Independent Builder · AI Product Management Capstone, Product Faculty (8-week course)
The challenge
Independent dealers get leads from every marketplace, dumped unstructured into one Gmail inbox, with no tooling at all
What I did
Over an 8-week AI PM capstone at Product Faculty, built a 4-agent system, solo, that drafts and screens replies inside Gmail
Result
A 594-run, cross-provider model comparison cut pipeline latency 40%
"A 594-run, cross-provider model evaluation cut pipeline latency 40% — rigor that paid for itself in the metric that actually matters to a dealer: speed."
Problem
I'd already seen this problem up close — I shipped the AI-response feature described in Case Study 2 for mobile.de's largest, most professionalised dealers. But that tool only exists for dealers already inside a Lead Management System. 71% of independent dealers have no LMS at all — just Gmail, and marketplace leads arriving with no structure, no tracking, and no help. The average response time is 9.2 hours; 82% of buyers expect a reply within 10 minutes. The dealers who need this most are the ones no enterprise tool is built for.
Approach
Rather than another dashboard dealers would have to adopt, I built Carbii to meet them exactly where they already are — a Chrome extension that activates inside Gmail itself. Underneath, it's a 4-agent pipeline, not a single model call: a Guard agent screens every incoming message for manipulation before anything else touches it and fails closed on any doubt; a Conversation Agent drafts the reply and classifies intent signals; an independent Integrity Critic re-checks every draft against guardrails before a dealer ever sees it; a Bulk Reply agent batches genuinely similar leads.
I treated model selection as an empirical question, not a default: a 594-run comparison (9 models × 66 eval cases, cross-provider LLM judging so no provider grades its own family) drove the final production model choices — and directly cut the synchronous reply pipeline's latency by 40%. Nothing sends without the dealer's explicit approval.
Shipping the enterprise version taught me the pattern works. Building the independent version taught me the harder lesson — for the dealers with nothing, the bar isn't a better dashboard, it's zero behaviour change. Evaluation rigor wasn't academic overhead here — it's what actually found the 40% speed-up.