For $10M–$250M companies who need AI deployed to production — not slide decks promised. I build enterprise-grade applications at roughly 25% of the cost of a traditional agency engagement, shipped in weeks instead of quarters.
Most solo AI people have one or two of these. This work wants all three.
Act 01 · The Operator
From the Dot-Com Boom to the AI Boom.
My career has spanned both — I started and ran my own digital agency in the early commercial-web era, doing SEO, SEM, and web development, carrying the P&L through the dot-com cycle. Same operator instincts now applied to AI.
Act 02 · The Enterprise Practitioner
Fifteen years in Fortune 500 rooms.
Spent the last decade and a half implementing, integrating, and selling enterprise software at Fortune 500 scale. That quality bar is what I carry into every build.
Act 03 · The Builder
336k+ lines of AI code, in 2026 alone.
Voice, SaaS, iOS, enterprise dashboards, multi-model deliberation, call intelligence. Velocity comes from operator instincts plus modern AI tooling — not a five-person team coordinating through a PM.
Legal MarketingProductionActive ClientCustomer Data Platform
AI-powered legal intake across 230 law firms.
The Problem
The agency manages intake for 230 personal-injury and consumer law firms. Calls come in across 13 CallRail accounts. Contact forms from 230 firm websites piped into a single inbox. Their operations team was manually triaging thousands of submissions a week. Real leads were slipping through.
What I Built
A unified data platform stitching calls and forms to canonical firm identities across multiple source systems. Real-time webhooks from 13 CallRail accounts feed an identity resolution graph that maps numeric IDs, website domains, vendor slugs, and sender emails to a single canonical identifier per firm. AI enriches every record with paralegal-voice summaries, lead classification, and urgency scoring.
A deterministic routing engine auto-approves submissions with zero manual triage. A multi-tenant denylist structurally prevents shared notification senders from misrouting leads across firms.
Generative Engine Optimization across 5 AI platforms.
The Problem
Generative Engine Optimization is a real category most agencies can't speak to. Buyer-intent queries increasingly get answered inside ChatGPT, Perplexity, and Gemini — users never click through. GA4 labels nearly all AI-referred traffic as "direct." There's no established playbook, and the existing tools each solve only one piece.
What I Built
A SaaS platform combining three previously-separate capabilities: AI visibility scanning across 5 platforms in parallel, AI traffic attribution that recovers ~90% of dark AI traffic, and a Bayesian marketing mix model treating AI as its own channel.
13 different deliverables generate per scan. The system picks the right AI model for each one, then runs every output through automated accuracy checks before it can be downloaded.
Marketing ScienceIn-BrowserFirst of its kindProduction
Hill saturation curves, Bayesian attribution, and efficient-frontier optimization — modeling AI as a marketing channel.
How the AI Mix Modeler Works
What if you knew exactly where every marketing dollar was wasted?
The Problem
Adobe, Nielsen, and every legacy marketing mix model was built for a world where Google was the front door. When a visitor arrives from ChatGPT or Perplexity, the referral header is stripped — so traditional tools dump that traffic into "Direct" or "Unknown." The fastest-growing channel in marketing is invisible in every budget model on the market. No CFO can evaluate AI visibility as a line item, because no model puts it in the same view as everything else.
What I Built
A self-serve Bayesian marketing mix model that runs entirely in the browser. Hill saturation curves fit per channel, MAP estimation across the time series, lagged cross-correlation matrix for halo effects, 200 iterations of marginal reallocation to find the efficient frontier. Upload three columns — channel, spend, conversions — and the model produces cost-per-lead rankings, response curves with 80% confidence bands, halo synergies between channels, and a specific reallocation plan in under ten seconds.
The differentiator: AI visibility sits in the model as a first-class channel alongside Google Ads, TV, SEO, and paid social — modeled with the same math. AI discovery stops being a vague strategic initiative and becomes a measurable budget line.
Account intelligence and territory planning for a Fortune 500 field sales organization.
The Problem
A North American field sales team of 38 account executives, each managing roughly 80 accounts in their patch. Every Monday, every seller faces the same question: which five accounts to call this week, which to deprioritize, and what to actually say when they get on the phone. Salesforce can produce a list and a CSV. It can't answer those questions. Sellers were guessing.
What I Built
A single-operator platform that turns Monday-morning planning into minutes instead of hours. Three capability surfaces under one filter bar: territory hunting (drop a 15–50 mile radius around any city, see every account fully hydrated), one-click intelligence (executive briefings and internal sales plans generated in ~3 minutes from internal data plus live 10-K and 10-Q filings), and a manager-level coverage heat map for partner gaps and white-space planning.
0
Accounts under management
0
Field sellers using daily
~80
Accounts per AE
~3 min
Click to executive briefing
0
Core functionality stood up
$250M
Enterprise Sales Segment
Cloudflare WorkersKVAnthropic APIClaude Opus 4.6LeafletJavaScriptGitHub PagesWeb Search
06 — Case Study
What an AI training engine actually ships.
Enterprise TrainingAI Content EngineLive Demo
A complete partner-training portal — content, custom interactive modules, narration, exam, certificate — built in days, not quarters.
The Problem
Enterprise partner-training programs follow a fixed recipe: a team of instructional designers, an LMS vendor, video producers, and a project manager spend two-to-six months building the courseware. The output is good. The cost is in the six figures. For a fast-moving consultancy onboarding partners onto a new platform, that timeline kills the deal — partners need to be productive on the engagement before the curriculum even ships.
What I Built
A complete partner-training portal generated end-to-end by one AI engineer. 11 lessons across 3 modules (data platform, journey orchestration, B2B). 17 interactive modules — some styled as "official vendor tutorials," some as custom strategy walkthroughs — each with slide navigation, narrated audio, and chapter markers. 47 hands-on sandbox tasks. 11 knowledge checks. A 20-question accreditation exam with auto-generated LinkedIn certificate. Everything orchestrated from a single hub with progress tracking, roadmap, and zero LMS dependency. The client and vendor in the demo are fictional. The engine that produces the curriculum is real.
Clean code isn't what you write. It's what stays running.
Pre-Deploy GateProduction MonitoringDefense in Depth
Gated on the way in. Monitored on the way through.
01 · Pre-Deploy — Multi-Model Ensemble
The Problem
Most teams gate production with one reviewer or one tool. When that vendor regresses — and they do — single-tool buyers absorb the full hit. BridgeBench measured Claude Code accuracy dropping from 83% to 68% over six weeks in early 2026 before Anthropic published a postmortem. Anyone who depended on it for review lost fifteen accuracy points without knowing it.
What I Built
Four frontier-tier reviewer models run in parallel on every diff. Their outputs are synthesized into a single triaged review, calibrated against a benchmark I maintain — 35 production-shape bugs seeded into three Cloudflare Workers, weighted by severity, refreshed when any reviewer in the lineup is replaced.
Cost per pre-deploy check is roughly thirty cents. The ensemble catches what no single model catches — different reviewers have different blind spots, and stacking them closes most of the gap to the restricted frontier.
92%
Pre-deploy bug catch rate
Days → Min
Review turnaround per pre-deploy gate
53%
Fewer bugs reach production vs. single-tool review
Proof · May 15, 2026 audit cycle
On May 15, 2026, this discipline took a 6,000-LOC production Cloudflare Worker — the system serving 230 law firms — through a full audit, refactor, and security-hardening cycle. Five deploy batches, six commits, review executed through Delphi — my proprietary multi-model code-review tool. Every major finding triaged and shipped.
Findings addressed included a multi-tenant block-sender failure class silently losing forms for a firm's own inbound leads, a full transactional rewrite of the firm-merge function that would have damaged the first merge under the old code, an atomic counter race in sender-trust recomputation, X-Forwarded-For spoofing in the auth rate limit, a KV check-then-set race, hidden-firm session revocation gaps, and a half-dozen defensive hardenings.
Industry equivalent
3 engineers · 105 hrs
Formal-inspection review to match the same bug-catch coverage
This discipline
1 AI engineer · 8 hrs
Review, fixes, and deploy — end to end
02 · Post-Deploy — Scheduled Invariant Assertions
The Problem
Pre-deploy review catches most bugs but not all. The rest reach production. And production drift can't be caught at deploy time at all — it happens later, against real data the deploy never saw. Most operational monitoring fires after a customer complains. The feedback loop from "something broke" to "we know it broke" is usually hours to days long.
What I Built
An AI agent that runs every 15 minutes, asserts that the production system is doing what it's supposed to, and sends structured alerts to Nooma Engineering Managed Services on violations. It catches mechanism failures — the AI pipeline skipped calls it shouldn't, a sync watermark drifted, an invariant quietly broke after a migration — before they cascade into outcomes the client sees.
What separates this from typical monitoring is the loop. Detection without resolution is just noise. Every alert routes to Nooma Engineering Managed Services, who triages, fixes, and verifies — usually before the client knows anything happened. The system catches the bug; we close it.
This is the back end of a discipline that starts before deploy — every change goes through multi-model review in Delphi before it ships. Pre-deploy review, post-deploy assertions, managed resolution: a closed loop, end to end.
Three products, shipped solo. Worth a closer look.
Spotlight 01 · Marketing Science
A Bayesian Marketing Mix Model for the generative engine era.
The Faneros AI Mix Modeler is a self-serve Bayesian MMM with Hill saturation curves, adstock decay, halo detection, an efficient-frontier optimizer, and a scenario planner.
How the world's major AI assistants see your brand.
Faneros GEO scans five AI systems and produces thirteen ready-to-deploy deliverables. Two-pass knowledge-base architecture keeps contamination at zero.
Classify, score, and route submissions with dual-pass verification and learning trust lists.
02
Call intelligence
Transcription, enrichment, lead scoring, and conversion attribution across your phone system.
03
Custom dashboards
Role-based views, real-time metrics, drill-downs, AI-assisted search. Your domain, your brand.
04
Voice AI agents
Real-time conversational agents for inbound calls, booking, and qualification. Sub-500ms.
05
Classifier pipelines
Dual-pass LLM classification with verification, confidence scoring, and human-in-the-loop fallback.
06
AI visibility (GEO)
How do today's major AI assistants see your brand? Scan all five and get ready-to-deploy fixes.
11 — How I Work
No slide decks. Just working software.
Step 01
Discovery call
Twenty minutes. You describe the problem. I tell you on the call whether I can ship in two weeks.
20 minutes · no cost · no slides
Step 02
Two-week free pilot
Free. Fixed scope. Working prototype in your team's hands by day ten.
Free · 2 weeks · no strings
Step 03
Production deploy
Harden the prototype, deploy to production. Billing starts the day your team logs in.
1 week · bundled into retainer
Step 04
Monthly retainer
Hosting, AI costs, monitoring, improvements. Full source access. Thirty-day exit.
Month-to-month · the code is yours
If we can't build what you want in two weeks, you pay nothing.
"Fifteen years across the buyer side, the builder side, and the operator side of enterprise software. The quality bar doesn't change just because the delivery timeline got shorter — it's the whole point."
— The operating principle behind every build
12 — Pricing
Transparent terms.
Pilot
Free Pilot
Free · 2 weeks · scoped on the discovery call
Discovery & scope definition
Two-week build to working prototype
Production deployment & handoff
Infrastructure & domain setup
30-day post-launch monitoring
Recurring
Monthly Retainer
Month-to-month · 30-day exit · code is yours
Hosting & infrastructure
AI model costs (LLM inference, embeddings)
Uptime monitoring & incident response
Continuous improvements from your usage
Direct line for feature requests
Enterprise
Scoped Individually
Annual engagements · multi-system builds
Multi-system architecture
Custom SLA & uptime guarantees
On-prem or customer-cloud deployment
Dedicated capacity (not shared)
Quarterly business reviews
13 — Questions I Actually Get Asked
The things you're probably thinking.
Why one AI engineer instead of a team of developers, architects, and a project manager?
Because most of that team exists to coordinate the team. A traditional shop needs a PM to translate, an architect to design, and developers who each see one slice of the problem. I do all four roles in one head — no handoffs, no telephone game, no status meetings about status meetings. One person owns the outcome.
What happens if you get hit by a bus?
You own everything. Every line of code lives in your repo, on your infrastructure, under your accounts. Every project ships with a handoff runbook a competent AI engineer can pick up in a day. The bus risk is lower with me than with an agency that could go under mid-project with your code on their servers.
Isn't this just vibecoding?
No. Vibecoding is prompting an AI until something kind of works and shipping it. This is architecting the system first, writing tests, and reviewing every AI-generated function the way a senior engineer reviews a junior's pull request — because that's exactly what it is. The AI is the junior dev. I'm accountable for the code. See the Clean Code section — the receipts are right there.
Do you work with my stack?
Almost certainly. Preferred: Cloudflare Workers + D1 + frontier LLMs. Also shipped on AWS, Azure, and GCP, with integrations across Salesforce, HubSpot, Twilio, CallRail, Slack, Gmail, and most major APIs. If you're on something exotic, tell me on the call and I'll give you a straight answer about fit.
What about data & security?
Your data never trains a model. Data stays in your region, code lives in your repo, secrets live in your secret manager. Upstream providers are SOC 2 Type II compliant. On-prem available at Enterprise tier, BAAs for healthcare. Send me your security questionnaire — I've filled out plenty.
Can I talk to a current client?
Yes. After a discovery call and a real fit, I'll connect you with a client who's solved a similar problem — including the legal marketing agency in the case study above. References aren't handed out cold; it's not fair to my clients and it's a screening step both ways. Once we both know we're a fit, you'll talk to someone who's where you want to be.
14 — Next Step
Let's talk for twenty minutes.
Describe the problem. I'll tell you on the call whether I can ship something useful in two weeks, or whether you're better served by someone else.
