The stack we actually
build production AI on.
We're not tied to one model or framework. Each project gets the provider, orchestration layer, and infrastructure that fits its constraints — cost, data residency, latency, compliance. These are the 23 we reach for most, and what we use each one for.
The models
we build on.
RAG pipelines, structured extraction, and LLM integrations — chosen per project, not by default.
Voice agents, RAG pipelines, and tool-using agents with long context and real guardrails.
Native multimodal workflows with tight Vertex AI and GCP integration.
A European model provider that matters when EU data sovereignty is a hard requirement.
Self-hosted, open-weight models when data residency, cost control, or no external API calls is the priority.
For clients already running on Microsoft's cloud and enterprise compliance stack.
For infrastructure and compliance postures that are already AWS-native.
LLM and ML systems for clients already on GCP, including Gemini access and MLOps tooling.
Sourcing, fine-tuning, and deploying open-source models via Transformers and the model hub.
How agents are
structured and routed.
Guardrailed, stateful agents — scoped tool access and human-in-the-loop checkpoints baked into the graph.
Role-based multi-agent systems — defined roles, defined handoffs, the same guardrail discipline throughout.
RAG pipelines and data connectors — indexing, retrieval, and evaluation built in.
Multi-provider LLM routing with Redis-backed caching — a provider outage becomes a config update.
Getting systems to answer
from your real data.
A self-hostable, open-source vector database suited to UK/EU data residency.
Clean, structured extraction from tables, multi-column layouts, and embedded charts for RAG.
Turning messy real-world documents into clean, structured text for RAG and extraction pipelines.
The pipeline behind
a production voice agent.
Real-time speech-to-text, LLM reasoning, and text-to-speech in one low-latency pipeline.
The telephony layer under production voice AI agents — real numbers, real call routing.
Speech-to-text, including self-hosted deployment when cost or data privacy rules out a hosted API.
Fine-tuning and knowing
what a system actually did.
Efficient LLM fine-tuning — LoRA/QLoRA without full-scale training costs.
Tracking training and fine-tuning runs so results are comparable and reproducible.
Tracing, cost tracking, and evaluation wired into production LLM systems.
Models that run
where the camera is.
We pick the tool
to fit the project, not the reverse.
This list isn't exhaustive. If you're running something else, tell us about your constraints and we'll tell you honestly whether it's a fit.