Context

At Solenium I worked on the software stack around a smart electricity meter. The role sat between embedded firmware, ML pipelines, backend APIs, and operational tooling, which made it a good fit for my background in engineering physics, STM32 systems, and applied AI.

The public version of this case study avoids proprietary implementation details and focuses on the architecture-level work I can discuss.

What I Worked On

• Smart-meter NILM: led ML development for real-time energy disaggregation at 10 Hz with sub-2-second latency targets.
• Data acquisition: tuned STM32 timer/ADC behavior and helped synchronize acquisition with downstream ML compute stages.
• Retraining loop: built workflows for neural-network updates using user feedback captured from dashboard and telemetry flows.
• Backend APIs: integrated Flask and Django services with JWT-based role control for protected operations and telemetry views.
• Fleet safety: contributed to firmware-facing features including OTA update flows, watchdogs, reboot logic, and secure data handling with OpenSSL.
• Device rollout: helped deploy and debug six field devices under real-world conditions.
• Operational tooling: worked around firmware version serving, calibration/validation scripts, Quoia webserver RBAC, and telemetry observability.

Why It Matters

This project is the clearest bridge between my embedded and AI work. The ML model only mattered if the meter sampled well, the firmware recovered safely, the APIs exposed the right state, and field devices could be updated and debugged without fragile manual steps.

Skills Demonstrated

• Edge ML systems that combine signal acquisition, model inference, telemetry, and human feedback.
• Cross-layer debugging across C/C++, Python services, MQTT, web tooling, and device deployment.
• Production habits: role-based access, OTA safety, watchdogs, calibration reports, and reproducible validation.