AI infrastructure guides

Practical decisions for planning, deploying and operating AI.

Where to start

BEYOND THE MODEL

Access, queues, versions and recovery to put a model into operation.

Measure memory, concurrency and latency before choosing infrastructure.

Define access, retention and review before connecting documents and conversations.

Documents and OCR
Prepare PDFs for search without hiding missing pages or extraction failures.
Incremental synchronisation
Apply changes and deletions without reloading the entire collection.
MySQL operating cost
Reduce repeated reads and writes before expanding the server.

Training data
Control versions, duplicates and the split between training and evaluation.
Distributed training
Limit the budget and resume work after losing a worker or GPU.
Model releases
Check files and evaluation before releasing a version for inference.

AI in existing systems
Connect databases, applications and telephony in verifiable stages.
Real-time audio
Separate capture, recording and transcription to recover each stage's failures.
Hybrid infrastructure
Place data and processing without multiplying network hops.
Infrastructure migration
Prove the destination and preserve data before retiring the old stack.
Agent permissions
Limit actions and access by context; check results outside the model.
PAP and PACT: what exists now
Separate announcements, published specifications and access that the service already offers.
Our discovery specification
Draft 0.1 for interfaces, access rules and prices.

Queues and safe retries
Recover tasks without duplicating actions or accepting results from stale runs.
AI observability
Connect queues, errors, versions and resources to completed work.
Deployment and rollback
Check the service and model before switching traffic; prepare recovery.
Backup and recovery
Plan the recovery of databases, documents, models and queues together.

I reply with scope and a proposal.