Technology
AI & Machine LearningOpenAI, PyTorch, Hugging Face
The AI stack we reach for, the trade-offs we name, and when we use something else — OpenAI and Anthropic for the frontier models, PyTorch and Hugging Face for the open-source, with the choice made against the use case, the data and the compliance posture.
What we use it for
OpenAI and Anthropic for the frontier models — the right answer for the use cases that need the frontier capability, the right tooling, and the right SLA. PyTorch for the open-source ML, with the right model and the right training pipeline. Hugging Face for the model hub, the inference and the evaluation.
We extend the stack with the orchestration the product needs — LangChain, LlamaIndex, Haystack. The choice is made against the use case, the data and the compliance posture.
When we choose it over the alternative
OpenAI and Anthropic are right when the use case needs the frontier capability, the team is not large enough to train a model, and the data does not justify the cost of fine-tuning. PyTorch and Hugging Face are right when the use case is well-defined, the team can train or fine-tune a model, and the data is the competitive advantage.
The right answer depends on the use case, the data and the compliance posture. We will say so on the call, with a written rationale for the choice and the trade-offs named.
When we do not choose it
We do not choose a frontier model when the use case is well-defined and the open-source model is good enough. We do not choose an open-source model when the use case needs the frontier capability and the team is not large enough to train it. We do not choose the default when the data posture (on-premise, air-gapped) makes the alternative the right answer.
Frequently asked
The questions the team asks
- OpenAI or open source?
- It depends on the use case, the data and the compliance posture. OpenAI is the right answer for the frontier capability, the right tooling and the right SLA. Open source is the right answer when the use case is well-defined, the data is the competitive advantage, and the team can run the model.
- How do you handle the eval?
- An evaluation suite is part of the architecture from sprint one. The suite includes a held-out test set, a rubric for the qualitative checks, and a CI gate that catches regressions before deploy. Every prompt change, every model swap, every fine-tune is run against the suite.
- Can the model run on our infrastructure?
- Yes. The model can run on your VPC, on your hardware, on your data centre, with the deployment model the data posture requires. The architecture is deployment-agnostic, with the choice made in the discovery.
Have a project in mind?Let’s scope it together.
Tell us the outcome you need. You get an approach, a rough timeline and next steps within one business day.