Model Providers

BtSavvy's posture on model providers

The model-provider question is the question every enterprise AI programme has to answer first. Frontier-model lock-in is real: model behaviour, prompt sensitivity, fine-tune economics, and provider-specific guardrail surfaces are all sticky. Provider neutrality is therefore not a marketing posture; it is an engineering discipline that has to be designed in deliberately, against the natural pull of single-provider depth.

BtSavvy's posture is provider-agnostic by design and provider-specific by competence. The firm works with frontier-model providers (Anthropic, OpenAI, Google, Meta, Mistral, AWS, AI21, Cohere) and with open-weight models hosted in client environments or via managed services. FORGE is identical regardless of provider; what differs is the implementation surface and the failure-mode catalogue. The accelerator library carries provider-specific calibration discipline so engagements can move between providers when the workload's economics or risk posture requires it.

Provider categories the firm works with

Frontier-model providers (closed-weight, API-accessed). Anthropic Claude, OpenAI GPT and o-series, Google Gemini, Meta Llama (where surfaced via managed services), Mistral Large, Cohere Command, AI21 Jurassic, AWS-hosted frontier models. The firm's frontier-model competence is broad rather than monogamous.

Open-weight models (self-hosted or managed). Meta Llama, Mistral, Falcon, Qwen, DeepSeek, and other open-weight families. Self-hosted on the client's cloud (Vertex AI, Azure ML, AWS SageMaker, on-premises GPU infrastructure) or via managed services (Together AI, Anyscale, Bedrock, Vertex AI Model Garden).

Specialised models. Embedding models, reranker models, vision-language models, code-generation models, and domain-specific fine-tunes. Selection per workload, not per portfolio standard.

FORGE on model selection

Factory. The customised builder is provider-aware. Provider-specific accelerators (prompt scaffolds, guardrail integrations, evaluation harnesses) live in the accelerator library, calibrated per provider.

Observability. Model behaviour is observable at provider-specific resolution. Anthropic guardrail trigger rates, OpenAI structured-output failure rates, Gemini safety-filter activation rates: each is a first-class SRE signal in the firm's observability discipline.

Review. Model selection is a signed gate. The choice of provider, model variant, and fine-tune approach: each is recorded against the accountable engineer who signed the decision into the gate log.

Governance and Trust. Provider-specific terms of service, data-handling commitments, and regulatory postures are tracked as policy-as-code. EU AI Act provider-of-record obligations differ between hosted-API and self-hosted-open-weight deployments; the firm's governance discipline tracks the difference rather than collapsing it.

Expert Intelligence. This is the discipline where model-provider selection lives. Context engineering, retrieval architecture, and targeted fine-tuning are the firm's working surface: and each adapts to the chosen provider rather than being abstracted away.

Engagement-pattern sketches

Representative patterns FORGE is built to deliver across providers:

A frontier-model deployment with a designed-in second-provider fallback. The primary deployment runs on one frontier provider; a designed-in fallback routes to a second provider on detection of provider-side incidents, with calibration-skew checks ensuring the fallback maintains semantic equivalence within tolerance.

An open-weight self-hosted deployment for sovereignty-sensitive workloads. A workload that cannot send data to a hosted-API provider, deployed against an open-weight model on the client's cloud with FORGE governance: including the same EU AI Act conformance evidence and ISO/IEC 42001 control posture a hosted-API deployment would carry.

A fine-tune programme that converts a frontier model into a domain expert. Targeted fine-tuning (LoRA, QLoRA, or full fine-tune depending on workload) on the chosen provider's substrate, with evaluation harnesses, calibration-skew regression, and signed model promotion across the lifecycle.