Backend Portfolio
Four production architectures that define my backend direction: maritime telemetry ingest, dynamic EV fleet charging, event-sourced greenhouse climate control, and industrial edge AI pipelines. Each card represents a system designed to stay predictable under failure. Flip them to see the engineering levers behind the scale.
Selected Work
Telemetry
Real-time monitoring for 340+ vessels handling 2M+ concurrent data streams via satellite.
Contribution: Built a Go ingest service with custom gRPC connection pooling, delta-compression, and ClickHouse materialized views.
Impact: Reduced satellite bandwidth by 68% and achieved sub-second fleet-wide aggregation latency.
Maritime Fleet Ingest Engine
The Hard Part
Satellite links are expensive and jittery. 2M concurrent streams hitting a naive ingest layer cause massive memory pressure and backpressure ripples that can crash the edge buffers on the vessels.
The Approach
We implemented a custom gRPC pool with explicit backpressure propagation. If the ClickHouse buffer fills, the ingest layer signals the edge devices to increase compression or pivot to priority-only telemetry. Materialized views pre-calculate fleet health, avoiding expensive runtime joins on massive datasets.
Scheduling
Intelligent load balancer for 80+ EV vans on a 2MW-limited grid connection.
Contribution: Designed a Go-based scheduling engine using Redis Streams for real-time state sync and PostgreSQL for persistent audit logs.
Impact: Guaranteed 100% route readiness while staying under 1.8MW peak demand, saving 800K EUR in grid upgrades.
Dynamic Fleet Charging Scheduler
The Hard Part
Total fleet demand (5MW) exceeds grid capacity (2MW). A single late arrival or battery fault can cascade through the schedule, leaving morning routes unfulfilled.
The Approach
We built a constraint-based scheduler that re-evaluates the charging graph every 30 seconds. Redis Streams handle the high-frequency state updates from ESP32 gateways. If a vehicle misses its charging target, the system dynamically reallocates current from non-critical vans to compensate.
Event Sourcing
Event-sourced climate control across 12 independent zones with deterministic timing.
Contribution: Built an Elixir/Phoenix backend using CQRS patterns and an EventStore to capture every control decision.
Impact: Eliminated climate jitter and provided 100% auditability for agronomists, reducing crop stress events to zero.
Autonomous Greenhouse Controller
The Hard Part
When a crop fails, agronomists need to know exactly why a specific vent opened at 3 AM. Standard logging is too sparse to reconstruct the state of 12 zones simultaneously.
The Approach
We pivoted to full event-sourcing with Elixir. Every sensor reading, control signal, and agronomist override is a first-class event. The "Decision Engine" is a pure function of these events, making it trivial to replay scenarios in a "shadow" mode to test new climate strategies without risk.
Edge Ingest
High-performance defect detection using Jetson Nano and custom C++ kernels on the manufacturing floor.
Contribution: Built a FastAPI ingest layer with custom C++ ring buffers for zero-drop image processing and TensorRT optimization.
Impact: Reduced model memory footprint by 40% and achieved sub-50ms local-first inference latency.
Industrial Edge AI Pipeline
The Hard Part
High-speed production lines wait for no one. Dropping a single frame means a defective part reaches the customer. Python's GIL and standard GC cycles are too unpredictable for 60fps inspection.
The Approach
We moved the heavy lifting to custom C++ ring buffers that sit outside the Python heap. FastAPI handles the high-level API and orchestration, while the C++ core manages the TensorRT inference and raw camera data. This hybrid approach provides sub-50ms latency with predictable memory usage.
Architecture Feed
Waiting for system reveal
Reliability Engineering Notes
When the character triggers a card lever, the system architecture reveal is activated, surfacing the design decisions behind high-vertical reliability.
In distributed systems, failure is the default state. Architecture is measured by how predictably the system recovers.
API schemas and event payloads require the same ownership and versioning as the services that emit them.
Logs are for people; traces and metrics are for systems. Designing explainable telemetry is part of engineering.
A healthy service must be able to refuse work. Bounded queues and load shedding prevent systemic collapse.
R&D Bench
R&D into Redis Streams vs Kafka for sub-100ms state synchronization in industrial fleet scheduling.
Exploring zero-copy memory patterns between C++ ingest cores and high-level Python/Go APIs.
Building tools to "time-travel" through event-sourced greenhouse climate data for agronomist auditing.
Personal Philosophy