ZEKLAR
EN / IT
← Work
2024 · Embedded Engineering / Electric Vehicles

E-ION

Every cell thinks for itself

A distributed Battery Management System forged on Formula Student circuits, where each cell module runs its own State-of-Health model and balances the pack chemically, extending service life by 30%.

CELLS: 96 / 96 OK■ BMS ACTIVESoH: 98.2%

Year

2024

Client

Italian automotive consortium (confidential)

Role

System architecture, embedded firmware, AI modelling

Duration

10 months

Team

5 embedded engineers, 1 ML engineer, 1 hardware designer

Category

Embedded Engineering / Electric Vehicles

// The context

The high-performance EV battery market is at an inflection point: centralised BMS designs treat the pack as a monolith, masking per-cell chemical drift until damage is already done. Italian Formula Student teams have spent years amassing high-density telemetry, an asset that rarely travels beyond the paddock. In 2024, an industrial consortium saw in that raw data the raw material to redesign battery intelligence from the ground up.

// The challenge

The real difficulty was not the AI model. It was making it run on low-cost, low-power microcontrollers in hard real time without compromising IEC 61508 functional safety. Distributing intelligence across dozens of modules meant rethinking the communication topology and maintaining global pack coherence with no omniscient central node. Chemical balancing demanded a physics-empirical hybrid that no off-the-shelf ML framework was equipped to host.

approach.rs

// The solution

Zeklar adopted a network-of-agents architecture: each cell module hosts an STM32 running a SoH model written in Rust, compiled no_std and compressed via 8-bit quantisation to fit within 48 kB of flash. Modules communicate over CANopen with a lightweight supervisor that aggregates local decisions without centralising compute. The model was trained in Python on 14 months of Formula Student telemetry and validated in Simulink against an NMC digital twin. Balancing uses a short-horizon Model Predictive Control loop each cell computes autonomously every 50 ms.

// What made it special

01

Inference in 48 kB

The 8-bit quantised SoH model runs on an STM32F4 in under 3 ms per cycle, bare-metal, adding less than 8 mW to the cell module's power budget.

02

Active chemical balancing

The distributed MPC acts on current and temperature to minimise SEI layer growth, extending pack life by 30% in accelerated ageing tests.

03

Race data, real product

14 months of telemetry, over 900 GB, supplied thermal and current-load distributions impossible to replicate in a lab, making the model genuinely robust.

04

SIL 2 functional safety

The full firmware chain was developed with static analysis, automated mutation testing, and 97% MC/DC coverage, meeting SIL 2 requirements.

// Results

+30%

Pack life extension

<3 ms

SoH inference latency

48 kB

Firmware footprint

97%

MC/DC coverage

900+ GB

Telemetry processed

96

Cells in pilot pack

// The process

01

Discovery & data

Audit of the Formula Student telemetry archive and selection of the distributed communication topology.

02

Modelling & validation

SoH model training, NMC digital twin in Simulink, and hardware-in-the-loop validation on a real cell test bench.

03

Firmware & hardware

Firmware in Rust/C with flash optimisation, cell module PCB design in Altium, and thermal qualification from -20 °C to +70 °C.

04

Integration & testing

96-cell pack integration, accelerated ageing campaign, independent SIL 2 audit, and certifiable documentation.

"Zeklar took an idea that existed only on paper and brought it to production, with the systems-engineering rigour we expected from a top-tier partner, and a pace we didn't think was possible from anyone."

R&D Director, automotive consortium (confidential)

// The impact

E-ION proved that cell-level distributed intelligence is industrially viable without sacrificing certifiability: the consortium launched a 500-unit pre-series programme. The architecture is now the subject of two European patent applications filed jointly with the client.

2 EP Patent Applications Filed
NEXTKAIROS

ZEKLAR software technology 4.0/5.0

Let's talk
ZEKLAR · SOFTWARE THAT MAKES THE DIFFERENCE.
ZEKLAR · VAT L81419031T · info@zeklar.com© 2026