Francisco Booth is a software developer and the founder of ClarifAI Consulting, a small automation practice in Hampshire. He develops ComputeFence, a preflight diagnostic tool for machine-learning jobs on rented GPU infrastructure.[1][2][3]
Start with ComputeFence – preflight checks for rented GPU training jobs →
GitHub Background
Booth’s original account describes paying for a GPU while a training process silently ran on the CPU. The practical problem was not choosing a larger model but noticing that the environment had failed to use the hardware. ComputeFence moves those configuration checks before the expensive launch.[4]
Career
At ClarifAI Consulting, Booth works directly with small businesses on workflow assessment, integrations, training and ongoing support. His practice connects existing business systems rather than requiring a replacement software stack. His public software projects also include Exolio AI and other consumer and business applications.[1][2]
Contributions
ComputeFence checks whether PyTorch can see a GPU, whether model caches and checkpoint output use persistent storage, and whether disk space and distributed-training configuration match the rented machine. It can compare an environment fingerprint between sessions. Drift produces warnings; the tool does not claim to validate the training script, model architecture or learning rate, and it does not monitor the running job.[3]
Tools, writing & talks
Writing, research & talks
Start here
Preflight checks for GPU visibility, persistent storage, disk headroom and environment drift before a training job starts.
Read the checks and non-goals first, then compare them with a disposable training environment. Pay attention to cache and checkpoint paths separately: fixing one does not make the other persistent.
Creator and maintainer[1][3]Original sourceA first-person account of a silent CPU fallback and the checks that could have exposed it earlier.
Author[4]