Zouhair El Koush

A desktop machine that prints a plastic object and builds working electronics inside it, in one job, with no one touching it halfway through. Four tools ride one motion platform and swap automatically. Against the commercial precedent, Voxel8, the difference is that the machine places the components itself instead of a person pausing the print.

• T0 · FDM hot end. Lays the dielectric structure.
• T1 · Pneumatic syringe. Writes conductive traces, then deposits conductive-adhesive dots.
• T2 · Vacuum head. Picks and places SMD components.
• P · Metrology tool. Contact probe plus down-camera. Every tool is re-probed against a bed datum after each change.
• The number that forces the architecture. A sub-50 µm tool-tip Z error budget across automatic tool changes. 5 K of thermal drift in a 500 mm aluminium frame alone accounts for 59 µm.
• Two MCUs, one Klipper instance. Axes are allocated so a multi-stepper axis always lives with its endstop. No homing move crosses a USB boundary.
• Hybrid CAM. Extending OctoPnP into toolchanger-aware CAM that interleaves printing, dispensing and placement in a single G-code job, with camera-based recovery of each tool's XY offset.
• Owner-repairable by design. No tool carries electronics, so no tool can fail in a way its owner can't diagnose. With T0 alone it is an ordinary FDM printer.

I lead the electrical and electronics subsystem of AUB's student-built off-road race car: three ESP32s across three boards covering data acquisition, the driver dashboard, and telemetry back to the pit over LoRa. Power is a LiFePO4 battery plus an alternator, sized for 4+ hour endurance runs.

• Three boards, three jobs. Data acquisition, the driver dashboard and pit telemetry each run on their own ESP32.
• 40+ sensor channels. Brake pressures and temperatures, engine and CVT temperatures, wheel speeds, battery voltage, fuel level and GPS.
• 20 Hz streaming. The firmware samples every sensor and pushes structured GPS and sensor packets over UART to the pit client.
• BAJA DAQ client. A Python desktop app with a live GPS map, a drag-and-drop dashboard, and race review with timeline scrubbing and lap playback.
• Custom module builder. Teammates build new dashboard widgets (gauges, graphs, status lights) from the UI, without writing code.
• Design report. I wrote the subsystem's technical design report: architecture, sensor selection, wiring, and coordination with the powertrain, brakes and chassis teams.

The tool the AUB engineering student body uses to get things printed. It runs inside AUB's WordPress makerspace page as an authenticated iframe, and it replaced a manual spreadsheet workflow with a permission-gated queue, inventory, analytics and access control.

• A real auth chain. WordPress passes the logged-in user, an Azure Auth Function issues a JWT, and a Permissions Function fills the session before any page renders.
• Seven permission-gated pages. Print queue, printers, analytics, access management, users and inventory, a read-only wall monitor, and a separate student portal.
• Two auth flows. Staff carry a permission record. Students get a simpler parallel flow that only ever shows their own requests.
• Danger zone. The semester reset is gated to the Head Jedi and Web roles.
• Estimates built in. 3DPME quotes filament weight the moment a student uploads, with no slicer in the loop.
• Fleet control. PCS drives the physical printers through a Cloudflare Worker and a Raspberry Pi relay.

Staff monitor every printer and control the Klipper ones from a browser tab: send a sliced G-code file, pause, resume, cancel, add or rename printers, and watch live webcam feeds, all inside the AUB WordPress iframe.

• Secrets never reach the browser. A Cloudflare Worker holds the fleet API key and the Access service token server-side, and gates every request by origin and session.
• Cloudflare Access in front of the Pi. The edge validates the service token and forwards a signed JWT, which the relay checks instead of trusting a header.
• FastAPI relay on a Raspberry Pi 5. Fans status out across the fleet; a print request passes a ready check and an identity check, returns 202, and uploads in the background.
• MAC-keyed self-healing. Printer IP addresses drift on every power cycle, so the relay keys each machine by MAC address and finds it again on its own.
• Degrades instead of failing. On a relay hiccup the Worker serves the last known good state from a 5-second cache.
• Adaptive polling. Every 30 to 90 seconds depending on activity, over the Red Room's own 4G LTE link.

Estimates a print's filament weight from geometry alone, so the portal can quote a job the moment a student uploads it. It reads the mesh itself (convexity, bounding-box fill, surface-to-volume and height ratio) to classify each part and apply calibrated corrections. It runs as an Azure Functions endpoint that the RedRoom PMS calls.

• Geometry rules. Ten detection rules derived from how parts actually print: flat plates, hollow shells and tall structures each get their own correction.
• Multi-part assemblies. Reads ZIP and RAR archives of STL parts and detects mixed assemblies, such as threaded screws next to large blocks with internal holes.
• Fine geometry detection. Face density picks out threaded bolts and nuts and raises their estimate.
• G-code ground truth. Parses slicer output for the real weight, which is how the benchmark was measured.
• Four input formats. STL, ZIP, RAR and G-code.