Concierge build process
How a Parlei is built.
Every Concierge Parlei is hand-built to standard over 4–6 weeks. Not a template, not a fine-tune, not a wrapper around a system prompt. What follows is exactly what happens between the day we sign a partnership and the day your first speaker sees her Parlei for review. If you are evaluating this on behalf of a sponsor or event, this page is for you. Send it to whoever needs to see the work before they say yes.
What a Parlei actually is.
A Parlei is a conversational spokesperson tied to a real person. It answers with sources — specific pages, papers, talks, or positions the speaker has published — and never extrapolates beyond them. When a visitor asks something the source material does not cover, the Parlei says so. Every conversation produces a brief for the speaker: who asked, what they wanted, whether it is worth her time to follow up.
The technical shape underneath is retrieval-augmented generation with curated source material, structured metadata about the speaker’s positions and preferences, guardrails written per-speaker during voice tuning, and a scoring layer for conversation briefs. But that is not what makes it work. What makes it work is the intake interview and the review round — the human parts.
The four phases.
Week 1
Phase 1 — Intake interview
One conversation, 60–90 minutes, with each speaker. Recorded, transcribed.
Not a form. Not a Notion template. A real conversation with a real person about her work, her audience, and what she wants her Parlei to be able to do — and what she does not want it to say.
- Her body of work: what she has published, where, when. Papers, talks, book chapters, guidance documents, blog posts, panels, interviews. We assemble the source list here.
- Her positions: the frameworks she uses, the stances she holds, the distinctions she makes. Not opinions in a general sense — the specific things she would want a stranger to be able to ask about and get her actual view.
- Her audience: who typically comes to her after a talk, what they typically want. This shapes the routing logic — which questions get answered on the spot, which get flagged as "she should see this."
- Her limits: what she does not want her Parlei to speak on. Regulatory grey areas outside her practice. Political positions. Personal life. Client work under NDA. Everything her Parlei declines to answer is declared here.
- Her voice: how she talks. Directness, formality, favored phrases, things she never says. We tune to this in Phase 3.
Deliverable: A source manifest (list of documents/URLs/talks to include), a positions document (her declared stances), and a boundaries document (what her Parlei will decline).
Weeks 2–3
Phase 2 — Source assembly
Every source in the manifest gets ingested, indexed, and quality-checked.
We do not scrape the entire internet for her. We index what she has explicitly approved. The rule: if it is not in the manifest, her Parlei does not know about it.
- Published papers, PDFs, and long-form writing: extracted, cleaned, structured with metadata (date, venue, co-authors, key claims).
- Talks and interviews: transcribed if not already, indexed with speaker attribution (a Parlei only draws from what she said, not what an interviewer said).
- Blog posts, newsletter archives, published guidance: pulled in via URL manifest.
- Structured positions from the intake: added as first-class source material with the speaker’s approval.
Deliverable: A fully indexed, source-attributed body of material. Every retrievable claim has a citation back to a specific document, page, and timestamp.
Week 4
Phase 3 — Voice tuning
We calibrate how her Parlei talks against how she talks.
Voice tuning is where hand-building actually earns its cost. A template answers everything the same way. A tuned Parlei answers a curious-generalist question differently than it answers a fellow-expert question, and both sound like her.
- Register: formal / warm / direct / hedged. Different speakers land in different places.
- Length calibration: her preference for short definitive answers vs. longer contextualized ones.
- Framework language: the specific terms she uses (and the ones she avoids) when explaining her work.
- Signature moves: how she pivots when asked something outside her lane. Some speakers redirect; some say "not my area, here is who to ask"; some just decline. We capture her actual pattern.
- Boundary language: how she declines. Sharp / warm / self-effacing / procedural. All valid; all different.
Deliverable: A tuned system with 20+ tested conversation samples across a range of visitor types, reviewed against "would she have said it that way?" — ready for her review.
Weeks 5–6
Phase 4 — Review round
She sees her Parlei before her audience does.
The review round has a real red-pen phase. Nothing goes live until she has signed off on samples across all four visitor types.
- Sample conversations across visitor types — a curious generalist, a fellow expert, a journalist, a potential collaborator, a booking inquiry. She reads how her Parlei handles each and tells us what is right, what is off, what is not her.
- Her positions as her Parlei represents them — every declared stance played back. She can adjust, remove, or add.
- Her boundaries in practice — we show her exactly what happens when someone asks something outside her declared scope. She confirms the decline pattern matches how she would handle it herself.
- Her landing card — the /talk page a visitor sees when they open her Parlei. Photo, title, one-line intro, and the "welcome" question that greets visitors.
Deliverable: Her Parlei live at her URL. She has admin access. She can update her sources, her boundaries, and her landing card anytime.
What we don’t do.
This is a page for evaluation committees. So this section is direct.
- We don’t train on her data. We index her sources for retrieval and quote from them with citations. Nothing is fine-tuned into a model that persists beyond her account.
- We don’t share sources across speakers. Every speaker’s source material is isolated per-speaker within her account. Her Parlei cannot accidentally speak from another speaker’s material.
- We don’t extrapolate. The system is instructed and tested to refuse to answer questions its source material does not cover. "I don’t know" is a valid and often-used answer.
- We don’t give the sponsor admin access. The sponsor funds the build. The speaker owns the account. This is architecturally enforced, not politely honored.
- We don’t retain conversation content she doesn’t want retained. Retention is configurable per-speaker during intake. Default: transcripts are hers; briefs are hers; nothing is shared beyond what she opts in to.
- We don’t use her Parlei’s conversations to train a general model. Full stop.
What the sponsor sees.
The sponsor’s engagement report contains:
- Aggregate metrics per speaker’s Parlei: conversations initiated, questions asked (categorized by topic), routing outcomes (answered / flagged for follow-up / declined).
- Named-partnership attribution counts: how many visitors saw the "presented by [sponsor]" attribution on which speakers’ Parleis.
- Cross-roster themes: what the audience wanted to talk about across the featured lineup. This is the sponsor’s actual value — real signal about the room’s interests.
- Speaker-consented follow-up leads (Program+ only): if a speaker opts in, high-fit inquiries from her Parlei can be shared with the sponsor for post-event outreach. Opt-in per-speaker, not blanket.
The sponsor does NOT see:
- Individual conversation transcripts (those are the speaker’s).
- The speaker’s rolodex or contact database.
- Any admin surface on the speaker’s account.
- Anything a visitor said that the visitor has not been made aware would be shared.
Technical details for the security review.
Some evaluation committees have a security review. Here is what to expect.
- Hosting: Next.js on Vercel; Postgres and object storage on Supabase (both SOC 2 Type II).
- Data residency: default US-East. EU residency available on request (Program+ tier).
- Auth: email + password or SSO on request. All admin actions audit-logged.
- Model calls: stateless — no data retention at the model-provider layer.
- Retrieval source material: stored in the customer’s isolated Supabase schema. Isolated per-speaker within an account.
- Backups: daily, retained 30 days, encrypted at rest.
- Deletion: speaker-initiated deletion removes account data within 72 hours; verification available.
- DPA / BAA: available on request for MedTech / healthcare / regulated-industry customers.
- GDPR: EU-visitor data handled per GDPR; speakers control retention and opt-outs on their account.
- SOC 2: Type II certification is on the aspirational roadmap; underlying infrastructure (Vercel, Supabase) is already Type II certified.
If your evaluation needs specific answers we have not listed here, email us and we will respond.
Timeline, one more time.
For someone skimming: this is the compressed timeline.
| Week | What happens | Speaker time required |
|---|---|---|
| -10 to -8 | Partnership scoping conversation between organizer/sponsor and Parlei team | none |
| -6 | Speaker intake interview (60–90 min per speaker) | 60–90 min |
| -5 to -3 | Source assembly, indexing, positions/boundaries documentation | maybe 30 min for source list confirmation |
| -3 to -2 | Voice tuning + internal test conversations | none |
| -1 | Review round: speaker reviews samples, red-pens, signs off | 60–120 min |
| Event day | Live Parlei, speaker hands out link | in the moment |
| +30, +60 | Engagement report to sponsor, source updates from post-event material | none |
Total speaker time investment: ~2–4 hours across 6 weeks. All the rest is the Parlei team’s work.
Ready?
You have read enough. If Concierge is the fit for your event, the next step is a scoping conversation.
Start the conversation →Or, if you have specific technical or security questions that aren’t answered above, email us directly.