Solo-built and shipped Flowmation — a production AI automation system for real-estate lead-to-cash. From problem to pipeline to a data-grounded optimization program: designed, engineered, and run by one person.
Agents lose deals to whoever replies first. A lead that waits hours is usually gone — but an agent can't leave a showing to answer every inquiry. Flowmation closes that gap: it responds to every inbound lead in under a minute, follows up until they answer, and runs the whole funnel — sourcing, engagement, and measurement — without manual work.
Everything. Problem discovery and positioning, the system architecture, the code, the outreach copy, the experiment design, and the analysis. I defined what success looked like, built the thing that pursued it, and instrumented it so the results could be read honestly.
A structured experimentation program — hypothesis-driven, statistically framed (sample sizes, base rates, expected outcomes), one variable at a time — so the funnel improves on evidence rather than instinct.
Each of these is a working component I designed and shipped — not a mockup. Together they form the pipeline above.
An inbound lead triggers a personalized reply that's written and delivered in under 60 seconds — webhook → Claude API → Gmail. Built and verified end to end, so no inquiry ever sits cold.
Routes leads to isolated variants by cohort while holding the message body as a controlled constant — so a subject-line test measures the subject line and nothing else. Designed for clean causal reads.
A documented, repeatable method that sources targeted leads and validates 100% of them for deliverability before a single send — with query variations rated by yield so the process compounds.
Time-triggered follow-ups that keep going until a lead replies — then stop the instant they do. Ships with a kill switch, failure logging, and an audit trail that reconciles every send. Fails loud, not silent.
One controlled content pipeline across web and email — generation, brand consistency, and versioned publishing — so what ships is on-brand and on-message everywhere, every time.
Pulled the LLM out of the send path when a fixed template served the experiment better and cheaper. Knowing when not to use the model is as much the job as knowing when to.
The habits that show up in everything above — and the reason the work holds up when you look closely.
Every call framed with the math — sample sizes, base rates, expected outcomes. Forecasts before results, so we optimize on evidence, not vibes.
Clean experiments only. When a subject line is under test, the body is a controlled constant — so a result actually means something and can be trusted.
Nothing ships on trust. Live test-sends, config checked against source, audit trails reconciled. "It should work" isn't the same as "I watched it work."
Kill switches, pristine rollback baselines, failure logging, syntax-checked deploys. Systems designed so the expensive mistakes are hard to make.
I'd like to help you ship it — and measure whether it worked.