Internal platform — access controlledPhase 3 in progress
Client
ATÖLYE — strategic design and innovation
Our role
Product design, engineering and ongoing development
Shape
Multi-phase — knowledge, then search, then reporting
Stack
Next.js · Claude API · Pinecone
01
Knowledge base. Scattered project documents unified into one indexed corpus.
02
Semantic search. Ask in plain language; get passages with their sources attached.
03
Reporting. Recurring questions become saved views instead of new requests.
01 — The brief
The answer already existed — in a folder no one could find
A studio that has run hundreds of projects accumulates proposals, research, decks and post-mortems. The knowledge is real; the retrieval is the problem. The brief was to make the archive usable by the people doing the next project — without asking them to learn a new system.
01Answers with citations — never an unsourced claim
02Permissions that mirror the organisation, not the file system
03Delivered in phases, each one useful on its own
04Costs that stay predictable as usage grows
02 — Under the hood
Retrieval first, generation second
Documents are chunked, embedded and indexed in a vector store. The model only ever answers from what retrieval returns.
An internal tool is only used if it can be checked
Every answer carries the passages it came from and a link to the source document. When retrieval finds nothing, the platform says so instead of improvising. That single rule is what moved the tool from a demo to something people open on a Monday morning.
Next.jsClaude APIPineconeVector search
Question
“What did we propose the last time a client asked for a maturity model?”
Answer
Three passages, three documents, each one openable.
04 — Delivered in phases
Each phase shipped something the team could use, rather than a step toward something they couldn't yet see.
Knowledge base → semantic search → reporting
05 — How it ran
Define to Operate, with the same team throughout
01DefineInterviews to find the questions people actually ask, then a phased roadmap.
02DesignInterface built around the citation, not the chat bubble.
03BuildIngestion pipeline, vector index and application, phase by phase.
04LaunchRolled out to a pilot group first, then widened as retrieval quality held.
05OperateContinued development, new sources and cost monitoring as usage grows.
