Electronic Brake Performance Monitoring for HGV trailers. It reads TrailerCAN (ISO 11992) data that a Teltonika FMC650 on the tractor reports, and scores how well the trailer's brakes actually performed on the road. Fleets can spot brake wear between scheduled inspections.

• Splits continuous telemetry into separate braking events using brake pedal rise and fall, then filters out events that are too short, too slow or too noisy to score.
• Scores each event with a Brake Performance Value: actual deceleration as a percentage of the deceleration expected from pedal input, axle load and road gradient.
• Scores trailers by VIN, so each score belongs to the physical trailer whichever tractor pulled it. A separate Cab EBPM view scores the tractor's own brakes.
• Telemetry dashboard, actual-vs-expected scatter chart and a downloadable PDF Braking Performance Report.
• Wrote the technical documentation and the customer guide.

Turned a single-file tachograph (.ddd) report into a staff tool and then into a panel inside Tracpoint. It reads vehicle and driver files straight from Teltonika's TachoSync API.

• Built a Flask service that proxies the TachoSync API and adds a browse view: customer, vehicles or drivers, search, then view or download.
• Designed the Tracpoint integration as one iframe panel. Tracpoint issues a short-lived signed token and the panel only serves files for that company. The reviewing developer approved the approach.
• Fixed the Gen1 file parser. It was matching marker bytes inside signatures and speed data, which produced fake days and hundreds of fake drivers.
• Deployed to a QA environment over HTTPS with a strict frame-ancestors policy.

A UI redesign prototype for a legacy PHP/ExtJS app whose users complained about buried menus and too many clicks. Live map, asset grid, route replay and tacho screens.

• Status filter chips with counts, one search box that filters the list and map together, and the six common asset actions on the detail panel instead of a right-click menu.
• Mock data uses the legacy app's own field names, so it can be pointed at the real status endpoint without renaming anything.
• Route replay generates repeatable journeys anchored to real depot coordinates.

A demo dashboard for Teltonika FMC650 eco-driving trip-end reports: scores by truck and driver across a sample fleet of 6 trucks and 8 drivers. Styled to match the tacho tool so the two read as one product.

Shows a vehicle's fuel use from its CAN bus data for any asset and date range.

• Detects whether the vehicle reports fuel as a percentage or as litres. When it reports litres, it estimates tank capacity from the history.
• Draws the route on a map, coloured by fuel level. Hovering over the fuel chart highlights the matching point on the map.
• Plots nearby fuel stations from OpenStreetMap so refuels can be matched to places.
• A live vehicle diagram shows door, bonnet, light, brake and parking brake status from CAN signals.

Validates and corrects supplier fuel card CSVs against the asset database before import, then writes them out in Tracpoint's 11-column fuel template.

When a reviewer asked for direct database inserts and odometer calibration, I proposed feeding Tracpoint's existing fuel importer instead. That importer already handles both, so the tool stays read-only and the logic lives in one place.

• Run an EC2 test server that hosts each demo as its own Docker Compose project behind one Caddy reverse proxy, with HTTPS and a landing page that shows which demos are up.
• Mapped a 6,400-commit PHP/ExtJS codebase into 23 linked architecture notes: request routing, module permissions and data storage. New features now start from those notes.

Most issues on this project get a PR within minutes, so I filled the gap that was missing: test runs on Linux with AMD's RADV driver. I set up two remote test machines (an RX 570 desktop and a Steam Deck on SteamOS) and run the full test suite under Vulkan validation.

• Wrote and merged my own fix: image tests now skip min/max reduction cases on GPUs that lack the feature. Verified fail to pass on the RX 570.
• Found a DCC clear layout regression in a format PR during review. The author fixed it and it merged.
• Filed a multisample depth bug and confirmed the cause was depth buffer (HTILE) decompression by toggling one driver flag. Another tester then reproduced it on Vega.
• Filed a test failure on RADV that led to a fix PR. The merged fix credits my RX 570 and Steam Deck runs.
• Got a homebrew game running at 59.94 FPS on the Steam Deck and wrote up the bring-up report.

A homebrew game ran at about 40 FPS on the Steam Deck's GPU driver (RADV) but crashed every time on llvmpipe, Mesa's software Vulkan driver. Nobody had a cause, so I traced it out of AnyPS5 and into Mesa itself.

• Got a symbolized backtrace with gdb and debuginfod. It showed llvmpipe crashing inside LLVM while compiling a mesh shader.
• The emulator's own shader dumps fired too late to catch it, so I used Mesa's SPIR-V dump to capture the shader. spirv-val confirmed it was valid, which pointed the bug at Mesa and LLVM, not the emulator.
• Wrote a standalone Vulkan C harness that loads just that shader, then shrank it with spirv-reduce and an automated crash test.

Contributor to a GPL mod for the game Skate. Wrote the troubleshooting guide and bug report template (merged), found and fixed a Windows plugin test failure in someone else's PR (merged), and opened a build fix so a probe target links.

A stock and crypto bot tested live on a paper account. The work was mostly safety engineering: each test day produced incident reports, and each incident became a safeguard.

• A self-healing audit every 15 seconds that finds positions without a stop order and protects them.
• Retry logic for stop orders rejected during the broker's post-fill settlement hold.
• A pause flag, a cool-down after stop-losses, and a pause-first liquidation script that removed a race condition.

Source-code auditing for the Nextcloud bug bounty program on HackerOne, using a local lab instance. Audited the OpenID Connect login app and tested one lead with a 200,000-case differential test. That test showed it was a robustness bug, not a security issue, so it was not reported as one.