What we build with
Not a partner wall and not a badge collection. These are the tools the work runs on and what each one is for here. We have no partnership, reseller or certification arrangement with any of these companies, and a page that implied otherwise would be the first thing worth distrusting on this site.
The whole set
Pick anything to read how we use it. Drag it if the graph is in its own way.
technology graph
Drag any node to pull the graph apart. A refresh puts it back.
49+ tools we build with. Pick one to read how we use it, and drag it if you want the graph out of its own way.
The working vocabulary
The same practice read by layer rather than by vendor. Every term here is one we would expect to be asked about on a technical call.
Retrieval
- ingestion and chunking
- embeddings
- pgvector
- Qdrant
- hybrid keyword and vector search
- reranking
- source citation
- evaluation sets
Agents
- MCP servers
- Claude Code
- Codex
- custom runtimes
- routing and shared state
- agent handoffs
- permission boundaries
- human approval gates
- run tracing
Data
- ingestion pipelines
- warehouse and lakehouse
- transformation and modelling
- quality and freshness checks
- access control
- lineage
Delivery
- Harness.io pipelines
- CI/CD
- environment and secrets management
- feature flags
- progressive delivery
- rehearsed rollback
Why the model sits behind an interface
The frontier moves every few months. A system wired directly to one provider’s SDK inherits that provider’s pricing, deprecation schedule and outages, and unpicking it later is a rewrite rather than a decision.
So the model goes behind an interface of our own. Swapping one out is a config change and a couple of evaluation runs, and the evaluation set is what makes that swap safe to make: without it, changing models is a guess dressed up as an upgrade. It is the same reason retrieval is hybrid and the same reason the runtime is pinned in a container. None of it is exotic. It is the ordinary engineering that decides whether you still own your options in a year.
Tell us what you are trying to build
If the right answer is a rules engine and no model at all, we would rather say so early. The bill is smaller and the thing works.
Start a project