Skip to the content
2 layers60 × 40 mmMax3485-rs485, the printed circuit board in 3D

Max3485-rs485

embedrapp/deepee2

Open benchmark for evaluating AI agents on embedded firmware, code repair, schematics, and PCB engineering

Size
60 × 40 mm
Copper layers
2
Thickness
1.6 mm
Parts
7
Nets
8
Pads
25
Tracks, vias
8, 0
Routed
Yes
KiCad
KiCad 10.0 (file 20240108)
Created
17 Aug 2026
Updated
22 Oct 2026

Overview

Max3485-rs485 is an open-source printed circuit board by embedrapp, from the embedrapp/deepee repository on GitHub, published under the MIT licence. It is a 2-layer board of 60 × 40 mm with 7 parts on 8 nets, routed with 8 tracks and 0 vias.

It is built around the MAX3485ED. Its parts include 3 resistors, 2 connectors, 1 capacitor and 1 IC. It belongs with the interfaces and adapters boards of this library.

Main components

1

README

README.md

DeepEE

DeepEE is a small benchmark for AI agents doing embedded firmware and KiCad engineering. An agent receives a prompt plus starter files, edits a fresh working directory, and submits ordinary source files. An offline verifier checks the result. Conversation quality and screenshots do not count.

Scope

Version 1.2 has 48 tasks: 12 in each engineering category.

CategoryTask IDsTopicsPass gate
Firmwaredeepee-fw-001–012I²C recovery; SHT3x, BME280, MCP9808, INA219, BMP280, ADS1115, DS18B20, SCD4x, PCA9685, and MAX31855 drivers; Modbus CRCProtected API + native C tests + PlatformIO build
Code repairdeepee-repair-001–012Scheduler rollover, ring buffers, ADC scaling, SLIP, MQTT, UTF-8, CBOR, PPP FCS, HTTP chunks, Base64, Modbus CRC, and wrap-safe tick timingProtected API + native C tests
Schematicsdeepee-sch-001–012Power-input repair, MCP9808, INA219, ADS1115, BMP280, AP2112, USB-C sink, RS-485, CAN, I²C level shifting, microSD, and NE555Exact component/pin-net contract + exported netlist + ERC
PCBdeepee-pcb-001–012nRF24L01, MCP9808, and the same ten researched interface circuits used by the schematic setExact footprint/pad/outline contract + fully routed nets + DRC

The tasks are compact adaptations of real bug reports, open-source boards, official KiCad QA material, and published breakout use cases. Provenance is recorded in each task's SOURCE.md and summarized in docs/TASK_SOURCES.md.

The evaluation contract and its limits are documented in docs/VALIDATION_MODEL.md. Contributions must follow CONTRIBUTING.md, including the known-good and mutant release gates.

What is verified

Every task declares public, machine-readable critical requirements. A task score remains binary:

text
all checks pass -> 1
anything fails  -> 0

The primary aggregate is Pass@1: the fraction of tasks passed on an agent's first publishable attempt. Score files also contain each requirement outcome, its engineering layer, leaf evidence, a layer-level score vector, and a stable decision hash. A strong result in one layer cannot mask a failed critical requirement in another.

The verifier checks required parts and pin/pad nets without comparing an entire design to golden geometry. KiCad performs the final ERC/DRC gate. PCB contracts additionally measure declared physical constraints directly from native copper: all named nets use at least 0.25 mm track width, and the USB-C repair constrains routed D+/D− skew to 1.0 mm. The release gate runs a targeted mutant for every critical requirement and must reject all of them.

This is deliberately a KiCad-native benchmark, not a tool-agnostic EDA benchmark. An AI system may be compared if it can produce the required native KiCad 10 artifacts. Altium- or EasyEDA-only outputs are out of scope because silently converting them would weaken the assurance and make the comparison misleading.

Firmware is verified by native behavioral tests and a real target build. Version 1.2 does not claim runtime peripheral, analog, thermal, EMI, or fabrication signoff; those require suitable simulation or hardware-in-the-loop fixtures and are excluded from task claims.

Lightweight local preparation

No VM, Colima instance, Docker image, KiCad install, or PlatformIO install is needed on the development laptop. The repository checks use Python and a host C compiler only:

bash
python3 -m venv .venv
. .venv/bin/activate
pip install -e .
make test

The Dockerfiles and scripts are deployment material for a later x86_64 Ubuntu EC2 run. They are not started by the commands above.

EC2 execution

On an Ubuntu 24.04 x86_64 host:

bash
./scripts/bootstrap-ec2-ubuntu.sh
# Log out and back in once for Docker group membership.
export CODEX_API_KEY='...'
./scripts/run-ec2-baseline.sh --all-tasks

For a ChatGPT subscription-authenticated Codex run, complete codex login --device-auth on the host and use ./scripts/run-ec2-chatgpt-baseline.sh instead.

The deployment builds separate deepee-agent:1.2.0 and deepee-verifier:1.2.0 images. The agent image has KiCad 10.0.4, PlatformIO 6.1.19, compilers, and Codex but no benchmark tasks or verifier code. Agent runs use an internal-only Docker network and one explicit CONNECT-proxy profile: API-key runs permit api.openai.com:443, while ChatGPT-subscription runs additionally permit auth.openai.com:443 and chatgpt.com:443. Both profiles block every other authority and direct internet egress, and the effective profile and allowlist are recorded in the run. Verification runs without any network or agent credentials. Before a baseline starts, the Ubuntu host also proves one private reference submission for every task against that same verifier image. See docs/EC2.md and docs/RUN_PROTOCOL.md.

For a manual external agent that uses KiCad:

bash
deepee benchmark \
  --agent agents/manual-kicad.yaml \
  --task deepee-pcb-001 \
  --prepare-only

deepee verify \
  --container-image deepee-verifier:1.2.0 \
  --task deepee-pcb-001 \
  --run-dir '<printed-run-directory>' \
  --output results/scored/external-pcb-run.json

Manual/external runs record their authoring egress as external_uncontrolled; DeepEE still applies the same offline verifier and never mislabels that external environment as proxy-isolated.

DeepEE's original code and fixtures are MIT licensed. The KiCad-derived schematic starter identified in THIRD_PARTY_NOTICES.md retains its GPLv3 license. Compare leaderboard results only when the benchmark and task hashes match.

Parts

7
QtyValueFootprintReferences
2680RResistor_SMD:R_0603_1608MetricR2, R3
1100nFCapacitor_SMD:C_0603_1608MetricC1
1120RResistor_SMD:R_0603_1608MetricR1
1Conn_01x03TerminalBlock_RND:TerminalBlock_RND_205-00288_1x03_P5.08mm_HorizontalJ2
1Conn_01x06Connector_PinHeader_2.54mm:PinHeader_1x06_P2.54mm_VerticalJ1
1MAX3485Package_SO:SOIC-8_3.9x4.9mm_P1.27mmU1

More from embedrapp/deepee

11
  • 2 layers60 × 40 mmPca9306-i2c level shifter, the printed circuit board in 3D

    Pca9306-i2c level shifter

    embedrapp/deepee

    Open benchmark for evaluating AI agents on embedded firmware, code repair, schematics, and PCB engineering

    10 parts2MIT
  • 2 layers60 × 40 mmBmp280-environment sensor, the printed circuit board in 3D

    Bmp280-environment sensor

    embedrapp/deepee

    Open benchmark for evaluating AI agents on embedded firmware, code repair, schematics, and PCB engineering

    7 parts2MIT
  • 2 layers60 × 40 mmMicrosd spi adapter, the printed circuit board in 3D

    Microsd spi adapter

    embedrapp/deepee

    Open benchmark for evaluating AI agents on embedded firmware, code repair, schematics, and PCB engineering

    7 parts2MIT
  • 2 layers60 × 40 mmNe555-astable, the printed circuit board in 3D

    Ne555-astable

    embedrapp/deepee

    Open benchmark for evaluating AI agents on embedded firmware, code repair, schematics, and PCB engineering

    7 parts2MIT
  • 2 layers60 × 40 mmIna219-current monitor, the printed circuit board in 3D

    Ina219-current monitor

    embedrapp/deepee

    Open benchmark for evaluating AI agents on embedded firmware, code repair, schematics, and PCB engineering

    6 parts2MIT
  • 2 layers60 × 40 mmAds1115-analog input, the printed circuit board in 3D

    Ads1115-analog input

    embedrapp/deepee

    Open benchmark for evaluating AI agents on embedded firmware, code repair, schematics, and PCB engineering

    6 parts2MIT
  • 2 layers60 × 40 mmUsb c-5v sink, the printed circuit board in 3D

    Usb c-5v sink

    embedrapp/deepee

    Open benchmark for evaluating AI agents on embedded firmware, code repair, schematics, and PCB engineering

    6 parts2MIT
  • 2 layers18 × 15 mmMcp9808-breakout, the printed circuit board in 3D

    Mcp9808-breakout

    embedrapp/deepee

    Open benchmark for evaluating AI agents on embedded firmware, code repair, schematics, and PCB engineering

    6 parts2MIT
  • 2 layers60 × 40 mmSn65hvd230-can, the printed circuit board in 3D

    Sn65hvd230-can

    embedrapp/deepee

    Open benchmark for evaluating AI agents on embedded firmware, code repair, schematics, and PCB engineering

    6 parts2MIT
  • 2 layers60 × 40 mmAp2112-3v3-regulator, the printed circuit board in 3D

    Ap2112-3v3-regulator

    embedrapp/deepee

    Open benchmark for evaluating AI agents on embedded firmware, code repair, schematics, and PCB engineering

    5 parts2MIT
  • 2 layers30 × 25 mmNrf24-adapter, the printed circuit board in 3D

    Nrf24-adapter

    embedrapp/deepee

    Open benchmark for evaluating AI agents on embedded firmware, code repair, schematics, and PCB engineering

    4 parts2MIT

Similar boards

  • 2 layers100 × 100 mmExplorer, the printed circuit board in 3D

    Explorer

    anyshake/explorer

    🌏 Detecting seismic wave using 3 geophones and 3-axis accelerometer, pack & send data to AnyShake Observer over RS232 / RS485 serial.

    266 parts100
  • 2 layers120 × 80 mmRx72t 100 pga usb, the printed circuit board in 3D

    Rx72t 100 pga usb

    hirakuni45/RX

    Renesas RX Microcontroller, C++ framework, Library, Sample

    113 parts86BSD-3-Clause
  • 4 layers106 × 63.5 mmBrineomatic, the printed circuit board in 3D

    Brineomatic

    hoeken/brineomatic

    Open Source Watermaker Controller.

    323 parts47LGPL-2.1
  • 4 layers100 × 75 mmOpenAirScope, the printed circuit board in 3D

    OpenAirScope

    emertcakir/OpenAirScope

    Open-source air quality and environmental monitoring node with STM32H743 + ESP32-C3, LoRaWAN, RS485, USB-C, Li-ion charging, and multi-sensor support.

    282 parts26MIT
  • 4 layers80 × 60 mm01 main, the printed circuit board in 3D

    01 main

    wgbowley/OpenLSM

    Low-Cost Linear Synchronous Permanent-Magnet Motor

    99 parts11MIT
  • 2 layers92 × 51 mmRs485-spy injector, the printed circuit board in 3D

    Rs485-spy injector

    rmingon/rs485-spy-injector

    RS485 Spy is an open-source hardware tool for real-time monitoring, logging, and testing of RS-485 networks

    44 parts7CERN-OHL-S-2.0
  • 2 layers50 × 20 mmM5stack board dynamixel ttl rs3485, the printed circuit board in 3D

    M5stack board dynamixel ttl rs3485

    kim-xps12/m5stack_board_dynamixel_ttl_rs3485

    When using this repository in your own environment, please import HY2.0-4P.kicadmod into the appropriate library. I import it into the global library so that it can be used from other projects as well.

    20 parts6MIT
  • 2 layers50.8 × 22.9 mmFeatherWing 485, the printed circuit board in 3D

    FeatherWing 485

    W8LID/RS-485-FW

    An RS-485 FeatherWing for the Adafruit Feather system.

    22 parts5

Questions

How big is the Max3485-rs485 PCB?

It measures 60 × 40 mm and has 2 copper layers, 1.6 mm thick.

What parts does the Max3485-rs485 use?

7 parts, built around the MAX3485ED. The full list of its parts is on this page, with links to each part's symbol and footprint where PCBKit's library has them.

Where can I download the Max3485-rs485 KiCad files?

From the embedrapp/deepee repository on GitHub, at the commit this page was made from: the board is validation/solutions/deepee-pcb-008/artifacts/max3485-rs485.kicad_pcb. PCBKit opens the board in the browser, with nothing to install.

Can I use the Max3485-rs485 design in my own project?

Its repository publishes it under the MIT licence: you may reuse it on that licence's terms (keep the author's notice, and for a copyleft licence share your changes alike).