Softwaresimanticfirmwareembeddedcode-reviewstm32
Simantic's Review Bot Runs Your Pull Request's Firmware on a Simulated STM32 Chip, and Its Demo Catches a One-Bit Slip That Leaves an LED Stuck On
Simantic Bot, from a Fall 2026 Y Combinator startup, builds the firmware in a GitHub pull request, runs it on a simulated microcontroller, and posts a review whose every runtime claim comes from those runs; its homepage example flags a +15 that should be +16.
The demo is a pull request titled “blinky: drive the LED through BSRR,” against an STM32F072RB board. It’s an illustration on Simantic’s homepage, not a customer’s repository, but it shows the whole loop. The diff swaps a read-modify-write of the chip’s output register for a single write to its bit set/reset register: GPIOA_BSRR = on ? (1u << LED_PIN) : (1u << (LED_PIN + 15));.
Someone types “@simantic-bot review.” The bot reports “Build · cmake for STM32F072RB · 0 warnings,” then “Simulate · 3 s on STM32F072RB · UART and PA5 recorded.” The firmware’s own log says the LED went off at 1.0 seconds and 2.0 seconds. The simulated pin disagrees: “PA5 never goes low.”
The review’s verdict is “Changes requested.” Its explanation: “Using BSRR for the LED is the right idea, but the reset half is off by one bit, so the LED never turns off.” The suggested change on line 78 of main.c turns the + 15 into + 16. The review also lists what it checked, including “PA5 pad follows the LED on/off messages” over 3.0 seconds and that the boot banner and blink messages “keep their 500 ms cadence.”
The bot builds projects that use CubeMX or a Makefile, CMake, the Raspberry Pi pico-sdk, ESP-IDF or Zephyr. The simulator behind it is also a command-line tool. Simantic’s CLI docs list three chip models in sim models: the RP2040 with 264 KB of RAM and 2048 KB of flash, the STM32F072RB with 16 KB and 128 KB, and the STM32F401RE with 96 KB and 512 KB. The docs promise “the same ELF with the same inputs gives the same output, byte for byte,” on Apple Silicon Macs or Linux x64.
A simulator is only as good as its match to real silicon, and Simantic has a page for that. Its validation page says the company will “Run the same firmware on real silicon and in simulation. Compare the results and investigate every mismatch.”
Pricing on Simantic’s pricing page:
| Plan | Price | Included credits | Concurrent runs |
|---|---|---|---|
| Community | Free | $10 a month | 3 |
| Developer | $100 a month or $1,000 a year | $200 | 10 |
| Enterprise | Contact sales | Custom | Custom |
Reviews draw down the monthly included usage, and “with less than $1.50 left, it is refused.”
The company is part of Y Combinator’s Fall 2026 batch, with a YC profile that reads “Firmware simulation for AI agents.” It was founded in 2025, has a team of four and is based in Waterloo, Canada. CEO Seungmin Hong was a firmware engineer at Tesla, Simbe Robotics and Rapyuta Robotics. CTO Ahnaf Shahriar worked at Neuralink and on Tesla’s Dojo and Autopilot teams.
Firmware review is heading toward the place web code reached years ago, where a pull request doesn’t merge until something has actually run it. The difference is that firmware needs a chip, and simulators like this one stand in for the board. The thing to watch is that validation page: a bot is only as trustworthy as the gap between its simulated pin and a real one, and Simantic hasn’t yet published how many mismatches it has found.