A wrist-worn 60 GHz presence sensor, built on the LILYGO T-Watch Ultra reference design with one radar added. This deck states only what the physics, the ordinances and the distributor price lists actually support.
radaruhr.ch has been advertising a device that cannot exist. Before asking anyone to fund the real one, here is every claim that fails, and what replaces it. Each is killed by a measurement or a statute, not by an opinion.
The T-Watch Ultra cannot be modified. LILYGO's own pin map lists 50 signals and marks every one of them unavailable — no header, no pogo pins, no castellated pads. Adding a sensor therefore means our own board. What LILYGO gives us is a validated schematic, a working AMOLED driver stack and an MIT-licensed firmware library: six months of de-risking, not a chassis.
Six candidates were priced and read at datasheet level. Only one fits a watch case, a watch power budget and ships both a presence and a breathing library. The others fail on a specific, quotable number — not on preference.
This is the chart that ends the through-wall story. NIST measured these losses at 60.5 GHz with an 18.1 dBi horn at 36 dBm EIRP — 27 dB more power than a watch may legally radiate. Radar is a two-way path, so every bar counts twice.
A 10° beam needs a 2.9 cm aperture at 60 GHz — watch-sized. At 3 GHz, where walls are transparent, the same beam needs 57 cm. There is no engineering trick around diffraction.
Camero's Xaver 400 reaches 20 m through masonry at 3–10 GHz. It is 37 × 22.5 × 12 cm and weighs 3.2 kg. The 100 m version weighs about 20 kg — at under 1 mW, which proves the win comes from band and aperture, never power.
Swiss RIR1008-31 allows 100 mW EIRP at 57–64 GHz. The FCC caps field disturbance sensors — exactly what a presence radar is — at 10 dBm. Our chosen part radiates 11.4 dBm EIRP and does not even reach the Swiss limit.
Drag the occupancy up and watch what a wrist-mounted sensor can still resolve. Targets merge when they fall inside one angular bin, and disappear entirely when a nearer body blocks them — millimetre waves do not pass through a torso. The counter reports only what is genuinely separable.
20 tracks. TI's people-counting reference design hard-codes maxNumTracks = 20 — on a mains-powered 2 W board with a dedicated DSP.
3–6 people. TI's own table: frames with zero errors fall to 40.2% at three people, 33.4% at four, 29.1% at five. The software limit is never reached in practice.
1–3 targets. Infineon documents its wrist-sized array at 2–3 humans to 5 m. The A121 has no angle channel at all — it reports presence per zone, and does not count.
A cardiac chest displacement of 0.25 mm produces a 36° phase swing at λ = 4.96 mm. A 2.5 mm movement of the sensor is a complete phase wrap. Move the slider and watch a real DFT lose the cardiac line — this is live arithmetic on a synthesised return, not an animation.
Double-integrating a consumer MEMS accelerometer accumulates 4.9 mm of position error across the 10 s window vital-sign extraction needs — and that is at a best-case 0.01 mg bias. At a realistic 1 mg it is 490 mm. The error is 20× the signal.
Watch removed and set down, or forearm braced; subject 0.3–1.5 m away; 10–30 s dwell. In that mode the published accuracy is roughly 0.5–2 BPM for respiration and 5–12% for heart rate. Acconeer ships a breathing reference application; heart rate is our own R&D.
This is where the A121 earns its place. At 43 mA drawn from the cell it runs continuously for 18.5 hours; the Infineon part manages 7.9 and the TI part 3.4. Drag the duty cycle to see the whole curve.
Three routes, costed to the same BOM. The Swiss route is 3.2× the landed cost at 1,000 units and the gap does not close with volume — because it is the mechanicals, not the assembly labour, that carry the difference.
Summary below. The complete line-by-line bill of materials for a Swiss build — every component, the case set, Swiss labour, per-unit radar calibration and yield, with the assumptions exposed as controls — is at the Swiss manufacturing cost model.
18 February 2027. EU Batteries Regulation Art. 11 requires portable batteries to be user-replaceable without special tools. A sealed watch case is non-compliant by default. The derogation for washable wearables is narrow and must be designed for deliberately.
The Infineon part carries a 26-week manufacturer lead time; the Acconeer part 9 weeks. This is a second reason the A121 wins, and it should drive the funding timeline more than any engineering milestone.
UN 38.3 battery transport testing, RoHS self-declaration with supplier evidence, REACH plus a SCIP notification, WEEE/EPR registration per country (€5k–10k a year across five markets) and the Swiss SWICO advance recycling fee.
Four research firms put the 2025 fall-detection market at USD 347 M, 470 M, 517 M and 729 M, with CAGRs from 5.9% to 22.4%. No tier-one analyst publishes a radar-presence figure at all. So this deck does not quote a TAM — it counts buyers.
Board respin, RED file opened, 50 paid pilot units into lone-worker and mobile-patrol sites. Care homes are deliberately excluded from the first go-to-market: it is the segment where one salesperson saying "monitor" turns this into a Class IIb medical device. The deliverable is a test report and a signed pilot, not a sales figure.
CE complete, 600 then 1,800 units through safety distributors. Break-even sits at roughly 1,100 units cumulative on this route and price — reachable in year 3, not year 2.
Only if the pilots show a clinical signal worth 12–24 months and €60k–250k. If they do not, the wellness product is a real business at 3,500 units a year. Rushing this decision is the fastest way to lose the company.
To first certified shipment on Route C, with 24 months of runway. That is the NRE base plus a small Swiss team, and it is sized to reach a CE test report and two paying pilots — not to reach scale.
Presence detection at 0.2–5 m and a breathing trace in docked mode, on an XM125 prototype. We will not show you a through-wall demonstration, because there is not one.
Every figure in this deck traces to a datasheet, an ordinance, a distributor page or our own stated arithmetic. The workings are open, including the ones that are unflattering.
Lone workers and mobile security patrols — the two segments where the site changes daily and there is nothing to mount a sensor on. 50 paid units, six months. If mobility does not beat a CHF 70 ceiling sensor there, the honest answer is to stop.