Production-usedPackage
Airpack
Build immutable delivery bundles, execute resumable runbooks, and preserve install and upgrade evidence.
Private AI delivery assurance
ProofTools
Build, move, harden, and prove AI systems for customer-owned infrastructure. From model adaptation to air-gapped delivery, every acceptance decision should have artifacts and evidence behind it.
Built from paid private-AI delivery work, not a reference architecture.
Customer-ownedOn-prem, VPC, bare metal, air-gapped
Content-addressedArtifacts, manifests, checksums, evidence
Deployment-shapedTests against the runtime that will ship
Acceptance-readyResults engineering and procurement can inspect
The missing layer
Between a model release and customer acceptance
Build systems prove that software was produced. ProofTools addresses what happens next: adaptation, packaging, constrained transfer, installation, capacity testing, and evidence on the customer’s own infrastructure.
Each tool owns one verb and composes through versioned contracts. The result is a delivery path that can fail honestly, resume safely, and leave a record strong enough for an acceptance decision.
- 01AdaptBehavior and model fit
- 02PlanTyped execution intent
- 03PackageImmutable delivery units
- 04MoveConstrained transport
- 05MeasureRuntime and GPU capacity
- 06ProveReadiness evidence
One verb per tool
A composable private-AI delivery suite
Maturity is stated explicitly. Production-used tools are separated from active development and planned acceptance capabilities.
Validated product
Behavioral adaptation
AirForge
Train private behavior adapters and merged model artifacts against an explicit trust boundary, then ship the result with corpus, training, deployment, and evaluation evidence.
- Reference gates
- 6 / 6
- Evaluation probes
- 600
- Deployment shape
- MXFP4 + vLLM
Production-usedMove
Airbridge
Transfer split bundles through restricted WebDAV, Windows jump-host, and SSH paths with checksum verification.
Active developmentPlan
AirPipeline
Compile semantic delivery intent into typed execution plans, durable run state, and provider-independent jobs.
DevelopingMeasure
AirBench
Capture workload, latency, throughput, GPU utilization, memory, and cost evidence under reproducible conditions.
DevelopingProve
AirProof
Evaluate readiness contracts and assemble the final evidence needed for release and customer acceptance.
Engineering discipline
Evidence is part of the deliverable
Pin the inputs
Models, images, corpora, contracts, and runtime assumptions receive immutable identities.
Test the real shape
Acceptance evidence targets the merged model, serving runtime, GPU topology, and environment that will operate.
Preserve failures
Interrupted transfers, rejected samples, evaluator corrections, and no-go findings remain visible.
Hand off operations
Delivery includes configuration, smoke checks, limitations, monitoring assumptions, and rollback.
Origin
Built because the delivery path did not exist
ProofTools grew from delivering a private AI product into a German university hospital’s customer-owned, GPU-backed, air-gapped environment. The hard problems were not confined to model inference.
They were bundle integrity, constrained transport, repeatable installation, target-specific validation, model behavior, capacity, and evidence that security and procurement teams could review.
Current public result