Russel Mei
De Jesus

The problem

• Nine manual steps per profile. Download the resume, pull the interview transcript, prompt a model, extract details, format the deck, upload photos, validate links, run QA, publish.
• Ten hours a month gone. At twenty profiles a month, assembly alone consumed a meaningful share of admin capacity.
• It gated matching. Every new assistant needed manual work before they could be shown to a client, which slowed time to bench and forced a tradeoff between profile quality and speed.
• Off the shelf AI could not do it. The deck tool's own AI generates presentations from scratch but cannot populate an existing template. Every profile came out structurally different, which is exactly wrong for a client facing document.

What I built

• A pipeline that generates a client ready profile from recruiting data that already exists, while holding template structure exactly.
• Two triggers. A meeting webhook fires on the final interview. An admin can also submit a candidate email manually.
• One fetch step pulls resume, transcript, video links, case study, headshot, and recruiter notes from three systems.
• Claude does the reasoning: extracting achievements, quantified metrics, a skills taxonomy, and career trajectory from raw material.
• A human gate stays. The recruiter reviews and approves before anything is client facing. Automation removed the typing, not the judgment.

The result

• Same day turnaround replaced multi day backlogs, so the matching bottleneck is gone.
• Every profile now follows one structure, which is what made the consistency problem solvable at all.
• Comfortable at twenty profiles a month today, with headroom to roughly five times that before the process needs rework.

The problem

• Every new client started with the same manual checklist. Create the Drive folder, hunt down the Energy Audit, EA Role Scorecard and Ideal Week templates, copy each one in, fix the sharing permissions on all three, then tell the team it was ready.
• The gap between paying and being set up was dead time. A client who paid on a Friday could sit unprovisioned until someone noticed on Monday.
• Sharing permissions were the quiet failure. A template copied but not shared looks finished and is not, and you only find out when the client cannot open it.
• Attribution was reconstructed after the fact. Referral source, marketing tag, and whether a discovery call happened before or after the campaign all lived in different tools, so revenue was credited from memory.
• Then the same pattern again at kickoff: the sales transcript and intake form read by hand, the pre call email and call script written from scratch, the role scorecard pre built, every time.

What I built

• Triggered by money, not by a person. Stripe fires on checkout completed, a filter drops anything that is not a current client, and the rest runs unattended.
• Provisioning, 10 steps of it. Find or create the client folder, then for each of the three templates: locate it, copy it into the client folder, and set its sharing preferences. The permission step is its own step every time, because that was the thing that used to silently fail.
• The team finds out automatically. Slack messages at both ends of the run, so nobody has to check whether provisioning happened.
• Attribution resolved in the same run. A Typeform lookup checks whether the referral field was filled, then an ActiveCampaign contact lookup drives three branches: tagged PPC, tagged something else, and whether a discovery call was booked before the campaign, after it, or never. Each outcome posts to its own webhook.
• Custom JavaScript where Zapier could not do it, for date formatting and for slicing the folder identifier out of a longer string.
• A second automation carries it into kickoff: pre call email and call script drafted from the intake form and sales transcript, then the follow up and role scorecard suggestions after the call, with a 20 hour reminder if the email is never marked sent.

The result

• Provisioning stopped being a thing anyone remembers to do. It happens on payment, at any hour, including the Friday night ones.
• Sharing permissions are set by the same step that makes the copy, so the "copied but not shared" failure cannot happen any more.
• Revenue is attributed at the moment it lands rather than reconstructed later from memory.
• The account manager role moved from document assembly to judgment. They still own what goes out.