Problemcluster C-1
Duplicate reporting between shifts
≈ 11 hours / week · 14 people · 3 groups
“We type the same numbers into two systems every Friday. Nobody reads the second one.”
marked for reportconfidence: high
Employee listening, at scale
Everyone’s answer. In days, not weeks.
Not a weeks-long analysis where there’s no time to ask everyone — a quick, affordable one where you do. Dopta interviews your employees, fifteen minutes each, in any of 30+ languages, following up for concrete examples, frequencies and hours. Then identity is stripped from every transcript, and what was said is clustered into quantified findings your consultants turn into a report.
Process efficiency — auditEnglish
Group: Productionabout 4 minutes left
I’d rather not answer
Type your answer…
- median interview length
- 15 min
- people per audit
- 25–500
- identities stored with content
- 0
- on-premise capable
- 100%
Anonymous by architecture
Runs fully on-premise
30+ languages
Fully anonymous
Transcripts deleted after 90 days
Method
From a brief to a report in four moves.
Scroll through the moves — each one hands off to the next.
01current move
Brief
You split the company into groups and give each a brief: topics, tone, must-ask questions, target length. Briefs are versioned data — a new audit type is configuration, not code.
02current move
Invitations
Every employee gets a personal one-time link. Reminders and milestone messages run themselves. No accounts, no passwords — it works on a factory-floor phone.
03current move
Interviews
The AI asks one question at a time and follows up for examples, frequencies and hours. It never judges, never advises, and respects “nechci odpovídat”. Every interview ends with a summary the respondent confirms.
04current move
Findings & report
Transcripts are stripped of identity, findings extracted, quantified and clustered across groups. The model drafts the report — the consultant edits it and signs it.
Anonymity
We know who responded. We can never know who said what.
Anonymity is not a client option or a promise in a slide deck. Participation data and interview content live in separate database schemas, under roles that physically cannot read each other. The single bridge between them is deleted when the audit closes — after that, who-said-what is unrecoverable by design.
schema: identitywho was asked
- j.novak@firma.czcompleted
- p.svoboda@firma.czreminder sent
- k.mala@firma.czinvited
linkage severed
at close
at close
schema: contentwhat was said
- session 01J9X4K…34 turns
- finding: duplicate reporting11 h / week
- cluster: shift handover3 groups
Nothing below n = 5
Findings are only ever shown for groups above the minimum reportable size. Smaller slices are refused by the query layer itself — not hidden by the UI.
Names redacted before display
A redaction pass removes colleague names from quotes before anything reaches a dashboard, a cluster, or a report.
Transcripts expire
Raw transcripts are purged 90 days after delivery by default — earlier on request. Findings remain; the words themselves are gone.
Findings
Every conversation ends as evidence.
Each interview is distilled into findings — typed, quantified, quoted — then clustered across the whole company. Illustrative output from a 140-person manufacturing audit:
Problemcluster C-2
Maintenance knowledge lives in one head
≈ 6 hours / week · 9 people · 2 groups
“When Marie is on holiday, the line waits. The manuals are from 2014.”
marked for reportconfidence: high
Opportunitycluster C-3
AI-drafted handover notes
≈ 4 hours / week · 11 people · 2 groups
“If the system pre-filled the handover from machine logs, I’d actually read it.”
under reviewconfidence: high
Quotes pass a name-redaction pass before display. Slices below the minimum group size are never shown.
Deployment
Your network. Your models. Your rules.
Some clients cannot let data leave their network — so Dopta does not need one. The whole platform runs on-premise with a local LLM, from invitation to report.
Your GPU server, your models
Runs on Ollama, vLLM, LM Studio or llama.cpp. Cloud providers are an optional per-client plug-in — never a dependency.
One PostgreSQL, nothing else
A single database with pgvector holds identity, content and embeddings. No Redis, no SaaS, no phone-home.
Model roles, switched in the console
Interviewer, extractor, labeler and embedder are configured per client. A running interview pins the model it started with.
Evaluated in 30+ languages before it ships
A scripted persona harness scores candidate local models on follow-ups, compliance and grammar across every supported language — the choice is evidence, not taste.
on-premise manifestdocker compose
- postgres 16 + pgvectortwo schemas, three roles
- ollama · local LLMon your gpu0
- dopta-webconsole + respondent UI
- dopta-workercampaigns, extraction, purge
No cloud account required. No code path requires the internet. Upgrades ship as a compose file and a migration.
Authorship
The model drafts. You decide.
The tool
- Collects every voice at once, in 30+ languages
- Clusters duplicate findings across groups
- Quantifies hours, frequency and reach
- Drafts the report you then edit
Never
- Evaluates individual people — ever
- Proposes solutions or promises anything
- Decides what matters for your client
- Publishes a word without your signature