Case studies
Six pieces of work, each with the problem, what I built and the results. Client code is confidential, so client work is described at the level of detail my CV uses.
Hardware-in-the-loop verification of a safety-certified ignition controller
Read the full case study The bench, the eight gates, arm-early timing and how verdicts are kept honest.Context
A two-stage gas furnace ignition controller certified to UL/CSA, ANSI Z21.20 and NFPA 54. I was the sole owner of the verification test suite and of the automated bench that drives the real hardware.
Problem
Tests were hand-written and duplicated, stimulus windows were not actually being measured for a large family of cases, and bench failures were hard to attribute to firmware, harness or instruments.
Verification pipeline
- 01 State-machine catalogue YAML model of every state, event and transition
- 02 Test generator A comprehensive executable suite, no hand-written duplication
- 03 Bench device layer AC source (pyvisa/SCPI), CAN monitor, Modbus I/O — one thread per instrument, shared monotonic clock
- 04 8-gate validator Arrange / Act / Assert against the captured run
- 05 JSONL timeline Every command, signal edge and bus message on one axis
- 06 Evidence Excel workbooks, bilingual HTML dashboards, Mermaid failure graphs
The eight validation gates
- Preconditions
- Stimulus timing
- State sequence
- Alarms
- Signal history
- Cross-layer consistency
- Timer accuracy
- Requirement traceability
Results
- Traced a "hardware fault" reported for months to a leaked USB handle, which unblocked an entire test family.
- Moved a C#/.NET WinForms serial-bus monitor into version control, fixed a log-parser crash and reduced build and deployment to one command.
- Owned delivery end to end: requirement traceability, issue tracking, branch / PR workflow, Conan packaging, CI and monthly client reporting.
Agentic AI harness and Model Context Protocol servers
Context
Bringing LLM agents into a regulated engineering workflow without giving up traceability or control, and publishing the reusable parts as open source.
Problem
Ad-hoc prompting does not scale across a team or survive review: prompts rot, agents act without guardrails, and nothing is reproducible.
Agent harness
- 01 Engineer request Human in the loop sets the task and approves results
- 02 Specialised agents & skills Scoped roles with versioned prompt workflows
- 03 Policy hooks Guardrails on what agents may run or change
- 04 MCP servers Prompt templates and project tools exposed over Model Context Protocol
- 05 Reviewed output Code, reports and analysis back into the normal PR flow
MCP-Prompts, the open-source part
- Server for storing and serving prompt templates to any MCP client
- Published to npm, Docker Hub and MCP directories (Glama, MCP Market, AIBASE)
- Storage adapters and AWS deployment examples
- Rust reimplementation, mcp-prompts-rs
Results
Diagnostic test engine for an aircraft onboard maintenance system
Context
The Onboard Maintenance System monitors aircraft condition and diagnoses member systems during flight and maintenance operations, on ARM / RTOS targets under aviation standards.
Problem
Diagnostic tests must run asynchronously, never interfere with each other, and behave predictably on timeouts and failed preconditions.
DiagTest engine
- 01 Test request Data collection, passive, interactive or non-interactive mode
- 02 Interference check Blocks conflicting tests from running together
- 03 Job queue Jobs processed asynchronously through a message queue
- 04 State machine An extensive set of operational states, preconditions and timeouts
- 05 Monitoring & results Live parameter monitoring correlated with maintenance messages
Results
- Python integration tests for communication protocols, state transitions and failure scenarios.
- Flatbuffers serialisation, Conan dependencies, cross-compilation with Clang, UML in Enterprise Architect.
Inspection proxy and dynamic firewall for the Kernun Adaptive Firewall
Context
Linux network security appliance: an HTTP/HTTPS forward proxy with TLS inspection, plus daemons that keep firewall rules in step with the network.
Problem
Inspect encrypted traffic and authenticate users without breaking clients, and react to address changes on interfaces without leaving the firewall in an inconsistent state.
Proxy and firewall path
- 01 Client Kerberos authentication, LDAP user / group lookup
- 02 Forward proxy C++ on Linux, TLS inspection, generated CA and client certificates
- 03 Certificate checks OCSP validation
- 04 Security stack Suricata IDS, nftables, logs aggregated in PostgreSQL
ipmon → rust-network-mgr
- ipmon: C++20 daemon watching IPv4 / IPv6 address changes over Netlink and updating nftables rules
- RAII, thread-safe and atomic operations, Unix-socket IPC, systemd-hardened Debian packaging
- rust-network-mgr: Rust redesign with atomic rule updates, lock-free structures and epoll
Results
Citadela Resort: a booking site for a private villa near Brno
Read the full case study The brief, what was delivered, the Booking.com sync, design decisions and the NFC access layer.Context
Website and booking back end for a villa above the Brno Reservoir that is let as a whole to one party at a time, built for the owner, a local accommodation business. Live at citadela-resort.cz.
Problem
The original single-file page was English-only, sent no inquiries and was nearly invisible to search engines, and nothing stopped a weekend being sold directly and on Booking.com at once. Booking.com offers no public API for a property this size.
Booking.com calendar sync
- 01 Booking.com Extranet Exports the property calendar as iCal
- 02 Inbound sync Scheduled job replaces the Booking.com snapshot
- 03 One availability check Booking.com and direct stays read together
- 04 Staff confirm Occupancy written in one transaction
- 05 Outbound feed Direct stays published back as iCal
Results
MIA: vehicle telemetry and IoT control from phone to microcontroller
Read the full case study Architecture, the digital twin, simulation-first design and the safety posture for a real car.Context
One system spanning an Android app, a Raspberry Pi, ESP32 and Arduino boards and a car's diagnostic port, driven by an AI agent through Model Context Protocol tools.
Problem
Devices that speak CAN, OBD-II, serial, GPIO and HTTP usually end up glued together by scripts that only run with every device plugged in, and an agent must never be able to actuate anything safety-relevant in the car.
OBD-II digital twin
- 01 Potentiometers on an MCU Stand in for throttle and road speed
- 02 Serial bridge USB serial → ZeroMQ PUB
- 03 OBD worker Live car state from telemetry
- 04 ELM327 emulator Answers PIDs on a virtual serial port
- 05 Diagnostic apps Connect as if to a real car