I build the systems
revenue teams run on.

Problem

A revenue operations practice that argues systems fail quietly has to prove it on its own property. Every part of this had to capture before it called out, fail loudly when it broke, and be checkable from outside.

What I built

Six pages on Next.js, a form that writes capture-first to Postgres before any external call, a Discord notifier that fires only after the row exists, and a site assistant running through n8n with structured output and full conversation logging. Every external call has its own error path.

Result

A submission survives a database outage, a model timeout, or a failed notification, because the record is written before any of them run. Failures return a real error rather than a success the visitor can’t act on. Live at purviewops.com.

Problem

A form fill is not a qualified lead. It is only worth acting on if it is real, reachable, and ranked. Most intake skips that work, so junk contacts flood the CRM, reps chase bad emails, and nobody knows which lead is actually worth a call. The qualification has to happen before anything hits the CRM, not after.

What I built

An eight-stage pipeline that takes an inbound form submission and turns it into a verified, scored, CRM-ready lead. It matches the lead to its campaign, enriches it through a per-campaign configurable path (Apollo, UpLead, or both), dedups, verifies the email, scores and grades on firmographics and reachability, assigns outreach sequences, and only then syncs the clean record to the CRM. It finishes with an AI mini-audit that scrapes the prospect’s site, checks their visibility, and drafts a personalized email.

Result

Only verified, scored, routed leads reach the CRM, so the data layer stays clean by design. Ran live, and the same enrichment-and-scoring architecture also handled standalone lead sourcing at volume. Each stage logs a summary, so a failure surfaces instead of silently dropping leads.

Problem

A wellness and medical-tourism business captured leads at events but had no system behind it. Leads sat unworked, attribution was lost, and value stalled at the entry stage. The constraint was conversion infrastructure, not lead volume.

What I built

An end-to-end capture and nurture system across two isolated sub-accounts: QR-to-form intake, standalone email-consent gating with DND suppression, interest-based routing into vertical pipelines, a discovery-call branch for high-intent leads, multi-email nurture sequences, and live dashboards with smart lists for attribution.

Result

Live at a single 2,500-person event. 107 leads captured without losing one, 81 opted in to email marketing, and 29 requested a discovery call, routed to a closer in real time. Built and debugged solo, in three days, on two days notice.

Problem

Building a medical knowledge base by hand is slow and risky. A paid specialist wrote cards one at a time, and nothing checked whether an answer was actually supported by its source before it went live.

What I built

An eight-stage pipeline that generates cards across three content types, runs every one through a compliance auditor, then tests its own output: a separate model answers questions from each card and a judge scores those answers against the source. Only cards that clear the threshold auto-publish.

Result

Publication was gated at an 80% pass threshold and 95% of cards cleared it, so roughly 5% routed to manual review rather than shipping. The work replaced a paid outside specialist. Anything flagged unsafe never publishes.

Problem

Reps decide where to spend their week on vibes. Which open deals are actually heating up and which are quietly dead looks the same in a pipeline list, so time goes to the loud deal instead of the winnable one.

What I built

A model that learns the behavioral fingerprint of past won and lost deals, then scores the open pipeline against it. The signal weights are derived from each signal’s real separating power, not hand-picked, and the model is validated on held-out deals so it proves rather than memorizes.

Result

A ranked call list with a plain-English action per deal. Each row shows the signals behind the score, so a rep sees why a deal is hot, not just a number.

Runs on standard CRM engagement data. The queries and validation are in the repo.

Problem

Deal health lives in call notes nobody re-reads. A rep can miss that a deal is single-threaded, has no timeline, or that the buyer is quietly weighing an in-house build, until it’s too late to act.

What I built

A one-click audit that reads a deal’s call history, scores its risk, flags the red flags, and writes the assessment back to the record. Every run is logged with cost, latency and outcome, so the system can be checked, not just trusted.

Result

Reps see risk before the deal goes quiet. The score is advisory and the rep still runs the deal. Failures surface to a human queue instead of disappearing.

Then I built the queries that could prove the whole thing useless, and the report they generate. Both are in the repo.

Problem

Good outbound is an intelligence problem, not a volume one. Writing every email from scratch ignores what the market already proves works, and generic personalization at scale reads like exactly what it is. The hard part is knowing which angle fits which lead.

What I built

A three-stage agent pipeline. It monitors 200+ competitor newsletters and distills them into a deduplicated library of proven strategies using vector search, matches each lead to the best-fit strategy semantically, and drafts personalized outreach with a two-stage writer that generates then edits. Formatted to read human, with a full multi-touch sequence per lead.

Result

Competitor sources analyzed into a deduplicated, evidence-weighted strategy library, and lead-specific campaign plans with subject lines, angles, offers, and cadence generated end to end. Built, tested, and validated, and designed to move to a new vertical by swapping prompts and sources.