Sheet 01Rev C

Projects

A direct-to-chip liquid-cooling architecture for high-power AI processors, combining an aluminum cold plate, optimized internal flow channels, and a modular rack architecture—designed not just to cool the machine, but to explore how efficiently we can move heat before the silicon hits its limits.

  • The problem: AI accelerators are actively pushing towards power intensive procedures. The challenge is no longer simply removing heat. Nope. it is moving enormous amounts of heat uniformly and efficiently without creating localized hotspots, excessive pressure drop, or unsustainable pumping and cooling loads. 
  • The solution (potentially): A liquid-cooling system that optimizes itself. Done by using real-time thermal feedback and trained AI to continuously reshape flow distribution across parallel microchannels. Flow patterns would be determined by current condition of the chip for example identified hotspots or pressure drops. 
  • The progress: Simulation of a data centre condition has been done with aluminum extrusion racks. A small 1.5hp pump for water circulation installed and a liquid cooling radiator fan to dump excess heat. Potential micro channels to effectively remove heat are under testing.
SolidWorksANSYSMATLABHeat Transfer
Target heat flux
~100 W/cm² 
Plate material
Al 6061-T6 (testing)
Working fluid
Water
Measurement
Thermocouples
Drawings & photos · 04
CADAI Generated model of a potential cold plate flow channel
PHOTOModel data centre rack with aluminum extrusions
CADInstalled radiator dimensions
VIDEOSolidworks assembly 
Digital renderings · 00

Designed and fabricated a sprint vehicle around a two-stage 3.85:1 gear drivetrain. Every gram removed was a gram that stopped arguing with the motor.

  • Two-stage gear drivetrain (3.85 GR) tuned for the RPM–torque tradeoff over a 10 ft sprint.
  • Laser-cut chassis iterated for mass, shaft alignment, and bearing friction.
  • Tolerance stack-up and vibration mitigation validated through repeated timed runs.
  • Final result: 1.35 s over 10 ft — 2.26 m/s average.
SolidWorksLaser CuttingPrototyping
Gear ratio
3.85 : 1
Distance
10 ft (3.05 m)
Best time
1.35 s
Avg speed
2.26 m/s
Drawings & photos · 01
CADDrivetrain layout — two-stage reduction
Digital renderings · 00

Instrument families taken from user requirement to batch acceptance: GD&T, tolerance stack-up, DFM/DFA, tooling coordination, and in-process quality control.

  • Applied GD&T and stack-up analysis so assemblies survive real manufacturing variation.
  • Evaluated material selection, force transmission, and stress distribution per instrument.
  • 3D-printed prototypes iterated against functional and dimensional testing.
  • Built production workflow: process planning → tooling → in-process QC → batch acceptance.
SolidWorksGD&T3D PrintingCNC
Material
Surgical SS 420 / 304
Methods
CNC, forging, hand finish
QC
In-process + batch
Annual sales
$15,000+
Drawings & photos · 01
PHOTOInstrument set — finish inspection
Digital renderings · 00