User Insight Hub
Interview Transcripts to Cited Research Artifacts
A research assistant for teams working from interview transcripts. It answers questions with inline citations and generates 59 kinds of structured research artifact across multiple datasets at once.
Teams that run qualitative research accumulate transcripts faster than they can read them. The insight exists in the corpus but recovering it means re-reading everything, and any synthesis produced by hand cannot be traced back to what a participant actually said.
- Weeks of manual work between finishing interviews and having usable findings
- Hand-written synthesis cannot be traced back to source quotes
- Personal information sits in transcripts that need to be searchable
- Comparing findings across several studies means doing the work again per study
- A general-purpose assistant will answer confidently from outside the corpus
A stateful agent that classifies each request, retrieves from an anonymised transcript index, and either answers with validated inline citations or generates one of 59 structured artifact types across the selected datasets.
- Intent classification on every message, routing to question answering, artifact generation, or an out-of-scope refusal
- Transcript chunking with metadata and quote extraction, with PII removed during indexing rather than at response time
- Citation validation against retrieved chunks, so unsupported quotes cannot reach the interface
- A template registry driving 59 artifact types, making a new type a registry entry
- Multi-dataset querying with technology and geography context preserved per dataset
- A dedicated latency optimisation pass on the retrieval path, planned and measured rather than opportunistic
How it was built
A research team finishes forty interviews and then loses weeks turning them into something a product decision can rest on. Any synthesis written by hand is also impossible to trace: six months later nobody can tell whether a claim came from a participant or from the person writing the slide.
So traceability was the requirement, not a feature. Every claim the system makes resolves to a quote from a named transcript.
The workflow classifies each incoming message before it does anything else: a question, an artifact request, or out of scope. Routing on intent rather than answering everything means the system never quietly responds to something it was not asked.
When the corpus does not support an answer, it says so and redirects. That behaviour is easy to skip and it is the difference between a research tool and a plausible-sounding one.
Transcripts are chunked with metadata and quote extraction, and personal information is removed during indexing. Filtering PII at response time leaves it sitting in the index, which is the wrong place for it the moment anyone else gets access.
Citations are validated against the retrieved chunks, so a quote that does not exist in the source cannot survive to the screen.
Segments, journey maps, jobs-to-be-done, pain points, decision drivers, contradictions and the rest are generated from a template registry rather than hand-written prompts per type. Adding an artifact type is a registry entry, which is why there are 59 of them and not five.
Several datasets can be queried together for cross-dataset comparison, with technology and geography context preserved per set so the comparison means something.
- Findings recovered by re-reading full transcript sets
- Synthesis claims untraceable to source quotes
- PII handled ad hoc in searchable transcripts
- Cross-study comparison repeated per study
- Questions answered with validated inline citations
- 59 artifact types generated from the indexed corpus
- PII removed before anything is indexed
- Several datasets compared in one query with context preserved