Production RAG. Production agents.
Most AI feature work fails the same way: a clever prototype on a laptop, a slide that says “production ready,” and a six-month integration project that never ends. The two installations below are both fixed-price because we have shipped enough of them to know the shape.
A RAG Installation is the chat-with-our-docs feature, done well: ingestion, chunking, retrieval, citations, refusal, eval. Four weeks, $22,000.
An Agent Workflow Build is a multi-step automation that runs a real business process: tools, loops, observability, a cost ceiling, and a human override. Five to six weeks, $32,000.
Both deploy to your accounts. Both come with a written eval and a handoff video. If a third AI workload does not fit either of these shapes, we will tell you on the call.
RAG Installation
A production retrieval-augmented chatbot that answers from your own corpus, with citations, refusals, and a clear evaluation harness.
Read the Holdfast Lab case study →Scope
- Document ingestion pipeline (PDF, HTML, Markdown, Notion, Google Drive) with incremental sync.
- Chunking and embedding strategy tuned to your content shape, not a generic 1,000-character split.
- A vector store of your choice (Postgres + pgvector by default; Pinecone or Turbopuffer on request).
- Retrieval ranking with hybrid search (BM25 plus vectors) and a re-ranker for top-k cleanup.
- A chat interface using Claude Sonnet 4.5 with inline citations and a hard refusal gate.
- An evaluation harness (10 to 30 graded questions) so quality regressions are visible before they ship.
Deliverables
- Ingestion service, retrieval service, and chat UI, all in your production.
- Eval suite with a baseline you can re-run on every change.
- A 10-minute handoff video plus README that another engineer can take over from.
Agent Workflow Build
A multi-step agent workflow that runs a real business process end to end: fetches inputs, calls tools, writes back, and tells you when it gave up.
Read the Holdfast Lab case study →Scope
- Workflow design: the explicit list of steps and tools, written before any code.
- Tool implementations (typed inputs and outputs, idempotent writes, structured errors).
- A run loop using Claude Sonnet 4.5 with prompt caching and explicit step budgets.
- Observability: every run logs its prompts, tool calls, costs, and final state, queryable.
- A human-in-the-loop UI for the runs that the agent deferred to a person.
- A cost ceiling per run and a circuit breaker that halts the workflow on repeated failures.
Deliverables
- A working agent in your production, running on real inputs.
- A dashboard your team can use to inspect, retry, and override individual runs.
- A README that names every tool, every limit, and every known failure mode.
Two AI features we shipped.
AI Feature Build questions.
Because we have built enough of these now to know what they cost. The two installations on this page are the ones we have shipped multiple times. The shape of the work is known, even if the corpus or the workflow is new.
Want one of these in your stack?
Send a one-page brief describing the corpus or the workflow. We will reply by morning Chicago time.