idle 0.2 W, Wi-Fi/BT 0.2 W, e-paper and light 0.2 W) and 7.5 W at a call with the cell charging at 1 A (est.): 38 % of the supply.
Mean Well IRM-20-5 PCB-mount 5 V 20 W supply (class II, listed) on the panel's own 120 V control circuit, on the power board≤ 70 % of the supply (14 W)passes
PowerPower cut
Pulling the control circuit is part of the rev A tests at a station; a thermistor holds charging outside 0 to 45 °C (lesson: charge).
one soldered 18650 cell (Samsung INR18650-35E) on the power board: when the station loses power the RV1126B sleeps, the BG95-M3 sends the power-lost call at once and a level and pump report every 5 min, about 0.35 W, about 19 h on an aged, cold cell (est.)the power-lost call and level reports for 12 hto prove on the dev kit
HeatRV1126B temperature, a summer lift station (40 °C air) with the panel under its sun shield
The shaded stainless side wall on a hot day; counted at 55 °C. 1.5 W through a 20 × 20 aluminium block and gap pads to the 1.9 mm stainless right side wall (8 K/W, est.).
67 °C≤ 85 °C (rated)passes
HeatCamera and radar chips away from the processor
A warm sensor adds noise.
station camera 133 mm, radar 123 mm≥ 30 mmpasses
ShockShock and vibration
Door slams and vandal kicks on the door and the windows; about 50 g at the boards on their stand-offs (est.). MEMS clocks are fitted anyway, and the card socket is latched.
50 g (est.)standard partspasses
SoftwareStarts our software from the microSD card, no setup
Owner requirement: power it from its own supply and our software runs from the card. Rev A and sale boards boot from the card; the eMMC holds no system copy (it ships empty, so the card always starts first: Rockchip tries the eMMC before the card) and keeps the station log and the call log. Updates write to the card with two system copies, only when both pumps are idle and the level is low, never during a high-level call.
microSD socket with a hinged, latched lid (Amphenol 101-00660 class), fitted on every board: our software starts from the cardcard socket fitted on every board, latchedpasses
SoftwareCard power switch and protection
A card that hangs must be restarted by cutting its power, not the whole board; the card slot is the one place a person touches the electronics (lesson: sdstart). Change: U90, U91 added (4 Nov 2026, CricketOne card-start review).
1 power switch and 1 ESD array for 1 card socketevery socket: switched card supply, ESD on all 6 linesfixed by the review
SoftwareThe card wins the start
Rockchip's start order is SPI flash, eMMC, card, USB: a written eMMC starts before the card. Proved on the dev kit: power on with a card in and the eMMC empty (lesson: sdstart). Change: eMMC ships empty and the loader reads the card first (4 Nov 2026; it used to say the eMMC kept a copy, which would have started instead of the card).
eMMC shipped empty; our loader reads the card firstthe processor starts from the card every timefixed by the review
SoftwareCard grade
Consumer cards wear out and fail hot; the image is read-only with a separate data partition. Proved: 1,000 power cuts during writes on rev A (lesson: sdstart).
industrial pSLC microSD (Kingston Industrial or Swissbit S-56u), −40 to 85 °Crated for the heat, cold and writesto prove on the dev kit
SoftwareClip buffer for an unsure call
The rest of the memory holds Linux, the program and the AI model (lesson: clip).
19.8 s of 8-bit 5 fps video in 40 % of the 2 GB memory≥ 3 spasses
RadioWi-Fi/BT module clock
The AP6256 runs from its own 37.4 MHz reference clock; a wrong or missing clock means no Wi-Fi or Bluetooth at all (lesson: wificlock). Change: the AP6256 had no clock on the board; Y3, a 37.4 MHz MEMS oscillator, added next to it (4 Nov 2026, from the CricketOne test-board circuit).
37.4 MHz MEMS oscillator (Y3)37.4 MHz next to the AP6256fixed by the review
RadioLTE-M from inside the stainless panel
A closed 304 stainless box blocks LTE-M, so the antenna sits outside it; the steel sun shield 40 mm above the puck is measured on rev A. Antenna outside the metal: the 2-in-1 IP67 puck antenna through the top wall, under the station's sun shield (lesson: enclosure). Measured on rev A (lesson: antenna). Change: CAD rev T: the bench LTE-M board and its antenna sat inside the closed stainless panel; a 2-in-1 IP67 puck now goes through the top wall under the sun shield, its two coax leads soldered to the main board (4 Nov 2026).
BG95-M3 with a soldered eSIM; its antenna is the LTE element of the puck through the top wallcertified module; callouts reach the on-call operator from a rural stationto prove on the dev kit
RadioBLE for the operator's phone at the station
Only pairing and the "operator here" mark that clears an entry call; everything else goes over LTE-M. Antenna outside the metal: the same puck antenna through the top wall (lesson: enclosure). Change: CAD rev T: the bench LTE-M board and its antenna sat inside the closed stainless panel; a 2-in-1 IP67 puck now goes through the top wall under the sun shield, its two coax leads soldered to the main board (4 Nov 2026).
AP6256 on the main board, its antenna the 2.4 GHz element of the same puckpairing from the hatch, 4 m away (est.)fixed by the review
LayoutFast lines on a solid ground plane
The layout rules go to the maker in the quote; the routed board's design rule check (0 copper errors, every keep-out kept) and the maker's impedance coupon prove them (lesson: planes).
RV1126B memory; 1 camera MIPI link (100 Ω pairs); USB (90 Ω); antenna feed (50 Ω, keep-out on every layer); card lines (length-matched, no vias under the socket)ground on inner layer 1, supply on inner layer 2, no vias in pads, impedance lines routed firstto prove on the dev kit
MoistureConformal coating
Hose-down, condensation in a cold panel and hydrogen sulphide from the wet well corrode bare copper; all three boards and both tabs get the same coat. Asked for in the quote (lesson: moisture).
coatedcoated, sensors, sockets and antennas maskedpasses
FitBoard inside the product CAD, every part checked in 3D
The board is checked on its own; the rev T CAD run places it in the product and checks every part in 3D (lesson: fit3d).
not yet in the product CADnothing touchesto prove on the dev kit
CameraEvery station call in view
Every inView target from the CAD (the hatch centre and both corners, both riser gauges, the gate; a person at the hatch and at the riser; both breakers, both contactors and the terminal strip) from each lens or chip's real place on our boards.
Station camera: worst 33.9° / 10.4°; Hatch radar: worst 12° / 3.3°; heat camera: worst 16.5° / 44.6°inside each sensor's viewpasses
CameraStation camera window
The round 14 mm window was sized for the bench camera module; our M12 lens rim and the gate call, 34° to the side, need a 20 mm square (lesson: window). The radar window was already big enough. Change: CAD rev T: camera window Ø14 → 20 × 20 (4 Nov 2026).
20 × 20 (margin 2.3 mm); radar window 50 × 20 (margin 6.4 mm); steepest line of sight 38°each view inside its windowfixed by the review
FitSensor tabs on their brackets
Each tab hangs off the main board and is held on its sensor's axis by a printed bracket; the brackets go in the rev T CAD and are checked in 3D there (lesson: fit3d).
Station camera: 0.3 mm inside the wall, hinge 13.6 mm; Hatch radar: 0.3 mm inside the wall, hinge 19.6 mmchip or holder ≥ 0.3 mm inside the wall, hinge within the 40 mm flexto prove on the dev kit
FitDoor parts in their openings
The door board sits 4 mm inside the door skin, so the e-paper face, the LEDs and the piezo are just behind their openings.
e-paper active area inside its 26 × 52 window, 8 call-light LEDs inside the 46 × 6 slot, the piezo over the tone holeseach inside its openingpasses
SafetyControl circuit kept on the power board
The panel builder wires our power board like any listed component inside a UL 508A panel; creepage between the 120 V terminals and the low-voltage inputs is checked in the layout.
the panel's 120 V control circuit into a spring-clamp terminal, fuse, varistor, IRM-20-5 (class II); the pump run-call inputs through optocouplers; only 5 V, cell sense and SPI on lead L1class II creepage on the power board; no 120 V on any other boardto prove on the dev kit
SafetyThe station's own controls stay in charge
The level controller and alternator keep running the pumps; PumpOne only reads and calls. The 100 Ω shunt is a passive part in the loop with nothing active in series, so the loop keeps working with the board unpowered; taking the board out needs a jumper across the loop terminals, written on the panel label.
level loop read through an isolated amplifier across a shunt, run calls through optocouplers, current clamps with no contact; one relay contact to the panel's alarm lightPumpOne never switches a pump; a board fault leaves the level controller workingpasses
SecurityDoor tamper sensor
The bench used a plunger switch at the door rim; the 3D Hall chip reads the magnet's field to a fraction of a millitesla, and the switch point is set on rev A.
TMAG5273 3D Hall chip on the door board, 21.4 mm from a Ø10 × 5 magnet on a printed boss on the body top walldoor opened = logged and signedto prove on the dev kit
FitLeads and the antenna coax
L2 crosses the door hinge in a strain-relieved loop with high-flex wire; the ADC's SPI on L1 runs at 1 MHz with series resistors.
L1 about 0.12 m (0.008 V drop), L2 about 0.59 m in the hinge loop (0.015 V drop); coax about 0.15 mleads ≤ 0.15 V drop; L2 rated for 10,000 door openings; coax ≤ 0.25 mto prove on the dev kit

Before we order the first boards, we prove these on the RV1126B development kit or in the rev A tests:

  1. AI engine time per frame
  2. Power cut
  3. Card grade
  4. LTE-M from inside the stainless panel
  5. Fast lines on a solid ground plane
  6. Board inside the product CAD, every part checked in 3D
  7. Sensor tabs on their brackets
  8. Control circuit kept on the power board
  9. Door tamper sensor
  10. Leads and the antenna coax

Where the design stands

  1. 1Shape and placementdone

    Main board with two sensor tabs, power board and door board in the panel’s shape, all 62 parts placed at their real size, every view checked through its window in 3D and KiCad design rule check (0 violations).

  2. 2Schematicnext

    Every wire between the chips, from the V1 core, C1 camera, station input and power circuits.

  3. 3Footprintsnext

    The exact solder pads under each chip.

  4. 4Routinglater

    Copper tracks on all layers, the camera lanes over the tab hinge.

  5. 5Files for the makerlater

    Gerbers with copper, drill, mask, paste and the assembly parts list. These go to the board maker.

  6. 6Rev A boardsMar 2027

    Made, assembled and tested on our fixture.

Design files (rev A-0)

These are the stage 1 files: the board's shape and where every part goes. A board maker needs the stage 5 files to build it.

Who can make it

The panel set is an 8-layer main board with two tabs on flex hinges, a 4-layer power board and a large 2-layer door board. It needs a maker who builds rigid-flex tabs and fine-pitch ball-grid chips. We plan three quotes once the stage 5 files are ready: two from China for price, and one from the US for speed.

MakerWhereRigid-flexHDI / layersAssemblyFit for us
PCBWayfirst quoteShenzhen, Chinayes, 6 to 16 layersHDI up to 32 layers, 0.35 mm BGA pitchyes, turnkey, incl. BGA and X-rayBest fit for rev A: the rigid-flex board and its assembly in one order
NextPCBsecond quoteShenzhen, ChinayesHDI, 6 to 32 layersyes (PCBA)Second quote, to compare price and lead time
ALLPCBChina (a network of partner factories)yes, through partner factories1 to 30 layersyesThird quote; check which factory builds it
JLCPCBShenzhen, Chinano (flex only: 1, 2 and 4 layers, up to 490 mm)HDI added Sep 2026yes, incl. BGANot for the one-piece board today. Good for the test fixture, simple boards, and the fallback of separate rigid boards joined by soldered flex
Sierra Circuits (ProtoExpress)US quoteSunnyvale, California, USAyesyesyes, in-house, incl. BGA; turnkey in about 5 daysUS option: fastest turn and engineering help, at a higher price
CirexxSanta Clara, California, USAyesyesyes (turnkey or consigned)Second US option
Advanced CircuitsAurora, Colorado, USAyescheckyesUS quick-turn; check its rigid-flex assembly
EpecNew Bedford, Massachusetts, USAyescheckyes; also builds battery packsUS option that can also make a certified LiFePO4 pack

From each maker's own capability pages, checked 3 November 2026. Limits change, so each quote confirms them for our exact stack-up.

Full design notes

All 62 parts, what is on each piece, every connection, heat, power, the factory test, cost and timeline.