TEC.01
Model Context Protocol (MCP)
MCP server development and business-system integration: clients, transports, authentication, tool permissions, security checks and a scoped implementation approach.
Open pageTechnology · 37 pages
Our AI technology stack.
We do not have a house stack we sell you. We have a set of trade-offs we understand well enough to choose between — and the evaluation practice to know when we chose wrong.
Introduction transcript. Client: Which AI model should we use? Qubi: Ooh. Let’s try it on your actual work. Qubi: We connect the tools, test the answers, and watch what happens. Client: Even when your first idea is wrong? Qubi: Especially then. I like a useful surprise. Qubi: The right stack earns its place.
At a glance
How a model gets chosen here.
Stated so they hold without the rest of the page.
- 01 What decides it
- Four inputs, in roughly this order of weight: where the data may live, what the task demands, what latency the interface can absorb, and what a request costs at your volume.
- 02 How it is tested
- Against your workload, not a public leaderboard. A cheaper model that wins your extraction task beats a leader on reasoning benchmarks.
- 03 If a model changes
- The model is a swappable component. Prompts are not entangled with one model’s quirks, so a deprecation is a migration, not a rewrite.
- 04 Vendor position
- We resell none of these and earn no commission from any of them, so the recommendation is the only thing we sell.
- 05 Where it runs
- Hosted APIs, open weights on hardware you control, or both. Self-hosting is supported where the data cannot leave.
HOW WE CHOOSE
Four inputs. Measured, in order of weight.
Model selection is an engineering decision. Capability is rarely the binding constraint: most production tasks are within reach of several models, and the choice is decided by the other three.
So we benchmark against your workload rather than against a leaderboard. A model that leads on public reasoning benchmarks can be beaten on your extraction task by something a quarter of the price, and the only way to know is to measure. That measurement is cheap; the assumption is not.
We also design for substitution. Model providers deprecate versions, change pricing and occasionally change behaviour within a version. A system where the model is a swappable component survives that; one where prompts are entangled with a specific model’s quirks does not.
In order, first to last
-
01Weighs most
Where the data may live
Residency, confidentiality and whether anything may leave your network. This is the input that often rules a model in or out before capability is considered.
-
02Then
What the task demands
Long context, tool use, multimodal input or a strict output shape. Matched to the workload and tested against real examples of it.
-
03Then
What latency can be absorbed
What the interface can tolerate, at the percentile that matters. A faster, smaller model is often the right answer to a user-facing step.
-
04Then
What a request costs
Priced at your projected volume, not at a list price. The cheapest model per token is not always the cheapest per correct answer.
Model families
Which family fits what the system has to do.
We do not have a favourite: toggle what the system has to do and see which families we would reach for first. Drawn from our model pages — a starting point, not a benchmark.
It has to
| Family | Long-running agents & code | Very long documents | Images & visual docs | Runs on your hardware | European provenance | Many languages | Cheapest at volume | Fit |
|---|---|---|---|---|---|---|---|---|
|
Claude
Enterprise coding and long-running agents
|
Yes | No | No | No | No | No | No | — |
|
GPT
Computer use and a million-token window
|
Yes | Yes | No | No | No | No | No | — |
|
Gemini
Long-context and multimodal strengths
|
No | Yes | Yes | No | No | No | No | — |
|
Llama
Self-hosted, on your own hardware
|
No | No | No | Yes | No | No | No | — |
|
Mistral
European open weights, Apache options
|
No | No | No | Yes | Yes | No | No | — |
|
Qwen
Broad sizes, strong multilingual coverage
|
No | No | No | Yes | No | Yes | No | — |
|
DeepSeek
Visual documents, coding, open weights
|
No | No | Yes | Yes | No | No | Yes | — |
|
Small language models
1B–14B: most tasks at a fraction of the cost
|
No | No | No | Yes | No | No | Yes | — |
why we'd reach for it
not the reason we'd pick it
Shortlist
Llama · Mistral · Qwen · DeepSeek · Small language models
Jev · our router
You don't have to pick just one.
Jev routes each task to the model that fits it, with typed outputs and confidence gates. Published benchmarks included.
Jev model routing →THE STACK
What we build with, by job.
- Frontier models
- Anthropic Claude OpenAI GPT Google Gemini
- Serving & inference
- vLLM Ollama AWS Bedrock Azure OpenAI
- Retrieval
- pgvector Qdrant Pinecone Elasticsearch
- Orchestration
- Model Context Protocol LangGraph Temporal Celery
- Evaluation & observability
- Custom eval harnesses OpenTelemetry Langfuse Grafana
Chosen per engagement against your constraints. We are not a reseller for any of these and have no incentive to prefer one.
EVERY TECHNOLOGY PAGE
Every technology page.
-
-
TEC.02
Multi-Agent Systems
Supervisor and handoff patterns — and why most multi-agent problems are one-agent problems.
Open page -
TEC.03
Evaluating AI Systems
Building evaluation sets, LLM-as-judge failure modes, and gating deploys on regression.
Open page -
TEC.04
Context Engineering
Retrieval, compaction and memory — why prompt engineering is the wrong frame in 2026.
Open page -
TEC.05
Choosing a Vector Database
pgvector, Pinecone, Milvus and Qdrant compared on the axes that actually decide it.
Open page -
TEC.06
Fine-Tuning & Adaptation
LoRA, full fine-tuning and distillation — what each one buys, and why retrieval usually comes first.
Open page -
TEC.07
Inference Optimisation
Quantisation, batching, KV caching and speculative decoding — the levers that decide what serving costs.
Open page -
TEC.08
AI Observability
Tracing, cost attribution and drift detection for systems whose failures are silent by default.
Open page -
TEC.09
Multimodal AI
Vision, document and audio models in production, and where the accuracy claims stop holding.
Open page -
TEC.10 10 pages
Model Families
Every model family we build on, what each is genuinely good at, and how the selection is made per workload.
Open page -
TEC.11 4 pages
Jev
Build decision-making tools with TypeSafe AI Jev: parallel typed outputs, model routing, confidence gates and integration design. Includes dated provider specifications and links to published benchmarks.
Open page -
TEC.12
Managed Agent Runtimes
OpenAI Agents API public beta: persistent sessions, sandbox choices and practical enterprise pilots for research, document review and incident investigation.
Open page -
TEC.13
Agent-to-Agent Interoperability
A2A and MCP explained through enterprise handoffs, September 2026 platform announcements, shared context, permissions and a measured integration pilot.
Open page -
TEC.14 5 pages
Quantum Computing
What quantum computers can and cannot do today, how you reach one through the cloud, and where the honest overlap with AI actually is.
Open page -
TEC.15
Voice AI & Speech Systems
Speech in, speech out: latency budgets, interruption handling, telephony integration and where voice agents actually break.
Open page -
TEC.16
Structured Outputs & Tool Calling
Function calling, JSON schemas and constrained decoding — the mechanism most production AI systems are actually built on.
Open page -
TEC.17
AI Guardrails
Input and output filtering, policy enforcement and escalation design, and why a guardrail is a control rather than a product.
Open page
Qubi Your guideTechnology guide
01 / 9
Shall I show you around?
Explore with Qubi
