Model-agnostic by default
Prompts, tools and evaluation sit behind an abstraction, so switching or mixing models is a configuration change rather than a rewrite.
Technology
The stack behind
The stack behind
agents that hold up.
We are deliberately boring about infrastructure so we can be ambitious about capability. These are the tools we actually use in production — not a logo wall of everything we have heard of.
Agent frameworks
Claude APILangGraphLangChainMCPTool callingFunction calling
Models & inference
Claude (Opus, Sonnet, Haiku)OpenAIOpen-weight modelsModel routingStructured outputs
Retrieval & data
pgvectorPineconeQdrantHybrid searchEmbeddingsRerankers
Conversational channels
WhatsApp Business APITwilioRealtime voiceWeb chatTelegramEmail
Backend
PythonFastAPINode.jsLaravelPHPGo
Frontend
ReactNext.jsTypeScriptTailwindJavaScript
Mobile
FlutterAndroidiOSReact Native
Databases
PostgreSQLMySQLMongoDBRedis
Cloud & delivery
AWSGoogle CloudAzureDockerCloudflareCI/CD
Evaluation & observability
Eval harnessesTracingPrompt versioningCost monitoringGuardrails
Engineering choices
How we decide
How we decide
what to use.
Model and framework choices are reversible if you build the boundaries properly. We design for that from day one.
Cheapest model that passes
We route each task to the least expensive model that clears your quality bar, and cache aggressively. Quality is measured, not assumed.
Data boundaries first
We agree what leaves your environment before writing code — from no-training API commitments through to deployment inside your own cloud.
Deterministic where possible
If a step can be plain code, it is plain code. Models are used for judgement and language, not for arithmetic or routing that a rule handles better.
Evaluation as a first-class artefact
Every agent ships with a fixed test set, so changes are regression-tested instead of eyeballed.
Observable in production
Tracing, cost dashboards, latency alerts and quality sampling, because an unmonitored agent is an unmanaged risk.