01 / Manta Dive · dive tourism
Cecilie, an AI booking assistant that never sleeps through an enquiry
The problemThe team was manually answering the same Instagram DM questions about courses, pricing, and certification requirements around the clock, with no system to capture leads overnight.
What I builtCecilie: a booking assistant on ManyChat and Claude that handles Instagram enquiries 24/7 and hands off to the team via WhatsApp once a booking is ready to close.
- Answers against the full SSI course catalogue (Try Dive to Rescue) with real pricing and certification requirements
- Collects name, activity, date, group size, and WhatsApp number through natural conversation
- Writes every lead into a Google Sheets waitlist, tagged by date and source, ready each morning
- Never confirms bookings or takes payment, and routes medical questions straight to a named human
ResultSold as a productized service: build fee, 2-week trial, monthly retainer. Now Manta Dive’s first line of response, 24/7.
02 / Springbolt · SaaS support
A support agent that verifies before it ever answers
The problemThe support inbox mixed real customers with noise across Gmail and Intercom, and every decision to auto-reply or escalate carried real cost if it got it wrong.
What I builtAn AI support agent that verifies who’s actually a customer before responding, only ever cites real knowledge-base articles, and routes replies by confidence.
- Scores every inbound message against product keywords, filtering out non-customers before any AI budget is spent
- Searches Intercom’s live knowledge base, explicitly instructed never to invent an article URL
- High confidence auto-sends · medium drafts for manager review · low just logs
- Escalation pings a human on Slack with full context, but only when urgency and low confidence coincide
Result24-hour turnaround on support that used to need manual triage, with escalations reserved for what actually needs a human.
03 / comme_une · agency sales
Gmail follow-up infrastructure in the founder’s exact voice
The problemThe founder was manually replying to every warm lead and manually tracking who needed a nudge, with high stakes on getting pricing language exactly right, every time.
What I builtTwo connected systems: an AI reply-drafting workflow matched to the founder’s brand voice, and a 6-stage automated follow-up sequence.
- Decides per-email whether a proven template fits or the message needs a fully custom response
- Uses the founder’s exact pricing and process wording, consistent across hundreds of replies
- Every AI draft lands in Gmail for human review. Never auto-sent
- Follow-ups only nudge after 24+ hours of silence and never talk over an active conversation
ResultThe DM and follow-up load automated, and no qualified lead falls through the cracks.
04 / comme_une · content strategy
YouTube audits: a week per client, down to 20 minutes
The problemManually auditing a channel’s titles, thumbnails, and scripts against a content strategy took the team a full week of work for a single client.
What I builtAn automated audit system using Claude’s vision and text models.
- Scores thumbnails against exact brand guidelines (colors, fonts, text limits) using Claude vision
- Scores scripts against a Hook / Intro / Body / Outro framework, quoting the actual transcript
- Delivers an executive summary and the “5 highest-impact changes”, automatically
ResultA week of audit work per client cut to about 20 minutes, running for multiple clients simultaneously and delivered as a Google Doc link via email.
05 / Life Is A Beach Party · group travel
Lucy, an Instagram sales agent with a real memory
The problemA high volume of enquiries about Greece group trips needed qualifying (destination, dates, group size, country) before sales could act. Leads were going cold overnight.
What I builtLucy: a Claude-powered Instagram DM sales agent with persistent memory across conversations.
- Voice shaped through careful prompt engineering: friendly, casual, and tuned to the brand’s 18 to 25 year old audience
- Remembers the full conversation via Supabase and never re-asks a question the customer already answered
- Extracts booking details naturally, accepting “June” as a complete date
- Once qualified: creates the lead in Zoho CRM, updates ManyChat, emails the team an alert
ResultQualified leads reach the CRM automatically, with zero manual data entry and zero overnight leads lost.
06 / Gloeilamp Goedkoop · e-commerce
An SEO engine that writes in the brand’s own voice
The problemFinding real keyword opportunities and producing SEO content across multiple markets, without it reading as generic AI copy.
What I builtAn SEO automation system for founder Edward Bartels.
- Discovers keyword opportunities via DataForSEO, scored by search volume against difficulty
- Scrapes the client’s own live site, where Claude extracts their real technical terms, categories, and brand phrases
- Generates SEO content per keyword, per target language, matched to the brand’s actual voice
ResultOn-brand, multi-market SEO content generated automatically, grounded in the client’s own site, not templated copy.
