✠ POCKET ORACLE ✠ ORACULUM SACCI ✠ BUILD CODEX V1

00 // PRINCIPIA FIRST PRINCIPLES

▸

A 1990 Sega Game Gear shell, fully gutted, rebuilt around a modern edge-AI compute module that runs a small local LLM at 30–50 tokens per second. The device is a dedicated, single-purpose code oracle: a portable terminal that drives a Claude-Code-style agent loop using its own on-board model, with the option to switch to a wireless tunnel and hand the work off to the home rig (Localeksi) for heavier tasks.

The primary input surface is 4 physical buttons plus the D-pad — approve, deny, scroll, prompt. The primary output is a modern square / portrait IPS or AMOLED panel mounted in the original screen window, rendering a TUI-style log of agent activity. It is a quiet, always-on artifact you carry like a Game Boy and use like a pager — only it codes. Only it divines.

Why this is worth building

Cross-domain demo. One object proves out edge LLM inference, embedded Linux, custom PCB design, mechanical retrofit, and modern agentic tooling. It is the synthesis piece that justifies everything you've been building toward.
Aesthetic singularity. Nobody is publishing this. A retro handheld that talks to itself in code is content-shaped for X and YouTube without trying to be.
Local-first AI is a thesis. Owning the silicon and the model on a portable device makes the case for sovereign compute in physical form. Timechain-adjacent narrative.
Time-bounded scope. Unlike a Trading Trinity refactor or a federal contract, this build has a hard mechanical floor — the shell. It cannot scope-creep past a certain volume.

First principles we accept

The local model is a receptionist, not an engineer. A 1.5–3B model on-device at 30–50 t/s can route, summarize, complete obvious patterns, and approve/deny on rules. It cannot architect a refactor. We design the UX around that truth.
The shell is the constraint. Internal volume of a Game Gear cavity is roughly 180mm × 95mm × 32mm. Every component must fit. We treat the shell as a hardware spec.
Thermals are the second constraint. Sustained inference at 7–25W in a sealed plastic shell requires active cooling. We will plan it from day one, not retrofit it.
Battery is the third constraint. Useful runtime ≥ 1 hour active inference, days of standby. We size the cells and the power tree to that target.
Reproduction shell first. Do not touch an original 1990 shell until the build is proven in a clone. Originals are brittle and irreplaceable in good condition.
Build in public, but ship one finished v1 before any "v2" energy. The narrative is "I built it." Don't fork into variants before v1 boots.

A working device, in a Game Gear shell, that I can hand to someone — they press a button, watch a coding agent run a real task, fully offline. That demo is the deliverable. Everything else serves it. Ex tenebris, lux.

01 // ANATOMIA SACRA THE SACRED ANATOMY · BLOCK DIAGRAM

▸

Treat the device as seven cooperating subsystems. Each one has a champion in the BOM and a milestone in the canonical hours.

Data flow · the divination loop

Wake. Power switch or D-pad press fires GPIO interrupt; NX wakes from suspend in <1s.
Capture intent. Button-driven menu OR press-and-hold mic for voice. Whisper-tiny renders speech to text in <500ms.
Local oracle loop. On-device Qwen2.5-Coder 3B (or successor) plans, calls local tools (shell, file edit, search) on the NX itself.
Approval surface. For any tool call flagged as destructive, render a diff on the window and require a button press. Mapping: A=approve · B=deny · L=defer · R=show-detail.
Escalation. If the local model returns low confidence or detects multi-file/architectural change, route the conversation over Tailscale to Localeksi. Stream tokens back.
Persist. Every session writes a JSONL log to NVMe. Push to GitHub on the next wifi connection.

02 // ELECTUS SILICII THE CHOSEN SILICON

▸

The hard requirement is 30–50 tok/s on a useful coder model. That is the screening filter. Anything that does not clear it becomes a thin client, not a brain.

Honest options · May 2026

OPTION	USEFUL MODEL	EXPECTED TOK/S	POWER	NOTES
Jetson Orin Nano Super 8GB	Qwen2.5-Coder 3B Q4 + spec decode	30–45	7–15W	Cheapest to clear the bar. 8GB caps future model growth.
Jetson Orin NX 16GB CHOSEN VESSEL	Qwen2.5-Coder 7B Q4 · or 3B with deep context	30–55	10–25W	Sweet spot. 16GB lets you host a permanent draft model in RAM for speculative decoding.
Hailo-10H + Raspberry Pi 5	Qwen2-1.5B INT4	~9	2.5W	Fails throughput. Beautiful chip, wrong fit for the spec.
RK3588 (Orange Pi 5+, etc.)	Qwen2.5 1.5B	8–12	5–8W	Fails throughput. Fallback only.
Jetson Thor Nano (announced)	Qwen3-Coder 7B+ confidently	60–100+ (est.)	~25W	If timing aligns, swap to this. Same SODIMM-260 socket as Orin NX.

Buy a Jetson Orin NX 16GB on a SeeedStudio reComputer J401 carrier for development. Design the custom carrier PCB around the NX SODIMM-260 socket. If Jetson Thor Nano lands before final layout, swap in. Otherwise NX 16GB is the floor.

Why the NX over the Nano Super

The Nano Super is ~$249 and clears the 30 t/s bar. The NX 16GB is ~$599 and gives you (a) double the RAM, which lets you permanently host a 0.5B speculative draft alongside the 3B target — the same MTP pattern you're already running on Localeksi, (b) 100 TOPS vs 67 TOPS, which buys headroom for larger context windows, and (c) the ability to run a 7B coder model at usable speed, which is the difference between "toy" and "actually useful for non-trivial tasks." Given the build cost is dominated by labor and mechanical (~$300+ in shells, prints, and finishing), saving $350 on silicon is the wrong economy.

When to actually buy

Order the Jetson at the start of Phase 2, not now. NVIDIA refreshes Jetson roughly every 18 months. If you're 4–6 months out from needing it, a Thor Nano announcement could justify waiting another month. The carrier PCB doesn't need final silicon until layout, and the SODIMM-260 footprint is consistent across Orin and Thor.

03 // RELIQUIAE THE SACRED RELICS · BILL OF MATERIALS

▸

Estimated total at v1 completion: $1,400 – $2,100 in parts, depending on Jetson tier and how many display/battery variants you prototype with. Costs assume 2026 US pricing and are approximate.

CEREBRUM · COMPUTE & STORAGE

RELIC	QTY	~$	NOTES
Jetson Orin NX 16GB module	1	599	SODIMM-260. Order direct from NVIDIA partners (SeeedStudio, Connect Tech).
SeeedStudio reComputer J401 dev carrier	1	199	Develop on this before committing to custom PCB.
NVMe SSD M.2 2230, 256GB	1	45	WD SN740 or similar. 2230 saves space.
microSD 64GB (fallback boot)	1	12	Sandisk Extreme. Keep one with stock JetPack.

FENESTRA · DISPLAY

Waveshare 3.5" MIPI-DSI IPS, 640×480	1	55	Primary candidate. Fits Game Gear screen window.
1.91" AMOLED RM67162 (SPI)	1	28	Alt for aesthetic: dark OLED behind tinted bezel.
Sharp Memory LCD LS027B7DH01 2.7"	1	42	Optional low-power "reading mode" secondary display.

IGNIS · POWER

21700 Li-ion cells, Molicel P42A or Samsung 50G	4	52	Protected cells, matched pair-of-pairs.
2S BMS, 20A continuous, balance leads	1	18	Daly or similar.
USB-C PD trigger board (20V/3A)	1	9	Wall charging input.
Buck converters (5V, 3.3V rails)	2	12	TPS-series modules.
Coulomb counter MAX17260	1	6	Real battery % over I2C.
Power switch + reset button	1	5	Slide for power, tactile for reset.

RESPIRO · COOLING

30mm 5V PWM blower fan	1	14	Sunon or Noctua NF-A4x10 if it fits.
Copper shim 30×30×2mm	2	8	SoC-to-shell heat path.
Thermal paste + pad set	1	12	Arctic MX-6, Thermalright pads.
Graphite thermal sheet	1	15	Spreads heat across rear shell.

VOX · MANUS · AETHER · INPUT/AUDIO/WIRELESS

Tactile switches (Kailh low-profile)	8	9	D-pad + 4 face + start + power.
Silicone button pad	1	10	Print or salvage from donor.
INMP441 I2S MEMS mic	1	5	Voice prompting.
MAX98357 I2S amp + 8Ω speaker	1	11	Audio out, TTS readback.
Intel AX210 M.2 wifi 6E + BT (optional)	1	22	If Jetson onboard is insufficient.

RECEPTACULUM · MECHANICAL

RetroSix or Castlemania reproduction GG shell	2	60	One prototype, one final. Black or clear.
FDM filament (PETG, ABS)	2 spools	50	Internal frame, button caps, vent grilles.
Resin printer + resin (optional)	—	—	Injection-quality button caps.
Screen bezel acrylic / tinted plexi	1	8	Custom-cut to screen window.
M2 brass heat-set inserts + screws	1 set	14	For securing the printed frame.

PCB FABRICATION · CUSTOM CARRIER

JLCPCB / PCBWay 4-layer board, 5pcs	1	40	~150×80mm carrier.
SMD components (passives, regulators, ICs)	—	~80	JLCPCB EasyEDA assembly to skip hand-placement.
SODIMM-260 socket	1	12	Match the exact NVIDIA reference.
Connectors (FPC, SH 1.0, USB-C)	—	25	Order in bulk to avoid downtime.

Order in clusters, not one-at-a-time. Group SMD + connectors + PCB in one JLCPCB order to share shipping. Group Jetson + carrier + storage in one NVIDIA-partner order. The whole BOM lands in roughly 3 shipments.

04 // INSTRUMENTA THE INSTRUMENTS · PHYSICAL KIT

▸

If you don't already own these, total tool investment is $700 – $1,400. Most of these you will keep using forever. None are optional; trying to skip core soldering or measurement gear leads to a dead build.

SOLDERING & REWORK CORE

Soldering station with hot air rework. Hakko FX-951 + 850D rework, or the all-in-one Aixun T3B/T420D. Hot air is non-negotiable for QFN/BGA work. ~$200–400
Microscope or USB digital scope. AmScope SE400-Z + ring light is the workshop standard. For mobile, Andonstar AD409 is fine. ~$200–350
Solder paste, leaded + lead-free. ChipQuik SMD291AX for leaded prototyping. Stored in a fridge. ~$30
Flux pen + flux gel. MG Chemicals 8341, AmTech NC-559. ~$25
Solder wick, ESD tweezers, kapton tape, isopropyl 99%. Boring but everything. ~$40
PCB holder / third hand. Stickvise or PanaVise Jr. ~$50

MEASUREMENT & DEBUG CORE

Bench multimeter. Owon B41T+ or Brymen BM235. ~$90
USB logic analyzer. Saleae Logic 8 if budget allows; otherwise a $15 8-channel clone for I2C/SPI. ~$15–400
Bench power supply. Riden RD6018 or any 0–30V/0–10A. Critical for safely bringing up the carrier PCB. ~$120
USB-C PD tester (KM003C). Confirms PD negotiation while developing the input stage. ~$50
Thermal camera or IR thermometer. InfiRay P2 Pro is the cheap modern pick. Essential for confirming cooling under load. ~$250

MECHANICAL PROJECT-SPECIFIC

FDM 3D printer. Bambu A1 or P1S. If owned, skip. ~$300–700
Dremel rotary tool + carbide bits. Shell cuts (vents, USB-C window). ~$70
Digital calipers. Mitutoyo absolute or equivalent clone. ~$30–120
Heat-set insert tool + iron tips. For M2/M3 brass inserts in the printed frame. ~$25
Fine files, deburring blade, sanding block. Finishing the shell cuts. ~$20

SAFETY NON-NEGOTIABLE

Fume extractor with carbon filter. Hakko FA-400 minimum. ~$90
ESD wrist strap + ESD mat. Protects every Jetson you touch. ~$25
Safety glasses + N95 mask for shell cutting. Plastic dust is bad to inhale. ~$15
Fire-safe LiPo bag. For 21700 packs. Cells are dangerous when mismanaged. ~$15

SOFTWARE / VIRTUAL INSTRUMENTS

KiCad 8+ (free) — schematic + PCB layout.
Fusion 360 (personal license free) or OnShape — internal frame CAD, button caps.
FreeCAD + KiCad StepUp — verify PCB fit inside the shell via STEP import.
JetPack SDK + SDK Manager — NVIDIA, free, required for Jetson flashing.
Docker + buildx for ARM cross-builds — already in your toolbox.
VS Code remote SSH — edit on the Jetson from your main machine.

05 // NOVITIATUS THE NOVICE'S PATH · SKILLS TO LEARN

▸

You already hold roughly seventy percent of the skill stack. Below is the gap list, ordered by when you'll need each one. Time estimates assume your background and your 5–10 hr/wk budget.

Already in your kit

Linux on embedded ARM — Unraid, systemd, ipmitool.
LLM quantization theory — INT4 + TurboQuant + MTP stabilized.
GPU/accelerator routing and PCIe debugging — your Localeksi POST code session.
Speculative decoding architecture — MTP draft + target, the same pattern reused here.
Docker, container networking, reverse proxies — OpenClaw / ClawdMac.
Front-end UI design — Y2K aesthetic, infinite canvas, IRON.AMP.

Gaps to close, in build order

1. KiCad schematic + 4-layer PCB layout ~20 HR

You need to design the carrier PCB. KiCad is free and dominant. Focus on: (a) reading the Jetson Orin NX carrier reference, (b) routing differential pairs (PCIe, MIPI-DSI, USB), (c) impedance-controlled traces on 4-layer JLCPCB stackup, (d) using KiCad StepUp to verify mechanical fit.

Best resource: Phil's Lab on YouTube — the "STM32 Hardware Design" and "PCB Design Crash Course" series. Free, dense, exactly the right level.

2. Mechanical CAD for retrofit ~15 HR

You'll model an internal frame that adapts the Game Gear shell to your PCB, fan, and battery. Fusion 360 personal is fine. Skills: sketching from caliper measurements, parametric extrusions, lofts for the battery cradle, STL export with proper tolerances for FDM (~0.2mm clearance).

Best resource: Lars Christensen on YouTube for Fusion 360; Maker's Muse for FDM-aware design rules.

3. Jetson platform · device tree + JetPack ~10 HR

The Jetson is not a Pi. Custom carriers usually require a custom device tree so the kernel knows your GPIO/I2C/SPI peripherals. NVIDIA's documentation is dense but accurate.

Best resources: NVIDIA's Jetson Linux Developer Guide; JetsonHacks YouTube channel; the RidgeRun wiki.

4. MIPI-DSI display bring-up ~10 HR

The display bus is the most fragile thing on the carrier. Learn DSI lane counts, panel init sequences, plumbing a panel driver in the kernel. Plan B: HDMI-to-MIPI bridge IC (Lontium LT8911EXB) and skip native DSI — easier, slightly more power.

Best resource: Toradex blog posts on DSI integration; NXP application notes on Lontium bridges.

5. Battery + BMS engineering ~6 HR

You're managing ~74Wh of lithium chemistry inside a sealed plastic shell. This is the one part of the build where a mistake is dangerous. Learn cell-matching, balance-charging, thermal cutoff design, BMS datasheet reading.

Best resource: Battery University; GreatScott! for the practical side; Big Clive for failure modes.

6. Edge LLM inference · NVIDIA-specific ~8 HR

You know vLLM. You don't (yet) know TensorRT-LLM for Jetson. That's the highest-throughput path on Orin NX. Alternative: MLC-LLM for ARM is mature and easier. Run both, pick the one that hits the 30–50 t/s bar.

Best resources: NVIDIA's Jetson AI Lab tutorials; MLC-LLM docs; llama.cpp Jetson build guide.

7. Whisper.cpp / on-device STT ~3 HR

Trivial once you've done Jetson bring-up. whisper.cpp on Jetson with CUDA is well-documented and fast.

8. Mechanical finishing & documentation ~5 HR

Least technical, most overlooked. Plan to learn clean Dremel cuts in ABS, hiding seams, and photographing the build under controlled light. The polish is what makes it post-worthy.

06 // HORAE CANONICAE THE CANONICAL HOURS · PHASED PLAN

▸

Your time budget is 5–10 hr/wk. This plan front-loads the cheap, low-risk learning so expensive silicon arrives exactly when you're ready for it. Total elapsed wall-clock: ~6–9 months. Each hour is a rite.

MATINS
WK 1–4

FOUNDATION & ANNOUNCEMENTHora prima · the first hour

Buy: caliper, soldering station if missing, KiCad install, Fusion 360 personal.
Acquire a donor Game Gear (eBay, $40–80, dead screen is fine) for measurements + button salvage.
Acquire two reproduction shells from RetroSix or Castlemania.
Disassemble the donor. Photograph every step in good light. This is your day-one X content.
Measure the internal cavity; draw a 2D CAD baseline.
Announce project on X with the disassembly photos + one-paragraph thesis.

LAUDS
WK 5–10

BREADBOARD PROTOTYPEHora secunda · the brain awakens

Buy: Jetson Orin Nano development board ($249) as a cheap proxy for software work — do NOT buy NX 16GB yet.
Flash JetPack. Boot. Confirm CUDA + GPU.
Install MLC-LLM. Benchmark Qwen2.5-Coder 1.5B / 3B locally. Confirm tok/s targets achievable.
Wire breadboard: 4 buttons → GPIO, INMP441 mic → I2S, MAX98357 + speaker → I2S, MIPI-DSI display via dev carrier.
Write the first version of the local oracle loop in Python: button → menu → LLM → tool call → screen render.
Mid-phase: post a video of the breadboard talking — "the brain that's going inside the Game Gear."

TERCE
WK 11–18

PCB DESIGN + MECHANICAL CADHora tertia · drawing the vessel

Order Jetson Orin NX 16GB module.
Design the carrier in KiCad. Reference NVIDIA Orin NX carrier. 4-layer JLCPCB, ~150×80mm, SODIMM-260, fan-out for: display FPC, button matrix, USB-C, mic/speaker I2S, M.2 SSD, M.2 wifi, 5V/3.3V regulation, BMS interface.
Run DRC. Full STEP export → Fusion 360 → confirm mechanical fit inside the shell.
Design 3D-printed internal frame: PCB cradle, fan duct, battery cradle, display mount, button standoffs.
Print v1 of the frame in PETG. Iterate on tolerance.
Order PCBs (5pcs) with JLCPCB assembly for bottom-side SMD. Hand-place anything sensitive.

SEXT
WK 19–24

FIRST INTEGRATIONHora sexta · the spark

PCB arrives. Bring-up sequence: power rails first (3.3V, 5V on bench, current-limited), then SODIMM, then storage, then display, then peripherals. Never plug a Jetson into an unverified board.
Build the custom device tree for your carrier. Boot to a serial console first.
Wire up the screen. Confirm framebuffer renders.
Mount everything in the printed frame. Test fit inside reproduction shell without screwing it shut — confirm fan clearance, port alignment, button registration.
Cut vents and USB-C window in the back shell. Slow Dremel pass, multiple test fits.
End-of-phase video: "first boot in the shell."

NONE
WK 25–30

SOFTWARE STACK + ORACLE LOOPHora nona · the voice is given

Port the oracle loop from Phase 1 onto the in-shell device.
TUI renderer: header (task), body (scrollable diff/log), footer (button hints). Use ratatui or a custom framebuffer renderer.
Wire whisper.cpp tiny for voice prompts. Push-and-hold on D-pad center.
Implement tool-call approval surface: A=approve, B=deny, L=defer, R=show full diff.
Tailscale install. Build escalation path: local model returns confidence + size estimate; if either threshold tripped, route to Localeksi.
Power profiling under sustained load. Validate runtime estimates. Add fan PWM curve.

VESPERS
WK 31–34

POLISH + V1 LAUNCHHora vespertina · the unveiling

Move from PETG prototype frame to final-quality print (ABS or resin).
Final shell finishing: vent grilles, button cap fit, USB-C window deburred.
Boot logo, splash screen, on-shell branding.
Record the launch video: cold boot → button press → real coding task → completion → power off.
Publish: thread on X, video on YouTube, full write-up on personal site with all schematics + STLs + dockerfiles open-sourced.

COMPLINE
BEYOND

FUTURE RITESHora completa · do not start early

Jetson Thor Nano swap if available.
Atari Lynx and Nomad shell variants (same architecture, different chassis).
Cellular modem add-on (LTE-M) for true off-grid mode.
Small-batch kit for selling to other hobbyists.

07 // LITURGIA CODEX LITURGY OF THE CODE · SOFTWARE STACK

▸

Stack overview

LAYER	PICK	NOTES
OS	Ubuntu 22.04 (JetPack 6.x)	NVIDIA-provided. Don't fight it.
Inference runtime	MLC-LLM (primary) or TensorRT-LLM	llama.cpp as the fallback safety net.
Target model	Qwen2.5-Coder 3B Q4 (or Qwen3-Coder when stable on Jetson)	Local everything-default.
Draft model	Qwen2.5 0.5B Q4	For speculative decoding — same MTP pattern as Localeksi.
STT	whisper.cpp tiny.en (CUDA)	Real-time on Orin NX.
Agent framework	Custom Python loop, or fork of aichat / aider	Don't run Claude Code locally — write a Claude-Code-shaped agent that calls the local model.
UI	Ratatui (Rust) or framebuffer + Pygame	Ratatui for the terminal aesthetic.
Wireless tunnel	Tailscale	Free for personal use, drops onto your existing tailnet.
Persistence	SQLite + JSONL session logs	Push to GitHub on wifi.

The oracle loop · invocation

# the four-button rite, simplified
while True:
    intent = await input_layer.next_intent()  # button, voice, or remote
    plan = local_model.plan(intent, context=session_ctx)

    if plan.confidence < 0.6 or plan.scope > LOCAL_BUDGET:
        plan = await remote.escalate(plan, target="localeksi")

    for step in plan.steps:
        if step.destructive:
            await ui.show_diff(step.preview)
            if not await input_layer.await_approval():
                continue
        result = await tools.execute(step)
        ui.append_log(result)

    session_ctx.commit()

The escalation contract

Local model and remote (Localeksi vLLM) speak the same OpenAI-compatible API. The escalator is a tiny router that sets base_url and adds a Tailscale auth header. Your existing OpenClaw proxy on ClawdMac is the closest equivalent and can be reused with minor changes.

The Node.js XML-to-JSON tool-call proxy you wrote for Qwen tool calls is directly portable. Drop it on the Jetson and it converts Qwen XML responses to OpenAI JSON for the local oracle loop — same MASTER PATTERN extraction strategy.

08 // RECEPTACULUM THE VESSEL · SHELL + COOLING

▸

Shell strategy

The original Sega shell is fragile, yellowed, and irreplaceable in good condition. The right approach:

One donor Game Gear for measurements and salvage (buttons, screen lens).
Two reproduction shells from RetroSix (UK) or Castlemania Games (US). $25–35 each. Black, clear, frosted, pastel.
Use shell #1 as cut-up prototype. Shell #2 is v1 final.
Don't touch any original 1990 shell unless you decide to build a "patina" variant in v2.

DROP-IN · side-by-side photo of donor shell + reproduction shells, calipers visibleshoot this in Phase 0 — your launch image

Required shell modifications

Vent grille on rear shell, ~30mm × 60mm aligned with the fan blower's intake side. Dremel, deburr, line with a 3D-printed grille insert (cosmetic + dust filter).
USB-C window on the top edge, near the original AC adapter cutout. Use the existing opening if possible.
Optional HDMI debug port on the side. Block off if not used for launch.
Screen window — original is ~80×60mm. Your new IPS panel may be smaller; print a custom bezel insert to fill the gap and hide PCB edges.
Speaker grille — keep original. Mount MAX98357 + 8Ω driver behind it.

Internal frame · 3D-printed

A single PETG or ABS print that sits inside the rear shell and locates: carrier PCB, 4× 21700 pack, fan + duct, display, button FPC routing. Designed in Fusion. Held by original shell screw posts plus four added M2 heat-set inserts.

Cooling math

The Orin NX 16GB dissipates up to ~25W under sustained inference. Inside a 32mm-deep sealed cavity, that's a thermal failure in <5 minutes without active airflow.

Heat path: SoC → copper shim → graphite sheet bonded to rear shell interior. Spreads heat across ~80cm² of plastic.
Active airflow: 30mm 5V blower fan, PWM-controlled by the Jetson based on thermal sensor. Off below 55°C, ramping linearly to 100% at 80°C.
Intake: rear vent grille (cold air across PCB top side, exits through fan).
Exhaust: fan blows out through original speaker grille area with an air gap routed around the speaker, or through a second top-edge slot.
Validation: thermal camera under 10-min sustained 100% load. Target: SoC junction <85°C, external shell <42°C (uncomfortable but not dangerous).

Do not seal the shell for the first thermal test. Run with the back off, thermal camera on the SoC, and a hand near the kill switch. Cooling failure inside a sealed shell with 74Wh of lithium nearby is the worst-case scenario for this build.

09 // IGNIS SACER THE SACRED FLAME · POWER SYSTEM

▸

Pack design

4× 21700 Li-ion cells in 2S2P (two series, two parallel) → 7.4V nominal, ~10Ah, ~74Wh. Buck-converted to 5V for the Jetson rail and 3.3V for peripherals. At ~15W sustained, expect ~4–5 hours active inference, ~24 hours intermittent, days of standby.

Cell choice	Molicel P42A or Samsung 50G
Configuration	2S2P · 7.4V nominal
Capacity	~10Ah · ~74Wh
BMS	2S 20A continuous · balance leads · over-temp cutoff
Fuel gauge	MAX17260 over I2C (real % readout)
Charging	USB-C PD 20V/3A → 4.2V/cell CC-CV via dedicated charger IC

Power tree

Charging behavior

Default input: USB-C PD 20V/3A wall brick (60W headroom).
Charges from empty in ~90 minutes; trickle-tops in 30.
Pass-through power: device can run on wall + charge simultaneously.
Cell balancing on every charge. Over-temp trips at 50°C cell temp.

Build the pack in a fire-safe LiPo bag on a non-flammable surface. Never spot-weld in the same room as your main computer. Test the BMS with a dummy load before installing it inside the shell.

10 // FENESTRA THE WINDOW · DISPLAY

▸

The original Game Gear screen window is approximately 80mm wide × 60mm tall. Three viable options, ranked by aesthetic-vs-effort:

OPTION A · 3.5" IPS via MIPI-DSI RECOMMENDED

Resolution 640×480 or 720×480, 4-lane DSI.
Pros: highest resolution, modern color, native Jetson support.
Cons: DSI bring-up is the riskiest software task in the build. Mitigate with a known-good Waveshare panel that has public device-tree examples.

OPTION B · HDMI panel with mini-HDMI input EASIEST

3.5" 800×480 HDMI-input panel (Adafruit, Waveshare). Driver board attaches.
Pros: dead-simple bring-up, just an HDMI cable from Jetson.
Cons: thicker stack (panel + driver board adds 8–12mm). Tight inside the shell.

OPTION C · AMOLED via SPI AESTHETIC

1.91" or 2.4" RM67162 OLED. Smaller than the original window — fill the rest with a printed bezel insert.
Pros: true blacks (Y2K chaos aesthetic dream), low power, vibrant.
Cons: small. Limits how much TUI fits at once. Mitigate with "press R to expand" detail view.

Start with Option B (HDMI) for the breadboard phase to de-risk software. Switch to Option A (MIPI-DSI) for the final carrier board. Reserve Option C for a v2 "OLED EDITION."

UI rendering approach

The screen is portrait-ish (80×60mm, original aspect ~4:3). Render a TUI with three regions:

HEADER (10%) — current task name, model in use, wifi/Tailscale status, battery %.
BODY (75%) — scrollable log, diffs highlighted in phosphor green (additions) and crimson (deletions).
FOOTER (15%) — button hints, always visible: A=approve · B=deny · L=defer · R=detail · ↕=scroll · Start=new.

11 // VOX & MANUS VOICE & HAND · INPUT & AUDIO

▸

Button mapping · final liturgy

BUTTON	PRIMARY	HELD	NOTES
A (Button 1)	Approve	Approve + remember rule	"Allow this kind of action without asking next time."
B (Button 2)	Deny / Cancel	Deny + add to blocklist	Stops the oracle.
Start	New session / prompt	(held = voice)	Press-and-hold while speaking.
D-pad ↑↓	Scroll log	—	Free movement.
D-pad ←→	Switch view (log ↔ diff ↔ detail)	—	Like browser tabs.
D-pad center (added)	Defer / next pending	Show queue	Modern shells allow adding a center button.
Power switch	On / Off with confirm	—	Mechanical slide preserved from the original look.

VOX · the microphone

INMP441 I2S MEMS mic mounted behind a small grille hole (drilled in the shell, discreet). Whisper-tiny on the Jetson handles STT in real-time. Voice activation is push-to-talk on Start — never always-listening (privacy + battery).

VOCEM REDDIT · the speaker

Salvage the original 8Ω paper-cone driver from your donor; it's perfectly sized. MAX98357 I2S amp drives it. Uses: TTS readback of long results, audio status cues (boot chime, approval confirmation tone), and silly Easter eggs.

12 // OMINA OMENS & WARNINGS · RISK REGISTER

▸

OMEN	PROB	IMPACT	WARDING
Thermal runaway in sealed shell	Med	Catastrophic	Active fan; thermal camera validation in Phase 3; never seal for first 10hr cumulative test.
MIPI-DSI bring-up fails on custom carrier	High	High	Use HDMI panel for breadboard; ship v1 with HDMI if DSI not ready.
Battery pack fails / cell mismatch	Low	Catastrophic	Matched cells from reputable seller (18650batterystore, illumn); cycle-test before integration; BMS with thermal cutoff.
Carrier PCB has a fatal trace error	Med	Med ($40 lost)	Order 5 boards. Inspect under microscope before populating. Power up rails individually on bench supply.
Shell cuts ruin the prototype shell	Med	Low ($30)	Buy 2 reproduction shells. Plan cuts in CAD first; transfer with template.
Local model can't actually do useful work	Med	High	Validate in Phase 1 on dev kit before committing $600 to NX. If 3B model is too dumb, escalation tier saves the project.
Jetson supply chain hiccup	Low	Med	Order through SeeedStudio + a US distributor as backup. Not constrained as of mid-2026.
Scope creep into v2 features	High	High	Hard rule: nothing labeled "v2" enters the codex until v1 has booted and been posted publicly.
Time budget collapses (work + family)	Med	Med	Wall-clock plan assumes 5–10hr/wk. If a week is missed, slide; don't make it up by sacrificing family time.

13 // EVANGELIUM THE SPREADING · CONTENT STRATEGY

▸

This project is content-shaped. Build in private and post once at the end, and you lose 90% of the reach. Document every phase as it happens and turn the build itself into a recurring cadence. The Phoenix algorithm rewards replies and dwell time — build threads, not single tweets.

X / Twitter cadence

Day 1 (Phase 0 launch): a 6-tweet thread. Hero photo: donor Game Gear + reproduction shells, calipers visible. Headline: "I'm rebuilding a Game Gear into a local LLM coding oracle. Targeting 30-50 tok/s. 4-button UI. Offline-first. Build log starts now."
Per phase: one anchor thread when a milestone hits (breadboard talks, PCB arrives, first boot in shell). Each thread: 4–8 tweets with photos.
Weekly: a small "shop note" — one photo, one observation, one open question. Asking questions juices replies, which juice algorithm reach 75×.
Reply-mode: spend 2× the time replying to people who reply to you than you spend posting. This is your authority loop.

YouTube approach

Three videos. Not more, not less.
VIDEO 1 · "Why I'm putting an AI inside a Game Gear" (Phase 0): 4–6 minutes. The thesis, the donor disassembly, shells on the bench. Hook: handheld + AI + retro is irresistible.
VIDEO 2 · "The brain works (breadboard)" (end of Phase 1): 6–8 minutes. Live demo of breadboard talking, model running, button mapping working. The skeptic-converter.
VIDEO 3 · "It boots." (end of Phase 5): 8–12 minutes. The full v1 launch. Cold boot → real coding task → close shut → walk outside → still working. The one that gets shared.

Personal site (evanhawkins.com or timechaintech.com)

Dedicated build-log page. All phases, photos, schematics, STLs, dockerfiles open-sourced. Y2K chaos.
RSS so the technical audience can follow without algorithm intermediation.
One long write-up at the end (~3000 words). The canonical reference. What gets cited on Hacker News.

The project doubles as Timechain Technologies portfolio content. "Sovereign compute in a handheld form factor" is exactly the narrative the federal contracting market is buying right now. Show this to a sovereign-AI-curious procurement officer and you're not pitching anymore — you're demonstrating.

Scribes of the build (sample guestbook)

xX_SILICON_MONK_Xx

this is the most cursed bishop i've ever seen 10/10

カオスマシナ

finally a handheld worth carrying 🔥

VOID_PROCURATOR

i need this in my pocket immediately

CRTBYZANTINE

oraculum sacci unironically based

14 // BIBLIOTHECA THE LIBRARY · LEARNING REFS

▸

PCB & HARDWARE

Phil's Lab — PCB Design Crash Course (YouTube)
NVIDIA Jetson Orin NX Carrier Board Reference Design (NVIDIA docs)
JLCPCB capabilities + stackup reference (jlcpcb.com)
EEVblog — fundamentals + tear-downs (YouTube)
Texas Instruments WEBENCH Power Designer (for buck regulator sizing)

JETSON / EDGE AI

NVIDIA Jetson AI Lab (jetson-ai-lab.com) — tutorials directly relevant
JetsonHacks (YouTube + blog)
RidgeRun wiki — device tree, MIPI bring-up
MLC-LLM docs (mlc.ai/mlc-llm)
TensorRT-LLM Jetson examples (NVIDIA GitHub)

MECHANICAL / CAD

Lars Christensen — Fusion 360 tutorials (YouTube)
Maker's Muse — FDM design rules (YouTube)
KiCad StepUp documentation
RetroSix shell store + spec sheets

BATTERY / POWER

Battery University — articles by Cadex (batteryuniversity.com)
GreatScott! — practical builds (YouTube)
Big Clive — failure modes (YouTube)
TI BQ25798 datasheet + reference design

LLM / AGENTS

Anthropic — Building Effective Agents (anthropic.com)
aider GitHub repo (paul-gauthier/aider)
Qwen2.5-Coder model card (Hugging Face)
Speculative decoding paper (Leviathan et al., 2023)

CONTENT / NARRATIVE

Marques Brownlee's framing structure for tech-object videos
Bunnie Huang's Hacking the Xbox and Adafruit's project write-ups — gold standard for build-log voice
The Phoenix algorithm notes you compiled in January 2026 (your own doc)