Hardware roadmap · 15 yrs software, ~0 yrs solder

Don't book Shenzhen yet. Book it for February.

Powerhouse's bootcamp — the reason to go — launches Q1 2027, per their own site. That gives you ~24 weeks. Spend them at a desk in your own city getting to a board that boots, then land in Shenzhen with a prototype instead of an idea. The plane ticket is worth far more on the second visit.

24
Weeks of home lab
Sep 2026 → mid-Feb 2027
$1,855
Total gear to buy
Staged across 3 purchases
$2
5 bare PCBs, JLCPCB
~$180 populated + shipped
$2.4k/mo
Shenzhen, all-in
Living + desk, solo

Gear at US retail, Aug 2026. Bootcamp fee not published.

The shape of it

Six phases, each ending in an object you can hand someone

Every phase gates on a physical artifact, not a finished course. Reading is not progress; a thing on your desk that works is.

P0
Weeks 1–2 · Wire it, then measure it

Stop trusting the code. Start trusting the meter.

  • Ohm's law, voltage dividers, pull-ups, decoupling caps — by measuring, not reading.
  • I²C, SPI, UART on a scope. See a real bus transaction.
  • ESP32-S3 with Arduino IDE + PlatformIO. Skip the Uno except as a teaching prop.
Gate: sensor → ESP32 → MQTT broker you run, and you can say why the pull-ups are 4.7 kΩ.
P1
Weeks 3–6 · Power is where software people die

Batteries, regulators, and the current you didn't budget for

  • LiPo + charge IC + protection, USB-C power delivery basics, LDO vs buck.
  • Deep sleep, wake sources, µA-level current measurement.
  • Level shifting, brown-outs, inrush, why your Wi-Fi resets the board.
Gate: a battery node that survives 7 days on one charge — with the measured sleep current written down.
P2
Weeks 7–10 · Make it an object

v0 of the voice puck: ugly, plastic, and used by someone else

  • CAD (Onshape or code-CAD via CadQuery), FDM tolerances, heat-set inserts, snap fits.
  • Mic array + I²S audio + speaker; wake word on-device with microWakeWord/ESPHome.
  • Buy an ESP32-S3-BOX-3 first as the reference you're allowed to copy.
Gate: someone who isn't you uses it for a week without asking you to reset it.
P3
Weeks 11–16 · The real unlock

Your own PCB, three spins, at $180 a spin

  • KiCad 10: symbols, footprints, ERC, ground planes, antenna keepouts.
  • Design for JLCPCB's assembly library — pick parts they stock, not parts you like.
  • Bring-up procedure, hot-air rework, cutting traces and bodging wires.
Gate: rev C boots. You found your own layout mistake twice and fixed it.
P4
Weeks 17–20 · Think like a manufacturer

A BOM, a landed cost at 1,000 units, and a DFM review

  • BOM costing at 100 / 1k / 10k, part lifecycle and second sources.
  • Injection molding vs CNC vs printed; draft angles, wall thickness, mold ownership.
  • Certification map: FCC/CE, battery UN38.3, and what a pre-scan costs.
Gate: a costed BOM plus written DFM feedback from a real factory contact.
P5
Weeks 21–24 · Shenzhen, on purpose

Arrive with Gerbers, not a pitch deck

  • Huaqiangbei sourcing runs; three prototype cycles in the time one takes elsewhere.
  • Bao'an factory visits, quotes, and a small EVT-quality build of 5–10 units.
  • Powerhouse coworking now; the 8-week bootcamp when it opens in Q1 2027.
Gate: 10 units in the hands of 10 users who aren't your friends.
The AI method

AI is a brilliant tutor and a confident liar — split the work by that line

The failure mode is well documented: models invent pin labels and register maps with total fluency. Give AI everything upstream of the physical world, and let instruments grade every claim about a physical quantity.

Hand it over

AI is genuinely faster than you

  • ✓ Firmware scaffolding: ESP-IDF, ESPHome YAML, Arduino
  • ✓ Explaining a datasheet section you paste in
  • ✓ Toolchain, build, Kconfig and linker errors
  • ✓ Parametric enclosures as CadQuery/OpenSCAD code
  • ✓ Turning an idea into a candidate parts list to evaluate
  • ✓ BOM spreadsheets, test scripts, supplier emails in Chinese
Verify every time

AI will lie to you with a straight face

  • ✕ Pinouts and register maps — check the datasheet, always
  • ✕ Whether a circuit actually works electrically
  • ✕ Current draw, thermals, EMC, RF behaviour
  • ✕ Part availability and today's real prices
  • ✕ Mechanical fit — calipers decide, not the model
  • ✕ Ground planes, return paths, antenna keepouts

The loop that actually compounds

01
Over-specify
Exact part numbers, board rev, toolchain version, clock config. Vague context is what makes it hallucinate.
02
Build small
One subsystem per iteration. Never a whole board from one prompt.
03
Measure
Meter, scope, logic analyzer, calipers. Numbers, not vibes.
04
Paste it back
Feed the measurement in. The model is only as good as your last reading.

Close the loop for the agent, not just for yourself

Your unfair advantage over a hardware native: you can give a coding agent shell access to the board and let it iterate unattended.

# give the agent the whole build-flash-observe cycle
arduino-cli compile --fqbn esp32:esp32:esp32s3 .
esptool.py --port /dev/ttyACM0 write_flash ...
python3 read_serial.py --port /dev/ttyACM0 --seconds 10   # it reads its own output

# then: "flash it, watch the serial log for 10s, fix the I2C timeout, repeat"

Compile errors and serial logs are ground truth the model can't fake. Register maps still need the datasheet.

The product

One device, four times — each version drops a crutch

Build the same open, programmable voice puck four times rather than four different toys. Repetition on one artifact is what turns tutorials into competence.

V0 · WEEK 10
The ugly puck
  • Dev board + breakout modules
  • Perfboard, hand-soldered
  • Printed box, visible screws
  • Wall power
Crutch dropped: none. Just finish it.
V1 · WEEK 14
Untethered
  • LiPo + USB-C charging
  • Deep sleep, wake word local
  • Printed shell, heat-set inserts
  • OTA firmware updates
Crutch dropped: the wall socket.
V2 · WEEK 18
Your silicon
  • Custom 2-layer PCB, ESP32-S3 module
  • JLCPCB assembled, 5 units
  • Half the volume of v1
  • Real connectors, no jumper wires
Crutch dropped: the dev board.
V3 · SHENZHEN
Ten of them
  • 4-layer board, proper antenna
  • Factory-quoted enclosure
  • Assembly + test fixture
  • Costed BOM at 1k units
Crutch dropped: you doing the assembly.
Money

$265 gets you started. Don't spend the rest until phase 2.

Buying the full lab up front is the classic way to convert enthusiasm into a shelf of boxes. Each tier unlocks only when its phase begins.

Item Unlocks at USD
ELEGOO UNO super starter kitP043
ESP32-S3-DevKitC ×3P045
Sensors, OLED, I²S mic, servosP060
Soldering station + hand toolsP080
Multimeter + USB logic analyzerP037
Tier 0 — desk kitWeek 1265
Bambu Lab A1 3D printerP2299
Filament, PLA / PETG / ASAP290
Entry 2-ch oscilloscopeP1150
Bench PSU, 0–30 VP175
Hot-air rework stationP370
LiPo cells, chargers, protectionP160
ESP32-S3-BOX-3 reference deviceP260
Calipers, inserts, ESD mat, lightP285
Tier 1 — the workshopWeek 3889
KiCad 10P30
Bare PCBs, 5 pcs + shippingP328
Assembled PCBA spin ×3 (stencil, NRE, parts)P3540
Bodge wire, spare parts, second attemptsP3133
Tier 2 — first custom boardWeek 11701
Everything, weeks 1–20—1,855

US retail, Aug 2026. Roughly 30–40% cheaper bought in Huaqiangbei.

Optional · Track R

Robotics, if it pulls you

  • → SO-ARM101 open 6-DOF arm kit — from $199, 3–4 h build, trains with Hugging Face LeRobot
  • → Second arm for leader–follower teleoperation — ~$400 the pair
  • → Rover chassis + motor drivers + depth cam — ~$200

Skip the Raspberry Pi 5 unless you need Linux — the memory shortage pushed the 16 GB board past $200.

Shenzhen · 3 months

Roughly $8–11k plus flights

  • → 1-bed city centre — ~$1,150/mo
  • → All-in comfortable solo — $1,800–2,870/mo
  • → Coworking desk — $400–800/mo
  • → Round-trip flights — ~$1,200

No digital-nomad visa exists. M (business) visa is the usual route; Shenzhen is a pilot city for online residence-permit extensions. Confirm with the consulate.

Reality check

Everything above costs less than 2% of your first production run

The learning is cheap. Tooling is what ends hardware companies — and a steel mold is machined long after the design must be frozen.

Injection mold: $3k–$30k+ Get mold ownership in writing Freeze geometry before steel is cut PVT ships 5–10% of the run
Operating rhythm

12 hours a week, split three ways

~6 H
Build
Hands on the bench. The only hours that count toward the gate.
~3 H
Read the primary source
One datasheet or app note a week, unassisted. This is the skill AI can't lend you.
~2 H
Publish
Post the build log. It's how you meet the people who answer your next question.
~1 H
Source
Browse LCSC/JLCPCB stock. Learn what's actually cheap and available.
This week

Four moves, none of which need a plane ticket

  • 1 Order Tier 0 — $265, arrives in days.
  • 2 Email Powerhouse: confirm the Q1 2027 bootcamp dates, cost, and whether a coworking week in October is worth it.
  • 3 Ask your hardcore friend for one thing only: a DFM review at week 18.
  • 4 Pick the puck's one non-obvious feature now — the thing an Echo can't do. It decides every part choice downstream.
Sources & assumptions
  • Powerhouse Engineering — services, spaces, bootcamps launching Q1 2027 — powerhouse.engineering
  • JLCPCB — PCB prototypes from $2 for 5 boards; PCBA setup, stencil and NRE breakdown — jlcpcb.com/blog/pcba-cost-breakdown
  • Home Assistant — ESP32-S3-BOX-3 and Voice Preview Edition, microWakeWord on-device — home-assistant.io/voice_control
  • Flux.ai / ProtoFlow — AI in ECAD 2026; LLMs invent invalid pin labels and violate electrical constraints — protoflow.ai/blog/ai-pcb-design-2026-guide
  • Adafruit — giving Claude Code direct access to Arduino hardware — blog.adafruit.com, Mar 2025
  • Seeed / Hugging Face — SO-ARM101 LeRobot arm from $199, 3–4 h build
  • Bambu Lab pricing 2026 — A1 $299, P1S $699 MSRP (frequently $399)
  • Raspberry Pi — memory-driven price rises, 16 GB model above $200 — raspberrypi.com/news
  • Encata / KD Product Dev — EVT $5–20k, DVT + tooling $50–200k+, PVT $20–100k; molds $3–30k+
  • Numbeo, Wise, Expatistan — Shenzhen cost of living, Aug 2026; coworking $400–800/mo
  • China visa guidance 2026 — no digital-nomad visa; M visa route; Shenzhen residence-permit extension pilot
  • Assumption: 12 focused hours/week, start Sep 2026, three PCB spins budgeted.