Leopoldo Rodríguez
449588649a
Leo's post-C12 DRC run left exactly one error: [unconnected_items] Zone 'GND_POUR_FCU' [GND] on F.Cu <-> Zone 'GND_POUR_FCU' [GND] on F.Cu KiCad prints only the zone anchor (60.1917, 40.3231) for both endpoints, so the report cannot localise the stranded island. Located it analytically instead. Root cause: U11 pad 9 [GND] @ (91.1275, 55.7600). It is the only GND item on the board whose item-only connectivity cluster never reaches the In2.Cu GND plane (two independent fill-independent proofs agree: layer-reach and In2-anchor). An island survives KiCad's isolated-island removal only if some item props it up, so pad 9 was propping up a GND_POUR_FCU island that touches nothing else. island_removal_mode cannot fix that - it needs copper. Pad 9 is fenced on all four sides (LQFP-48 lead ring, pad 8's westward /3V3_ETH escape, pad 10's eastward /ETH_RSET escape, the /ETH_RXP chain under the body). A 0.05mm grid router reached only 227 of 20883 legal cells and found no way to touch existing GND copper without a via, so a through via is mandatory. A 0.005mm sweep of every legal via site in that pocket picked the best one on the board's existing drill inventory - no third drill size added. + F.Cu GND track w0.3 (91.1275,55.7600) -> (90.1000,55.7600) -> (89.9950,55.8900) + F.Cu-B.Cu GND via 0.6/0.3 @ (89.9950,55.8900) * U11 pad 9 -> (zone_connect 2) Both track legs hold +0.2000mm clearance - the structural ceiling for a 0.3mm track between 0.3mm leads on 0.5mm pitch. The via holds +0.1758mm copper and +2.9705mm hole. The site sits inside GND_POUR_FCU, GND_POUR_BCU, P3V3_PLANE_IN1 and GND_PLANE_IN2 and outside MAGJACK_NO_POUR, with the nearest In2 signal copper 1.10mm away, so the barrel lands on solid In2 GND plane. zone_connect 2 mirrors the landed C12 treatment of pad 29: pads 8 and 10 each knock 0.5mm of pour out around themselves, covering the whole pad-9 lane (0.10mm tall < min_thickness 0.25), so the pour cannot reach pad 9 in any connect mode. Solid mode removes the pad from the minimum-spoke test rather than leaving a starved thermal behind. Verified on the output file: drc_c12.py 0 hard clearance violations, same 49 sub-epsilon marginals as C12 proof_c13.py 0 GND clusters without an In2 GND-plane anchor (was 1) fcu_only.py 0 F.Cu-only GND clusters (was 1 = pad 9) 24V / /24V_IN / IGN_IN copper byte-identical (45 segments) segments 1984->1986 vias 351->352 zones 5->5 footprints 189->189 Backup: .mcp-backups/hbox-watchdog-v4.kicad_pcb.pre-C13 (md5 042efd5d) Pushed: md5 94d36711d60985eef109913fcee36b09, 3502885 bytes Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GSKKes8GsS4YHAMWJ44tCm |
||
|---|---|---|
| _brand | ||
| .claude/skills/project-drogheda | ||
| 00-Project-Management | ||
| 01-Requirements | ||
| 02-Architecture-Software | ||
| 03-Architecture-Hardware | ||
| 04-Source-Code | ||
| 05-Analytics-Reports | ||
| 06-Testing-Validation | ||
| 99-Archive | ||
| Marketing | ||
| mcpmarket-me | ||
| scripts | ||
| .gitignore | ||
| .gitmodules | ||
| CLAUDE.md | ||
| mcpmarket-plugin-me.zip | ||
| README.md | ||
Project Drogheda (GPS RTK)
Helgen Technologies' HBOX RTK GPS tracker fleet for Hutchison Ports heavy
equipment (ICAVE / TIMSA / EIT / LCT) — centimetre-class position over 4G, streamed to the
Helgen backend (retranslator → Wialon/Traccar → VERONICA), built to survive a hot, salty,
high-vibration port environment.
New here? Read
00-Project-Management/Engineering-Handover.mdfirst. It is the deep context: product, failure analysis, fixes, repo map, open items, glossary. This README is the 2-minute orientation;CLAUDE.mdis the terse machine index.
Where we are (2026-06-20 · updated 2026-07-02)
2026-07-02: backend deployed + stable; thermal chamber run done; 3-board protoboard campaign in flight (units to ICAVE 7-jul). Watchdog v3 PROD schematic is complete (ERC 0) — LTC4040 battery-backup sized to the 12 Ah pack, fail-safe Modo-B relay bypass — and the PCB is now a 120×100 carrier board: Blackpill + simpleRTK2B Pro (UNO-shield sockets) + 4G NTRIP Master (on-board XBee socket) all on female headers. Layout in placement review; fab gate = proto bench validation. Canonical:
03-Architecture-Hardware/watchdog-electrical-design-state.md.
We deployed a first batch of HBOX trackers on ICAVE machines in Veracruz. They worked "ish," then failed. We are mid-crisis, hardening for scale — target 30 units by August. The failure is three independent fronts, fixed in parallel:
- Backend (confirmed in code) —
retranslator.pyhead-of-line blocks the whole fleet on slow upstream, and an anti-jump filter can permanently black-hole a healthy unit. - Thermal — an ASA case (UV-stable, ~95–100 °C — adequate) sealed in a sealed IP67 cabinet that also holds the heat-generating Mean Well PSU, in full sun → nested heat trap for the battery / boost / modem (not the case).
- Prototype hardware — Dupont + hand-wiring + 18650/PowerBoost; fragile under vibration.
Units: all pulled — 4 in hand (3 dead, 1 working-ish) in Veracruz; team in Querétaro → remote debug. Field firmware is v1 (no thermal log).
Free triage: a unit ONLINE in htracker but OFFLINE in Wialon is a backend victim, not dead hardware. Check that before condemning a radio.
The product in one breath
Rover (HBOX) on the machine + Base Station (Ardusimple AS-PRO-CORS, 110/230 V) publishing RTCM3 corrections over NTRIP. The rover gets corrections over the 4G modem, not the GNSS → recover a dead link by power-cycling the 4G stack, never a GNSS reset. HBOX core: simpleRTK2B Pro (ZED-F9P) · 4G NTRIP Master · STM32F411 watchdog · Mean Well SD-15B-05 (24→5 V) · PowerBoost + 18650 · GX16-5 to the cabinet. 24 V is the production variant.
Invariants (never violate)
- Power-off is the safe state. 2. Never vanish, never loop. 3. Corrections ride the 4G modem, not the GNSS. 4. No material that softens below ~80 °C (PLA-class) in the thermal path — the ASA enclosure complies; design to measured cabinet temperature. 5. Backend forwards without head-of-line blocking; anti-jump never locks out a healthy unit. 6. No secrets in the repo (the committed MSSQL password is a live violation to rotate).
Repo layout
00-Project-Management (plan, ethos, financials, handover, Drive tooling) ·
01-Requirements (SRS, product spec, tracker) · 02-Architecture-Software (backend review) ·
03-Architecture-Hardware (cabinet/HBOX, BOM, PCB brief, watchdog, manual) ·
04-Source-Code (firmware/ · htracker/ · notifications/) ·
05-Analytics-Reports (SQL views — placeholder) · 06-Testing-Validation (RCA, diagnostics) ·
Marketing (datasheet, brochure) · _brand (LaTeX style) · 99-Archive (raw inputs).
Quick start
# Branded docs (XeLaTeX; League Spartan + MADE Evolve Sans):
cd 01-Requirements/HBOX-Ficha-Tecnica && latexmk -xelatex main.tex
# Firmware (own repo 'hbox-firmware', at 04-Source-Code/firmware):
# open blackpillrtkcode_v2/blackpillrtkcode_v2.ino in the STM32duino toolchain.
# ⚠ Bench-test before any field flash — the sandbox cannot build STM32.
# Backend: 04-Source-Code/htracker (ssh://git.helgen.dev:6611/htracker, feature/test500unidades)
Storage model (one home per artifact)
This repo = working source of truth (editable .md/.tex, code). Google Drive
Drogheda (GPS RTK) = finalized exports (branded PDFs, native Docs/Sheets). Second Brain
vault = distilled decisions/lessons. Author in repo → export final to Drive → distill to
vault. Drive upload how-to + mapping: 00-Project-Management/Drive-Upload/README.md. Never
change Drive sharing — that's Leo's call.
Go deeper
- Deep context / onboarding →
00-Project-Management/Engineering-Handover.md - Current priorities →
00-Project-Management/PLAN.md - Machine index →
CLAUDE.md - Failure analysis →
06-Testing-Validation/RCA-2026-06-field-failures.md