Hardware as fast as software

On 27 August I replied to Kyra’s founder Sahil Dhull that hardware design should move as fast as software^[1]. On 2 September we tried to prove it with Kyra’s own public roadmap, which is two sentences. We took two boards to part selection before lunch. This is the record, with the commits.

Rank 27 in JEE Advanced. 230 in JEE Mains out of 1.5 million fellow candidates.

Nah! Wasn't enough.

Raised $1.2M, backed by a16z Speedrun, building AI agents.

Still Nah! Wasn't enough.

I wanted hardware too - the actual thing on your body, not just the software behind it.… pic.twitter.com/IVbMheUFzZ

sahil dhull (@_sahildhull), August 27, 2026

The reports

Design package	Revision	Pages	File
Kyra R1 EV1, the escalation glasses engine	3297c80	88	PDF, 0.9 MB
Kyra R1 EV1, the escalation ring engine	c19ab02	102	PDF, 1.0 MB

Drawing	Sheet	File
FIG. 1, the glasses exploded	patent style, reference numerals	SVG
FIG. 2, the ring exploded along its axis	patent style, reference numerals	SVG
KYRA-GL-001, the glasses, dimensioned	A3, 1:1, third angle	SVG
KYRA-RG-001, the ring, dimensioned	A4, 2:1, third angle, section A-A	SVG

The input

Kyra’s entire public hardware specification, quoted in full from kyrainterface.com^[2]:

Mac and iOS first, because that is where your work happens. Then ambient glasses controlled by a ring: reply, dismiss or confirm with your thumb.

Plus a privacy commitment: your data stays with you. No part numbers, no budgets, no dimensions, no display or camera decision. The evening before, Sahil Dhull had posted a working V0 of the glasses, demonstrated by making a phone call through them^[3]. That post settled one question the roadmap did not: the glasses have an audio path.

The record

Hashes are commits in the two design repositories; each stage of the pipeline commits when its completion contract holds, so the hash is the evidence^[4][5]. The pipeline is copperhead create. The scope was its first three stages, specification, architecture and part selection, stopped by design at the schematic boundary.

	Ring	Glasses
Workspace created	f4a625b	ed41083
Stage 1, specification	25m 19s, 786b723	28m 17s, 0c2f005
Stage 2, architecture	11m 01s, 370678f	7m 51s, abad67f
Stage 3, part selection	43m 28s, c19ab02	48m 44s, 3297c80
Stages 1 to 3	1h 20m	1h 25m
Placed symbols	90	90
Design package	102 pages	88 pages

The ring’s brief was written the night before, from the roadmap above and a market study, with ten engineering decisions taken in the absence of a client and the consequence of each written down. The glasses’ brief was written while the ring ran, after every candidate part had been checked against the 222 symbol libraries on the machine, because a part without a symbol makes a run unwinnable. Between the two sits the contract that ties the boards together: glasses as the escalation surface, ring as the confirm surface, four links between them and the phone. Both design packages were assembled from the stage commits: brief, specification, architecture, bill of materials, decision log, change log and every requirement file.

What each stage cost

Every figure here is read off the Claude Agent SDK’s own usage record, one line per API call and priced at the published Claude Opus 5 rates: $5 per million input tokens, $10 per million written to the one hour cache, $0.50 per million read back from it and $25 per million output tokens^[6]. The model was claude-opus-5 for every call.

Uncached input across the six stages is 164 tokens. The harness writes each turn’s context into the one hour cache and reads it back on the next turn, so the bill sits in two lines: cache writes at $18.57 and output tokens at $6.59. Reads are $0.86. Writes cost more than output because every turn appends to the context and the appended part is written at twice the input price. The five minute cache, at 1.25 times, would have cut that line by more than a third, but stage 3 turns ran past five minutes often enough that the one hour cache is the right choice.

Part selection is the largest line on both boards, 45 percent of the ring’s bill and 48 percent of the glasses’. It is also where the wall clock goes: 43 and 49 minutes against under half an hour for a specification.

What came out

Figure	Glasses EV1, 3297c80	Ring EV1, c19ab02
Placed symbols, against the ceiling	90 of 110	90 of 105
Parts, all symbol and pin verified	77	78
Radios	BM83 audio, nRF52840	nRF52833, VDDH mode
Standby	3.25 mA ledger	96.9 µA itemised
Design figure against target and ceiling	4.77 mA against 8 and 12 mA	116.3 µA against 150 and 177 µA
Predicted endurance	about 50 h on 240 mAh usable	about 6 days on 17 mAh usable
Stage-3 turns used, cap 80	46	35 then 38
Package	88 pages	102 pages

FIG. 1. Glasses, exploded. No display, no camera. Microphone power is a hardware switch under the application controller, off unless the wearer opens a call or a reply. Reference numerals key to the requirement register. Download the SVG.

FIG. 2. Ring, exploded along its axis. No microphone, no camera, no optical or electrical biosensor: the one window on the inner liner is a wear-detect electrode. Every shipping smart ring is a capture device^[7][8][9]; this one is a control device. The architecture follows the published Open Ring design record^[10]. Download the SVG.

The glasses, as designed

Two radios, because an iPhone carries a call over Bluetooth Classic and the Kyra app needs a Bluetooth Low Energy data channel and a central role toward the ring. The audio module owns the call path and the charger; the controller module owns the sensors, the touch strip, the microphone switch and both BLE links. Everything off the board is dashed.

Glasses EV1 block diagram. Download the SVG.

The 23 parts that are not passives, from the bill of materials committed at 3297c80. Every part number is a proposal, flagged unverified in the BOM itself; the package column is the KiCad footprint the symbol was checked against.

The other 67 rows: passives and test points

Refdes	Value	Package	MPN (proposed, unverified)
R1	1k	R_0402_1005Metric	RC0402FR-071KL
R2	5.1k	R_0402_1005Metric	RC0402FR-075K1L
R3	5.1k	R_0402_1005Metric	RC0402FR-075K1L
R4	330R	R_0402_1005Metric	RC0402FR-07330RL
R5	2k	R_0402_1005Metric	RC0402FR-072KL
R6	10k	R_0402_1005Metric	RC0402FR-0710KL
R7	10k	R_0402_1005Metric	RC0402FR-0710KL
R8	10k	R_0402_1005Metric	RC0402FR-0710KL
R9	100k	R_0402_1005Metric	RC0402FR-07100KL
R10	120k	R_0402_1005Metric	RC0402FR-07120KL
R11	120k	R_0402_1005Metric	RC0402FR-07120KL
R12	100R	R_0402_1005Metric	RC0402FR-07100RL
R13	100R	R_0402_1005Metric	RC0402FR-07100RL
R14	100R	R_0402_1005Metric	RC0402FR-07100RL
R15	100R	R_0402_1005Metric	RC0402FR-07100RL
R16	10R	R_0402_1005Metric	RC0402FR-0710RL
R17	100k	R_0402_1005Metric	RC0402FR-07100KL
R18	1k	R_0402_1005Metric	RC0402FR-071KL
R19	1k	R_0402_1005Metric	RC0402FR-071KL
R20	1k	R_0402_1005Metric	RC0402FR-071KL
R21	1k	R_0402_1005Metric	RC0402FR-071KL
R22	1k	R_0402_1005Metric	RC0402FR-071KL
R23	4.7k	R_0402_1005Metric	RC0402FR-074K7L
R24	4.7k	R_0402_1005Metric	RC0402FR-074K7L
R25	10k	R_0402_1005Metric	RC0402FR-0710KL
R26	10k	R_0402_1005Metric	RC0402FR-0710KL
C1	10uF	C_0805_2012Metric	GRM21BR61C106KE15L
C2	100nF	C_0402_1005Metric	CL05B104KO5NNNC
C3	100nF	C_0402_1005Metric	CL05B104KO5NNNC
C4	4.7uF	C_0603_1608Metric	GRM188R61A475KE15D
C5	10uF	C_0805_2012Metric	GRM21BR61A106KE19L
C6	100nF	C_0402_1005Metric	CL05B104KO5NNNC
C7	4.7uF	C_0603_1608Metric	GRM188R61A475KE15D
C8	100nF	C_0402_1005Metric	CL05B104KO5NNNC
C9	100nF	C_0402_1005Metric	CL05B104KO5NNNC
C10	4.7uF	C_0603_1608Metric	GRM188R61C475KAAJ
C11	10uF	C_0805_2012Metric	GRM21BR61A106KE19L
C12	100nF	C_0402_1005Metric	CL05B104KO5NNNC
C13	1uF	C_0402_1005Metric	CL05A105KA5NQNC
C14	4.7uF	C_0603_1608Metric	GRM188R61A475KE15D
C15	1uF	C_0402_1005Metric	CL05A105KA5NQNC
C16	220nF	C_0402_1005Metric	CL05B224KO5NNNC
C17	220nF	C_0402_1005Metric	CL05B224KO5NNNC
C18	220nF	C_0402_1005Metric	CL05B224KO5NNNC
C19	220nF	C_0402_1005Metric	CL05B224KO5NNNC
C20	100nF	C_0402_1005Metric	CL05B104KO5NNNC
C21	100nF	C_0402_1005Metric	CL05B104KO5NNNC
C22	10uF	C_0805_2012Metric	GRM21BR61A106KE19L
C23	1uF	C_0402_1005Metric	CL05A105KA5NQNC
C24	100nF	C_0402_1005Metric	CL05B104KO5NNNC
C25	100nF	C_0402_1005Metric	CL05B104KO5NNNC
C26	100nF	C_0402_1005Metric	CL05B104KO5NNNC
C27	100pF	C_0402_1005Metric	GJM1555C1H101JB01D
C28	100pF	C_0402_1005Metric	GJM1555C1H101JB01D
TP1	GND	TestPoint_Pad_D1.0mm	none (pad or hole)
TP2	GND	TestPoint_Pad_D1.0mm	none (pad or hole)
TP3	VBAT_PROT	TestPoint_Pad_D1.0mm	none (pad or hole)
TP4	VDD_3V0	TestPoint_Pad_D1.0mm	none (pad or hole)
TP5	MIC_VDD	TestPoint_Pad_D1.0mm	none (pad or hole)
TP6	BT_UART_TXD	TestPoint_Pad_D1.0mm	none (pad or hole)
TP7	BT_UART_RXD	TestPoint_Pad_D1.0mm	none (pad or hole)
TP8	SPK_L_P	TestPoint_Pad_D1.0mm	none (pad or hole)
TP9	SPK_R_P	TestPoint_Pad_D1.0mm	none (pad or hole)
TP10	SDA	TestPoint_Pad_D1.0mm	none (pad or hole)
TP11	SCL	TestPoint_Pad_D1.0mm	none (pad or hole)
TP12	MIC_EN_N	TestPoint_Pad_D1.0mm	none (pad or hole)
TP13	AMP_SD_N	TestPoint_Pad_D1.0mm	none (pad or hole)

The ring, as designed

One radio and no regulator: the nRF52833 takes the cell directly in high-voltage mode. A dedicated touch controller scans on its own and wakes the controller once per gesture, which is the whole power budget. Charging is inductive through a discrete rectifier, and the haptic pulse is drawn from a capacitor bank rather than the cell.

Ring EV1 block diagram. Download the SVG.

The 23 parts that are not passives, from the bill of materials committed at c19ab02, on the same terms.

The other 67 rows: passives, test points and mounting holes

Refdes	Value	Package	MPN (proposed, unverified)
R1	330R	R_0402_1005Metric	Yageo RC0402FR-07330RL
R2	2k	R_0402_1005Metric	Yageo RC0402FR-072KL
C1	100nF	C_0402_1005Metric	Murata GRM155R71C104KA88D
R3	66.5k	R_0402_1005Metric	Yageo RC0402FR-0766K5L
C2	4.7uF	C_0805_2012Metric	Murata GRM21BR61C475KA88L
C3	4.7uF	C_0603_1608Metric	Murata GRM188R61A475KE15D
R4	2.2k	R_0402_1005Metric	Yageo RC0402FR-072K2L
R5	100k	R_0402_1005Metric	Yageo RC0402FR-07100KL
R6	100k	R_0402_1005Metric	Yageo RC0402FR-07100KL
C4	560pF	C_0402_1005Metric	Murata GRM1555C1H561JA01D
C5	DNP	C_0402_1005Metric	none, position not fitted at build
C6	10uF	C_0805_2012Metric	Murata GRM21BR61C106KE15L
C7	100nF	C_0402_1005Metric	Murata GRM155R71C104KA88D
R7	1M	R_0402_1005Metric	Yageo RC0402FR-071ML
R8	470k	R_0402_1005Metric	Yageo RC0402FR-07470KL
C8	100nF	C_0402_1005Metric	Murata GRM155R71C104KA88D
C9	12pF	C_0402_1005Metric	Murata GRM1555C1H120JA01D
C10	12pF	C_0402_1005Metric	Murata GRM1555C1H120JA01D
C11	12pF	C_0402_1005Metric	Murata GRM1555C1H120JA01D
C12	12pF	C_0402_1005Metric	Murata GRM1555C1H120JA01D
C13	1uF	C_0402_1005Metric	Murata GRM155R60J105KE19D
C14	100nF	C_0402_1005Metric	Murata GRM155R71C104KA88D
C15	100nF	C_0402_1005Metric	Murata GRM155R71C104KA88D
C16	820pF	C_0402_1005Metric	Murata GRM1555C1H821JA01D
C17	47nF	C_0402_1005Metric	Murata GRM155R71C473KA01D
C18	100nF	C_0402_1005Metric	Murata GRM155R71C104KA88D
C19	100nF	C_0402_1005Metric	Murata GRM155R71C104KA88D
C20	100nF	C_0402_1005Metric	Murata GRM155R71C104KA88D
C21	4.7uF	C_0603_1608Metric	Murata GRM188R61A475KE15D
C22	100nF	C_0402_1005Metric	Murata GRM155R71C104KA88D
C23	4.7uF	C_0603_1608Metric	Murata GRM188R61A475KE15D
R9	10k	R_0402_1005Metric	Yageo RC0402FR-0710KL
R10	4.7k	R_0402_1005Metric	Yageo RC0402FR-074K7L
R11	4.7k	R_0402_1005Metric	Yageo RC0402FR-074K7L
C24	0.5pF	C_0402_1005Metric	Murata GRM1555C1HR50BA01D
C25	0.5pF	C_0402_1005Metric	Murata GRM1555C1HR50BA01D
C26	300pF	C_0402_1005Metric	Murata GRM1555C1H301JA01D
C27	300pF	C_0402_1005Metric	Murata GRM1555C1H301JA01D
R12	1k	R_0402_1005Metric	Yageo RC0402FR-071KL
R13	1k	R_0402_1005Metric	Yageo RC0402FR-071KL
R14	1k	R_0402_1005Metric	Yageo RC0402FR-071KL
R15	1k	R_0402_1005Metric	Yageo RC0402FR-071KL
R16	1k	R_0402_1005Metric	Yageo RC0402FR-071KL
R17	10k	R_0402_1005Metric	Yageo RC0402FR-0710KL
R18	10k	R_0402_1005Metric	Yageo RC0402FR-0710KL
C28	100nF	C_0402_1005Metric	Murata GRM155R71C104KA88D
C29	100nF	C_0402_1005Metric	Murata GRM155R71C104KA88D
C30	100nF	C_0402_1005Metric	Murata GRM155R71C104KA88D
R19	15R	R_0805_2012Metric	Yageo RC0805FR-0715RL
R20	100k	R_0402_1005Metric	Yageo RC0402FR-07100KL
C31	220uF	C_1210_3225Metric	Murata GRM32EC80J227ME05L
C32	220uF	C_1210_3225Metric	Murata GRM32EC80J227ME05L
C33	220uF	C_1210_3225Metric	Murata GRM32EC80J227ME05L
C34	220uF	C_1210_3225Metric	Murata GRM32EC80J227ME05L
C35	100nF	C_0402_1005Metric	Murata GRM155R71C104KA88D
C36	1uF	C_0402_1005Metric	Murata GRM155R60J105KE19D
TP1	TP	TestPoint_Pad_D1.0mm	none, copper pad
TP2	TP	TestPoint_Pad_D1.0mm	none, copper pad
TP3	TP	TestPoint_Pad_D1.0mm	none, copper pad
TP4	TP	TestPoint_Pad_D1.0mm	none, copper pad
TP5	TP	TestPoint_Pad_D1.0mm	none, copper pad
TP6	TP	TestPoint_Pad_D1.0mm	none, copper pad
TP7	TP	TestPoint_Pad_D1.0mm	none, copper pad
TP8	TP	TestPoint_Pad_D1.0mm	none, copper pad
TP9	TP	TestPoint_Pad_D1.0mm	none, copper pad
H1	M2	MountingHole_2.2mm_M2	none, mechanical
H2	M2	MountingHole_2.2mm_M2	none, mechanical

What this does not prove

• Anything past part selection. There is no schematic, no ERC, no layout and no measurement. Both packages say so on their first pages.

• That the parts are right. Every part number is proposed from catalogue knowledge and flagged UNVERIFIED, with the datasheet claim that must be checked written next to it. That check is a lookup, not a design change. It has not been done.

• That the budgets hold. The glasses figure rests on one assumed number, the audio module’s sniff current, entered at 3.00 mA and marked as assumed. The ring’s haptic bank came out smaller than the brief assumed, and the pulse it can carry is now an open item. Both are the first things the boards would measure.

• That Kyra wants any of this. The two biggest decisions, no display and no camera, were ours, taken from a three-day prototype and a two-sentence roadmap. They are listed as open items addressed to Kyra, unasked.

What it does prove

That the slow part of hardware comes before the schematic: turning intent into a specification precise enough to argue with, choosing an architecture and building a bill of materials in which every part is known to exist, to have a symbol, to have the pins you will wire and to fit a power budget itemised line by line. That part took a working day for two boards, and every claim in it traces to a commit, a decision number or an assumption with its consequence written down.

Software teams take that for granted: a spec, a design, a build, a diff, a log. Hardware mostly does not have it, and that is the gap copperhead is built for^[11]. The proof is not the pitch. It is the record above, which anyone can check out.

References

Each entry says which claim in the text it carries. Web sources were read on 2 September 2026; local sources are files in the engagement record, named by the commit they were built from.

1. Animesh Chouhan, reply to @_sahildhull on X, 27 August 2026.
   https://x.com/animeshsingh38/status/2093044844423946444
   The claim this page sets out to prove: imagine what you could build if hardware design moved as fast as software.

2. Kyra, product page, quoted in full.
   https://kyrainterface.com/
   The entire public hardware specification (The input) and the privacy commitment that became a hardware constraint on both boards. Re-checked the same day the boards were designed; unchanged.

3. Sahil Dhull, LinkedIn post, 1 September 2026.
   https://www.linkedin.com/feed/update/urn:li:activity:7500590483107344384/
   The glasses V0 built in three days and demonstrated by a phone call, which is the only evidence that the glasses carry an audio path. It is the reason the day’s scope moved from a ring alone to glasses and ring. Nothing else about the product is inferred from it.

4. Kyra R1 EV1, the escalation ring engine: design documentation package, revision c19ab02, 102 pages.
   /blog/kyra-ring-ev1-design-package-c19ab02.pdf
   Every ring figure in The record and What came out: the stage commits, the 90-symbol BOM, the itemised 96.9 µA idle draw and the 116.3 µA figure with contingency. Built from the workspace at that commit through git archive, so the revision on its cover is checkable.

5. Kyra R1 EV1, the escalation glasses engine: design documentation package, revision 3297c80, 88 pages.
   /blog/kyra-glasses-ev1-design-package-3297c80.pdf
   Every glasses figure: the stage commits, the 90-symbol BOM, the 3.25 mA standby ledger and 4.77 mA day average and the note that the audio module’s sniff current is entered as assumed, which What this does not prove relies on.

6. Anthropic, Claude pricing.
   https://platform.claude.com/docs/en/about-claude/pricing
   The Claude Opus 5 rates behind every cost in What each stage cost, including the one hour cache write and cache read multipliers. Token counts are the Claude Agent SDK’s session records, one usage entry per API call, deduplicated by message id.

7. Becky Stern, Oura Ring teardown (Gen 3 and Gen 2), 2022.
   https://beckystern.com/2022/04/17/oura-ring-teardown-gen-3-and-gen-2/
   The optical heart-rate stack, 16 mAh cell and inductive charging inside a shipping ring: one of three teardowns behind every shipping smart ring is a capture device (FIG. 2 caption).

8. iFixit, Samsung Galaxy Ring Chip ID.
   https://www.ifixit.com/Guide/Samsung+Galaxy+Ring+Chip+ID/176114
   Nordic nRF5340, NFC tag, external NOR flash and wireless charging in the Galaxy Ring: the second of the three teardowns and the source for Nordic silicon as the default in the ring’s part research.

9. DigiKey Maker, Ultrahuman Ring Air teardown, 2024.
   https://www.digikey.com/en/maker/blogs/2024/ultrahuman-ring-air-teardown
   nRF52840, flex PCB, LEDs and photodiode: the third teardown. The optical stack these three share is what the Kyra ring omits and what buys its power budget.

10. Memfault Interrupt, Smart ring development, parts 1 and 2, with the open-source hardware at https://github.com/stawiski/open-ring.
    https://interrupt.memfault.com/blog/smart-ring-development-part-1
    The published design record the ring’s architecture leans on: wafer-level packaging forced by a 2.6 to 2.9 mm band, rigid-flex, 6.78 MHz induction charging with a discrete rectifier and a capacitor bank for the haptic. It is a design one may read, not one being shipped; the Kyra ring departs from it on the touch controller and the second MCU.

11. copperhead.
    https://copperhead.sh
    The pipeline that produced the record: copperhead create, version 0.10.0, eight stages, of which the first three were run here.