Selected work · 03

Walk into every first meeting already warm.

Big Talk turns public, B2B context into a 60-second briefing — three real hooks per person and one opener that bridges from human to business. We delivered the whole surface: brand, marketing site, product app, billing, and the AI pipeline that writes the brief.

Big Talk — marketing hero with a floating briefing card

Role

Design, engineering, AI pipeline

Services

Brand · Product UI · Web · AI features · Billing

Duration

Shipped 2026

View the site

Background

Back-to-back first meetings — and every one of them starts cold.

Sales reps, founders raising, recruiters, consultants — there are people whose week is a wall of first meetings with strangers. The prep ritual is always the same: fifteen minutes of tab-hopping across LinkedIn, company news, and podcast bios per person, usually compressed into a hurried skim in the elevator. Most people skip it entirely and open with the weather.

Big Talk's bet: the first minute of a meeting is where trust is won, and that minute can be prepared in the time it takes to pour a coffee. Paste who you're meeting — or connect your calendar — and get three genuine, sourced conversation hooks per person, plus one opener that bridges from human to business, gracefully.

Challenge

An AI product that has to be warm without being creepy.

A briefing tool that digs up personal trivia is worse than none — one wrong "hook" and the meeting starts with an apology. The product had to work from public, B2B-appropriate sources only, cite where every hook came from, and pass every candidate through a scoring gate before it reaches the user. And it had to be fast: a brief that takes five minutes to generate is a brief nobody opens before a back-to-back call.

Under the hood that meant a multi-stage AI pipeline with per-stage model selection, deterministic cross-referencing of shared facts for group meetings, calendar-triggered automation, multi-channel delivery, and workspace multi-tenancy with pooled quotas — all wrapped in a brand calm enough to be trusted with meeting data.

Approach

Three models, one pipeline.
Claude Haiku extracts and scores hooks, Sonnet cross-references the room, Opus writes the opener. Each stage runs on the smallest model that clears its quality bar — and a "cringe filter" gates every hook before it ships. The prompts are versioned Markdown, treated as the product's IP.
Briefings that arrive by themselves.
Connect Google Calendar and a brief lands ~15 minutes before every external meeting — by email or Telegram, at a lead time the user picks. Recurring internal syncs are skipped automatically. The best prep UX is the one you don't have to remember.
Calm as a brand position.
Fraunces serif, a warm amber palette, and a voice with a written rule: no urgency, no rocket emojis, no "10x". Even the 404 page reads "Awkward silence." In a category full of AI noise, the anti-hype stance is the differentiator.
Privacy as a feature, not a footnote.
A strict source allowlist, personal emails and phone numbers dropped at ingest, self-serve GDPR deletion, and Postgres row-level security for workspace isolation. The trust story is engineered, not claimed.

Highlights

Three Claude models in one pipeline
Haiku for extraction and scoring, Sonnet for the group Room View, Opus for the final opener — orchestrated per participant in parallel, with partial results streaming into the UI as they land.
~20 seconds per briefing
Briefings run as background jobs; each participant's row is written the moment it completes, so the user watches the brief fill in. Three hooks, one bridge, before the coffee is poured.
Four languages, every surface
EN, DE, FR, ES across the app, the marketing site, and billing — localized sitemaps and hreflang included. Stripe runs as catalog-as-code with deliberate USD/EUR parity pricing.

Screens