Technology
The technology behind intelligent systems.
Listed because we work with them, not because they make a good logo wall. The specific choice in any engagement follows the workflow and the constraints you already have — which is why the older names are on this list too. Most systems worth adding AI to are already running on something, and meeting them there is usually cheaper than replacing them.
Nothing here is a partnership or a certification, and none of it is a default. Models in particular sit behind an abstraction, so provider choice stays an implementation detail rather than a commitment you inherit from us.
How we choose
Knowing a technology is not the same as knowing when to use it.
Six decisions that come up in almost every architecture session, what settles each of them, and the case where the obvious answer is the wrong one.
Which model runs this step
The task, the latency budget, the cost per call at your real volume, and where the data is permitted to be processed — evaluated per workload rather than once per company.
When we would not
A frontier model is the wrong answer for a high-volume classification step that a small fine-tuned model, or a classical classifier, does more cheaply and more predictably. Reaching for the largest model is a cost decision disguised as a quality one.
Whether this needs an agent at all
How much of the path is genuinely unknown at design time. Agents earn their overhead when the sequence of steps depends on what earlier steps returned.
When we would not
If the steps are known in advance, an agent is a slower, less testable, more expensive way to run a workflow you could have written down. Most requests we receive for an agent are requests for a workflow with two model calls in it.
How much orchestration to take on
Whether runs are long, need to survive a restart, or must be resumable and auditable step by step. That is what pushes a workflow onto a durable execution engine rather than a request handler.
When we would not
A framework whose control flow you cannot read is a bad trade for a three-step workflow. We have replaced more orchestration with plain code than we have added, and a graph library that hides the retry semantics is worse than a queue you understand.
Where the knowledge lives
The shape of the data and the query. Structured filters and joins belong in the relational database you already run; a vector index earns its place when similarity over free text is the actual access pattern.
When we would not
A dedicated vector database is unnecessary for most corpora that fit comfortably in Postgres with pgvector, and it adds an operational surface and a consistency problem that has to be worth something. Scale justifies it. Novelty does not.
Whether to self-host the model
Your data-residency position and whether a hosted provider is acceptable for the data class in question. This is settled in discovery, before anything is built against an assumption.
When we would not
Self-hosting is the wrong default: it trades a per-token bill for GPU capacity planning, evaluation, patching and an on-call rota. It is right when a compliance constraint or a genuinely predictable high volume makes it right, and expensive when chosen for control in the abstract.
Where it runs
Your environment, not ours. If the system has to live in your cloud account or on your own hardware, that is a constraint gathered at discovery rather than negotiated after the architecture assumes otherwise.
When we would not
We do not move an estate to a different cloud in order to add AI to it. Meeting a system where it already runs is nearly always cheaper than the migration that would make our side simpler.
Group 01
AI & Models
The reasoning layer. Provider selection is a per-workload decision — cost, latency, context window and where the data is allowed to be processed — and it is revisited rather than fixed at the start.
- OpenAI
- Anthropic
- Gemini
- Llama
- Mistral
- Hugging Face
- PyTorch
- scikit-learn
Open-source models where applicable. Classical models where they beat an LLM on cost and accuracy — most classification and forecasting work still does.
Group 02
Agentic AI
What turns a model call into a system: decomposition, tool calling, memory boundaries, and the evaluation harness that tells you whether a change made things better or only different.
- LangGraph
- LangChain
- LlamaIndex
- CrewAI
- MCP
- Temporal
- Tool calling
- Memory
- Evaluation
Group 03
Data
Where knowledge lives and how it is retrieved. Retrieval quality dominates generation quality in almost every business case, so this layer usually decides whether the system works.
- PostgreSQL
- MySQL
- SQL Server
- MongoDB
- Redis
- Elasticsearch
- pgvector
- Pinecone
- Qdrant
Group 04
Backend
The services that hold the business logic, the queues and the integration surface. Ordinary engineering, and the reason a prototype survives contact with production volume.
- Python
- FastAPI
- Django
- Node.js
- TypeScript
- Go
- Spring Boot
- .NET
Group 05
Frontend
The interfaces people actually work in — copilots embedded in an existing product, review queues, approval surfaces.
- React
- Next.js
- TypeScript
- Tailwind CSS
- React Native
- Streamlit
- Vue
- Angular
Group 06
Cloud
Wherever your compliance position requires the system to run, including inside your own account. That constraint is gathered in discovery rather than negotiated afterwards.
- AWS
- Azure
- GCP
- AWS Bedrock
- Azure OpenAI
- Vercel
- Cloudflare
- On-premise
Group 07
Voice
Telephony, speech recognition and synthesis. Latency is the binding constraint here, not accuracy, and the architecture is chosen around it.
- Twilio
- Deepgram
- Whisper
- ElevenLabs
- LiveKit
- WebRTC
- SIP
- Asterisk
Group 08
Engineering
The unglamorous work that decides whether a system survives its first month: containers, pipelines, tracing and tests.
- Docker
- Kubernetes
- Terraform
- GitHub Actions
- Jenkins
- OpenTelemetry
- Grafana
- Playwright
AI maturity
Where you are determines what to build next.
Most organisations move through these in order. Skipping a stage is usually why a pilot stalls — the tooling above only helps once the stage below it is real.
- Stage 1Experiment
- Stage 2Copilot
- Stage 3Automation
- Stage 4Agent
- Stage 5Multi-Agent
Your Next Intelligent System Starts Here.
Tell us what you’re trying to improve, automate or build. We’ll help you identify the right AI strategy and engineering path.