Analytics · Growth · Measurement
$1,041,637
in recoverable revenue, found inside one billing platform’s failed-payment retry window. It took a query across 349,244 rows that nobody had run. The finding was arithmetic. The work was looking.
That is the whole through-line. Precision and recall rather than a demo. A metric that came back at zero before anyone reported on it. Nine statistical tests with zero significant results, written up as such. Two sources disagreeing about the same site by a factor of five, both published with the disagreement intact.
I build the systems that produce and prove the numbers: instrumentation, taxonomies, dashboards, churn models, and the content operations that turn finding into pipeline.
Eight years of it, for FormDR, Scale Insights, Net Atelier, HackingHR, DialMyCalls, CYE, Spikerz, Tasq and Willow, and before that agency-side delivery for monday.com, Guardio and CirclesUp.
Ordered by how often they are the actual problem.
Instrumentation design, event taxonomies, property audits, dashboard programmes, and migrations between platforms.
Defining activation, finding time to value, onboarding across plan configurations, lifecycle email, win-back.
Rule-based, ML and behavioural models, driver ranking, dunning and recovery, expansion surfaces.
SERP research, answer-box structure, and getting cited by AI engines. Priced per query, logged per call.
Knowledge bases, claims registries, lint layers, reviewer roles and publishing gates. The machine, not the tips.
Classifiers, evaluation sets, calibration, confidence routing, and telling the difference between a model and a regex.
Market selection, beachhead, ICP before TAM, pricing architecture, sales motion as a constraint rather than a choice.
Warehouses, pipelines, joins that survive contact with reality, statistical testing, and privacy as a design constraint.
Compliance architecture, test harnesses, and the plumbing that makes the other eight possible at volume.
Named where the relationship allows. Every figure here comes from the same registry the reporting did.
FormDR
PostHog instrumentation and a property audit, Stripe revenue intelligence, a churn-prevention flow built inside a vendor’s experience limit, and JTBD research across 60 recorded sales calls.
281 events · $252,139 MRR modelled · $1.04M dunning window
Scale Insights
Amplitude verification against source data, an eleven-dashboard programme built from Python, a ClickUp board generated through an agent, and a deterministic tagging engine with a model explaining each recommendation.
11 dashboards · 39 charts · 2,100 lines
Net Atelier
GTM strategy, a three-scenario financial model validated against industry benchmarks, a competitive teardown with verified pricing, and a board deck fact-checked claim by claim.
30 claims checked · 6 fabricated · 6 competitors priced
CYE
A search and AI-citation programme: per-piece SERP gap analysis, an evidence registry with verbatim quotes, and a research stack metered to the cent.
19 pieces gap-analysed · 285 keywords · whole programme under $2
HackingHR
Reconciling two systems that disagreed about the same customer base by 2,868 users, an ML churn model, and an engagement analysis that separated machine clicks from human ones.
2,868-user discrepancy reconciled · 45% reported against 18% real
DialMyCalls
The first behavioural baseline the business had ever had, built from raw billing and usage, and the leak it surfaced on day one.
$1,300/month leaking, found in about two hours
Not productised packages. The diagnostics that decide what the work actually is before anybody commits to it.
A product’s structure determines what is possible. A strategy that fights it fails without ever being wrong.
Market, motion and channel fit are constraints before they are choices. Most go-to-market plans are a motion chosen for reasons unrelated to the product.
Your pipeline is not broken. It is disconnected. The value is stranded in the seams between the layers, because nobody owns a handoff.
One check cannot catch six classes of failure, because the failures do not resemble each other. So the operation is layers, each with one job.