Tech Stack

The AI stack we build with

No legacy tools. No preferred vendor. Just the best available option for each layer of every system we build — evaluated against performance, UK GDPR compliance, integration depth, and total cost of ownership.

When a better model ships, we evaluate it. When a vendor changes its pricing model, we move. What stays constant is the selection criteria — and the commitment to building systems that run reliably in production, not just in demos.

AI Models & LLMs

Foundation models we integrate and orchestrate for production workloads.

We treat LLM selection as an ongoing engineering decision, not a one-time procurement choice. Different models have measurably different performance profiles across tasks — long-context reasoning, structured output, tool use, and multilingual instruction following — and we route workloads to the right model for the job. We also build all agent systems to be model-agnostic by design, so a change in provider pricing or capability doesn't require a rebuild.

OpenAI GPT

General reasoning, structured JSON output, tool use, vision, and real-time multimodal tasks across most client products

Anthropic Claude

Long-context document analysis, regulated-sector reasoning, and nuanced multi-turn instruction following with fewer hallucinations

Google Gemini

Multimodal inference, Google Workspace data integrations, and cost-efficient high-throughput generation tasks

Mistral AI

European-hosted, GDPR-aligned LLM inference with lower latency and cost for sensitive or regulated workloads

Agent Frameworks

Orchestration layers that coordinate multi-step AI behaviour, memory, and tool use.

Single-model prompting only takes you so far. For production AI systems — ones that take actions, query external APIs, manage state across conversation turns, and handle failure gracefully — you need an orchestration layer. We use battle-tested agent frameworks to define how multiple AI components interact, how decisions branch, and how human oversight is preserved at critical steps. The framework we choose depends on the complexity, state requirements, and fault tolerance needed for each specific system.

LangChain

RAG pipelines, prompt chaining, tool-augmented agents, and document retrieval workflows with rich integration support

LangGraph

Stateful, cyclical multi-step agent workflows with conditional branching, persistent memory, and human-in-the-loop checkpoints

CrewAI

Role-based multi-agent collaboration with structured task delegation, sequential execution, and shared memory

AutoGen

Conversational multi-agent systems for enterprise automation, code-generation pipelines, and iterative problem-solving tasks

Voice AI

Speech recognition, synthesis, and real-time conversation infrastructure for voice-first AI products.

Voice AI has moved well beyond IVR menus. We build voice agents capable of handling full natural-language phone conversations — qualifying leads, booking appointments, resolving support queries — using a stack that combines telephony integration, real-time speech processing, and LLM-powered response generation. Every system includes escalation logic so calls reach a human when they genuinely need to, and full transcription for compliance and QA.

Twilio

Telephony integration, inbound and outbound SIP connections, call routing, and programmable voice infrastructure

ElevenLabs

Ultra-realistic voice synthesis and voice cloning for customer-facing agents where natural, human-quality speech matters

Deepgram

Real-time and async speech recognition with domain-specific fine-tuning, low latency, and speaker diarisation

Whisper

OpenAI's transcription model for batch audio processing, call logging, meeting summaries, and multilingual transcription

Automation & Workflow

Visual and low-code engines that connect systems, trigger AI actions, and automate high-volume tasks.

Not every automation needs custom code. For many high-value workflows — syncing CRM data on form submission, triggering AI enrichment on new records, routing support tickets by urgency — a well-configured workflow engine is faster to deploy, cheaper to maintain, and easier for non-technical teams to adjust. We choose between self-hosted and cloud-managed tools based on data sensitivity requirements and the client's internal technical capacity.

n8n

Self-hosted workflow automation with custom code nodes, AI model integrations, and complex multi-branch event logic

Zapier

Fast SaaS-to-SaaS automation for clients that need reliable integration without custom infrastructure or engineering overhead

Make

Visual multi-step automation with advanced data mapping, error handling, and scenario branching for complex operational flows

Data & Analytics

Storage, transformation, and BI infrastructure that powers AI-driven reporting and agent memory.

AI systems are only as useful as the data they can access and reason over. We design data layers that are structured for both transactional use and AI workloads — including vector embeddings for semantic search and RAG — and we build reporting infrastructure that surfaces AI-generated insights to business stakeholders without requiring them to write SQL. Every data architecture is documented, versioned, and designed with access controls appropriate for UK GDPR compliance.

Supabase

Postgres-backed backend with real-time subscriptions, row-level security, and pgvector for AI memory and RAG pipelines

BigQuery

Large-scale analytics warehousing, event stream processing, and SQL-based ML inference at enterprise data volumes

PostgreSQL

Primary relational database for transactional workloads, structured agent state, and vector search via pgvector

Metabase

Business intelligence dashboards built for client stakeholders — no SQL knowledge required, fully white-labelable

dbt

SQL-based data transformation, lineage documentation, and modular analytics pipeline management across warehouses

Web & Product

The engineering stack we use to ship fast, scalable, SEO-optimised digital products.

Every web product we build is engineered for commercial performance from day one — sub-second load times, Core Web Vitals that rank, accessible markup, and a codebase that a future developer can extend without archaeology. We use a consistent, opinionated stack that our team has optimised over dozens of production deployments. TypeScript everywhere, server components where they make sense, and no framework fatigue from chasing the latest release.

Next.js

Full-stack React framework for all web products — App Router, server components, API routes, and edge rendering at scale

React

UI component layer for web products, internal dashboards, AI chat interfaces, and cross-platform applications

Vercel

Edge deployment, pull-request preview environments, CI/CD pipelines, and Core Web Vitals monitoring for all frontend projects

Node.js

Backend API services, real-time WebSocket servers, background job processors, and server-side automation middleware

TypeScript

Type-safe development across every frontend and backend codebase — no exceptions, no gradual migration debt

Python

AI pipeline scripting, data processing, ML model evaluation, and back-office automation in regulated environments

CRM & Ops Integrations

Business systems we wire AI agents and automation workflows directly into.

AI only creates real business value when it connects to the systems where work actually happens. We have deep integration experience across the most common UK business stack — CRM, accounting, e-commerce, and sector-specific platforms — and we build these integrations with bidirectional data flow, proper error handling, and audit trails. Where a vendor provides an official API, we use it. Where it doesn't, we build the connector properly rather than screen-scraping around it.

HubSpot

CRM data sync, deal and contact creation, AI-powered lead qualification pipelines, and sales workflow automation

Shopify

E-commerce data access, product feed management, order lifecycle automation, and AI-personalised storefront integrations

Xero

Accounting integration for invoice processing automation, reconciliation agents, and AI-generated financial reporting

Salesforce

Enterprise CRM integration for high-volume customer workflow automation, AI enrichment pipelines, and reporting

EMIS Health

NHS patient record access and clinical workflow automation in regulated UK healthcare environments — DSPT-aligned

How we select tooling

Opinionated about outcomes. Agnostic about tools.

Every tool in this stack earned its place by solving a real problem in production — not by being fashionable, heavily marketed, or what we already knew. Before any new technology enters a client project, we run structured evaluations: accuracy at the specific task, data handling under UK GDPR, integration friction with the client's existing stack, and realistic total cost of ownership at production scale.

We also call out when simpler is better. Not every workflow needs an agent framework. Not every report needs a vector database. Part of our value is knowing where AI creates compounding returns — and where it adds cost and complexity for diminishing gain.

This stack is a snapshot of what we're building with today. It changes as the category changes. If you want to understand how any of these tools might apply to your specific project, the fastest way is a direct conversation.

No vendor lock-in

Every integration we build is abstracted so you can swap providers without rebuilding from scratch. Your AI stack should serve your roadmap, not constrain it.

LLM-provider independence

All agent systems are designed to run on OpenAI, Anthropic, Gemini, or locally-hosted models — so a pricing change or model deprecation doesn't break your product.

Compliance before capability

We only recommend tools with UK GDPR compliance documentation, EU/UK data residency options where required, and clear data processing agreements.

Cost modelled before build

We calculate API and infrastructure running costs upfront — not post go-live — so you know exactly what the system costs to operate before committing to an architecture.

Want to know which tools are right for your project?

Every project gets a tailored architecture. Tell us what you're building and we'll map the right stack, integration points, and delivery approach — before you commit to anything.