How we build

Most agencies hide the stack behind “proprietary technology.” That usually means one of two things: they’re wrapping an off-the-shelf platform and marking it up, or they don’t want you to know how replaceable they are.

The short version: we publish the whole stack — orchestration, models, memory, voice, evaluation, deployment — because you can’t easily verify technical claims and publishing lets you. Anyone hiding a stack behind “proprietary technology” is usually wrapping something you could name.

Last updated: 19 August 2026

The stack, published

Here’s ours. Verify it. Ask hard questions about it.

Orchestration

: LangGraph for stateful, multi-step agents where control flow matters and you need to inspect the graph. OpenAI Agents SDK and Anthropic tool use for direct model-native agent loops. CrewAI for role-based multi-agent coordination. Microsoft Copilot Studio when the client lives inside a Microsoft tenant and governance should follow existing tenant policy.
: We pick per project. Anyone who uses the same framework for every problem is selling their comfort, not your outcome.

Models

: Model-agnostic by default, and we design so a model swap is a config change rather than a rebuild. Deprecations happen; last year’s best model is next year’s migration.
: Selection is per-task: reasoning-heavy steps and cheap classification steps rarely want the same model. We benchmark against your data during the build and publish the comparison.

Memory and retrieval

: pgvector where you already run Postgres — fewer moving parts is usually the right call. Pinecone, Qdrant or Weaviate where scale or latency demands it.
: Retrieval is hybrid — semantic plus keyword — because pure vector search fails on exactly the specifics your business cares about: part numbers, permit IDs, unit references.

Voice

: Deepgram for speech recognition, LiveKit for real-time transport, Vapi or Retell for orchestration depending on integration needs.
: Target latency is under 1.2 seconds round-trip. Beyond about 1.5 seconds callers start talking over the agent, and the conversation degrades.

Evaluation

: LangSmith and Langfuse for tracing and observability. Braintrust for evaluation suites. Every agent ships with a regression test set built from your real cases. Every model or prompt change runs against it before deployment.
: An agent without evals is an agent nobody is measuring. It doesn’t fail loudly — it drifts.

Deployment

: Your cloud, your accounts, your data residency. AWS Bedrock, Azure AI Foundry or Google Vertex AI depending on where you already live. On-premise and hybrid where regulation or connectivity requires it.
: We do not host your agents on our infrastructure and rent them back to you.

Integration

: Direct API integration to systems of record wherever possible. n8n or Make for lightweight glue where a full integration isn’t warranted — and we’ll tell you which is which rather than over-engineering to inflate the invoice.
: MCP servers for tool exposure where the client stack supports it.

What this means for you

: No lock-in. Everything runs in infrastructure you control.
: No black box. You get the repository, the eval suite and the documentation.
: No surprise dependency. If a vendor in this stack fails or prices badly, the architecture allows substitution.

Frequently asked questions

Why publish your stack? Doesn’t that help competitors?

Competitors can already read our job ads. Clients can’t easily verify technical claims, and publishing lets them. It also filters out buyers who want magic rather than engineering.

What if we want a different framework?

Then we use it, assuming it’s sound. We have preferences, not religion.

Do you build with no-code tools?

Where appropriate. A workflow that genuinely fits n8n shouldn’t be hand-coded. But no-code platforms hit walls on error handling, state and testing, and we won’t pretend otherwise to make a build look cheaper.

How do you handle model deprecation?

Abstraction at the model layer plus the eval suite. When a model retires, we swap and re-run the regression set. Covered under AgentOps.

Can we run this entirely on-premise?

Yes, with open-weight models. Performance and cost differ; we’ll model both during the audit.