From saved interview to full overview
Once an interview is captured on a phone, a chain of background steps turns it into structured, enriched data — automatically. Here's the whole flow:
- Interview capturedOn the phone
The interviewer records the chat, taps the answers, adds notes and photos. Everything — audio, answers, notes, photos, location, who/when — is saved against one unique interview ID. Offline? It queues and syncs automatically. - Gemini extracts the key informationAutomatic
A background process runs every 15 minutes. It finds new recordings and sends each to Gemini with a fixed research prompt (interviewer, depot and type are injected from the record, so nothing is guessed). Gemini returns structured JSON: key facts & figures, decisions, issues, action items, notable quotes, insights, and every door name mentioned — with the surrounding conversation context. - ValidationAutomatic
The JSON is checked — if it doesn't parse cleanly it's retried once, otherwise it falls back to a plain-text version. Nothing is silently dropped. - Enrichment — adding product contextAutomatic
Each door name captured is matched against the Howdens product range using the conversation context (a "moulded Dordogne" ≠ "oak Dordogne"). Names get classified into product family, finish and price tier. Where the context genuinely can't tell, it's flagged for a human spot-check rather than guessed. - Stored & linkedAutomatic
The enriched result is saved back against the same interview ID — so the recording, photos, answers, notes and the Gemini extraction are all traceable to one record. Same for field observations (kept separate as extra context). - Collation — the bigger pictureAutomatic
The dashboard merges every interview's structured data into one overview. Group and filter by interviewer, depot, or interviewee type; pull every price point and percentage; stack the quotes and insights; count recurring issues. All consistent, because every interview used the same prompt and format. - Human reviewResearch team
Anything flagged as ambiguous (unusual door names, unclear context) appears in a short review list for the team to confirm — then it's final.
By the numbers: each interview is processed in roughly 30 seconds at a cost of about 1 pence. Field observations follow the same path with their own prompt, feeding in as extra context — they never count as interviews.
Every step is keyed to the original interview ID, so there's a full traceability chain from raw recording to the collated overview.