Magna Collective
Book a callGet in touch
CRUZ FLORES IV · PHOENIX, ARIZONA · AVAILABLE FOR ENGAGEMENTS

WHITEBOARD
TO REVENUE

I design and build the data and AI systems companies run on — ETL pipelines, warehouses, and the agent and reporting layers above them. Principal engineer on every engagement, from architecture through production.

Book a call →Client workWhat I do
Directcruz@magnacollective.xyz
Phone(480) 930-0515
Start an engagement24H RESPONSE
SENT DIRECTLY TO CRUZ · NO MAILING LIST
Companies run on it
240+
Portfolio scale, already in production
Engineer, start to finish
1
No handoff, no account manager
First response
24h
Direct to the person who builds it
Data egress required
0
Private and on-device when you need it
(01)What I doFIVE ENGAGEMENT TYPES

One engineer across the whole stack. Nothing is lost in a hand-off.

01

Fractional CTO

Founders without a technical counterpart · Teams whose roadmap has stalled

I own the architecture, the sequencing, and the trade-offs — then build the system myself. You get an engineer accountable for both the decision and the implementation, not a strategy document.

System architecture
Build vs. buy
Technical hiring
Code review
Roadmap ownership
02

Data Engineering

Data spread across systems nobody can query · Reporting that takes a week

ETL pipelines, data warehouses, and the analytics layer above them. I consolidate CRM, billing, and product systems into one queryable model, then build the reporting your operators use without asking engineering first.

ETL pipelines
Data warehousing
Source-system connectors
Unified data models
Analytics + dashboards
03

AI Enablement

Teams with AI licenses and no measurable change in output

I make your existing team faster with AI, rather than replacing them with it. Internal assistants wired to your real systems, retrieval over your own knowledge, and hands-on training — I write AI engineering curriculum, so the teaching is the same work I do.

Internal copilots
Company knowledge RAG
Slack + Teams agents
MCP tool servers
Team training
Adoption metrics
04

Agent & LLM Systems

LLM projects that never leave the demo

Agent infrastructure, retrieval, and LLM analytics taken to production with evaluations, cost controls, and human approval gates — including on-device inference where the data cannot leave the hardware.

Agent infrastructure
RAG + retrieval
Evaluation harnesses
Cost + token controls
Private + on-device inference
05

Product Engineering

Software that has to ship to real users on real platforms

Native Swift and Kotlin for mobile, Next.js on Postgres for everything server-side. Offline-first architecture, multi-tenant backends with genuine isolation, and delivery through the App Store and Play.

Swift · SwiftUI
Kotlin
Next.js · TypeScript
Multi-tenant backends
App Store + Play
(02)AI & LLMIN DEPTH

Most AI work stalls in the same three places. Adoption, reliability, and data residency.

The usual problem
Licenses are bought and a pilot gets built, but nothing in the workflow actually changes. Usage decays within a quarter.
What I do about it
Wire the models to your real systems, put a human gate where the risk is, and measure whether output moved.
How it's taught
I write AI engineering curriculum, so enablement sessions are the same work, taught — not a vendor deck.
A

Team AI Enablement

Getting an existing team to actually produce more with AI. The licenses are usually already bought; the workflow is what's missing.

  • Internal copilots wired to your CRM, warehouse, and document stores — not a chat window that knows nothing about you.
  • Retrieval over company knowledge: docs, tickets, call transcripts, contracts, and wikis, with citations back to the source.
  • Slack and Teams agents that complete work — draft the follow-up, update the record, pull the number — rather than answering trivia.
  • MCP tool servers so the team's assistants can safely reach internal systems through a reviewed interface.
  • Prompt and context standards so output is consistent across people instead of depending on who wrote the prompt.
  • Hands-on training. I produce AI engineering curriculum; the sessions are the same work, taught.
  • Adoption measurement — who is using what, and whether it moved a number worth moving.
CopilotsRAGSlack APIMCPTraining
B

Production Agent Systems

Agents that run unattended against real systems, with a human in the loop exactly where the risk is.

  • Autonomous agent loops with tool registries and structured, reviewable execution traces.
  • Human approval gates on anything that spends money, contacts a customer, or writes to a system of record.
  • Multi-step workflow agents that carry state across tools instead of one-shot prompting.
  • Voice agents on a Twilio, Deepgram and ElevenLabs pipeline — inbound qualification, booking, and follow-up.
  • Scheduled and event-driven agents with cron, webhooks, and durable retries.
  • Vision and document processing — classification and structured extraction from images, scans, and PDFs.
  • Observability: every call, tool invocation, and token logged and attributable.
Agent loopsTool registriesVoiceVisionWebhooks
C

Reliability, Evals & Cost

The difference between a demo and a system somebody is willing to put in front of a customer.

  • Evaluation harnesses — regression suites for prompts, so a model or prompt change cannot quietly degrade output.
  • Structured outputs and schema validation so downstream systems can rely on the shape of a response.
  • Guardrails on refusals, hallucination surface, and prompt injection through untrusted content.
  • Token and cost optimization: caching, context trimming, and routing cheap work to cheap models.
  • Model routing and fallback across providers so one vendor's outage is not yours.
  • Latency budgets — streaming, partial rendering, and speculative calls where the wait is the product.
  • Cost attribution per feature, per tenant, per customer.
EvalsGuardrailsStructured outputsCost control
D

Private & On-Device AI

For data that cannot go to a third-party API — regulated, confidential, or simply not worth the exposure.

  • Local inference with llama.cpp and GGUF models on Metal and Apple Silicon.
  • On-premise model serving over a private Tailscale network, with no egress to a vendor.
  • Fully on-device agents. I built Pegasus: an autonomous agent running entirely inside the iOS sandbox — no server, no jailbreak.
  • Local speech-to-text with whisper.cpp for transcription that never leaves the machine.
  • HIPAA-conscious architecture where the data's residency is a design constraint rather than a policy note.
  • Open-weight model selection and quantization tuned to the hardware you actually have.
  • Upstream fixes — I've contributed llama.cpp tool-calling fixes to make local agent execution reliable.
llama.cppGGUFMetalwhisper.cppTailscaleHIPAA
(03)Work2025 — PRESENT

Three engagements in full. The brief, the build, and what came of it.

RevCentric.AI

Head of AI & Data Systems
Aug 2026 — Present
The brief

RevCentric.AI is an AI-native B2B outbound agency: it designs and runs outbound programs for B2B SaaS and service companies, pairing SDR talent with an AI stack so clients don't carry the cost and ramp of an in-house team. Activity moves across three outreach channels and half a dozen GTM tools, and every claim about what produced a meeting has to survive a client asking.

What I built
  • Own the AI and data architecture end to end — source-system ingestion, through the agent layer, to the reporting a client actually sees.
  • Ingest engagement data across all three outreach channels and the surrounding stack — Clay, Apollo, Instantly, Trellus and LinkedIn Sales Navigator — into a single model.
  • Build the unified data model consolidating CRM, billing, and product data into one queryable revenue layer for operators and finance teams.
  • Ship LLM-powered analytics and conversation intelligence to production with evals, cost controls, and human-in-the-loop approval gates.
Outcome
  • Attribution that holds up under scrutiny: activity → conversation → meeting → pipeline, traceable per client.
  • The firm's $10M+ generated pipeline and 500+ booked meetings report through the data layer I own.
  • Daily and weekly client reporting runs off one model instead of per-tool exports.
3
Channels unified
6
GTM systems ingested
$10M+
Pipeline under reporting
1
Queryable revenue layer
IngestionConversation intelligenceAttributionEvalsAgents

Phoenix Strategy Group

AI & Data Systems Engineer
Aug 2025 — Present
The brief

Phoenix Strategy Group runs full-stack finance and revenue operations for founder-led companies — fractional CFO, M&A, and investment banking across a book of 240+ portfolio companies. Client financial and revenue data sat in disconnected systems, so every reporting question became a manual export.

What I built
  • Built the firm's data engineering practice from nothing — ETL pipelines, data warehouses, and the analytics and dashboard layer that now appears in their service catalog.
  • Wrote the source-system connectors — HubSpot, Salesforce, Square, Zoom, and custom REST and webhook integrations — consolidating fragmented client systems into a single queryable revenue model.
  • Delivered a multi-tenant backend rewrite for a health-tech client platform, including tenant isolation and access control.
  • Productized the AI work into a sellable line: voice and CRM automation, white-labeled sales intelligence, and autonomous lead-generation agents.
Outcome
  • Data engineering is now a standing service line at the firm — pipelines, warehousing, analytics, and dashboards.
  • Led the firm's first open-source release.
  • Moved a fractional CFO practice into shipping AI products.
  • Negotiated the AI practice's commercial terms with the managing partner: monthly retainer plus 60% of net revenue.
240+
Portfolio companies
5+
Source systems unified
1st
Open-source release
60%
Net revenue share
ETLData warehousingAnalyticsMulti-tenantAgents

Growth Capital

Head of AI & Data Systems
Apr 2026 — Aug 2026
The brief

Growth Capital needed an internal dashboard and the data layer beneath it, with executives waiting days on numbers that had to be assembled by hand. The work began as a defined flex engagement with a single deliverable.

What I built
  • Designed and built the data layer and internal dashboard as sole engineer, working directly with executive stakeholders rather than through an analyst layer.
  • Modeled the underlying data so operators could answer their own questions instead of routing every request back through engineering.
  • Set the firm's AI strategy and then implemented it — the same person on the whiteboard and in the repository.
Outcome
  • Expanded from the initial flex engagement into Head of AI & Data Systems.
  • Converted the work into a recurring engineering and SaaS retainer.
1
Flex scope to retainer
4mo
To expanded mandate
0
Layers to the executive
Data layerDashboardsAI strategyExecutive stakeholders
Founder & product roles
2025 — Present
Magna Collective
Founder

Agent architecture, automation, and custom LLM systems for clients in Phoenix, San Francisco, and New York. Principal engineer on every engagement.

Aug 2025 — Present
Save Our Souls
Co-Founder & CTO

Built the native iOS (Swift) and Android (Kotlin) applications for a digital wellness platform. NFC-activated Devotion Mode on an offline-first architecture requiring no login or connectivity, with group infrastructure and shared analytics. Owned product and technical strategy through beta, including WCAG 2.1 compliance.

Feb 2025 — Jun 2025
Monograph
AI Automation Strategist

Built a prompt-driven sales toolkit that lifted conversion for the AEC SaaS team, and applied LLMs to lead qualification, call summarization, and CRM enrichment.

EARLIERCARVANA · CRUZ BUYS HOMES · ANODIZE CAPITAL · EDWARD JONES · C4 MOBILEB.S. FINANCE · ARIZONA STATE UNIVERSITY
(04)StackIN PRODUCTION USE
Data & Pipelines
Postgres · BigQuery · ETL and pipeline design · data warehousing · data-lake connectors · multi-tenant architecture · REST and webhook API design · Firebase
Integrations
HubSpot · Salesforce · Square · Zoom · Apollo · GoHighLevel · Slack API · custom REST and webhook
AI & LLM
OpenAI and Anthropic APIs · local inference (llama.cpp, GGUF, Metal) · RAG · evaluations · structured outputs · vision models · whisper.cpp · token and cost optimization
Agents
Autonomous agent loops · tool registries · MCP · Claude Code · Twilio, Deepgram and ElevenLabs voice pipelines
Languages
Python · Swift · Kotlin · TypeScript · SQL · Solidity
Web & Mobile
Next.js · React · SwiftUI · Kotlin · StoreKit 2 · RevenueCat · NFC
Cloud & Infra
Google Cloud (Architect-level) · Docker · Kubernetes · GitHub Actions · Vercel · Tailscale
Blockchain
Solidity · Foundry · Viem · Wagmi · Base/EVM · tokenomics · ZK attestations
Compliance
HIPAA-compliant system design · WCAG 2.1 accessibility · financial modeling
(05)How an engagement runs
01

Scope

A working session on what needs to exist and what it touches. You leave with an architecture and a build order, whether or not the engagement proceeds.

02

Build

Running software early and often — pipelines moving real data, agents making real calls. Progress is demonstrated, not reported.

03

Hand over

Into production with evaluations, cost controls, and approval gates where people belong. Documented so your team can own it, or I stay on retainer.

BUILT TO RUN
WITHOUT ME

Start an engagement →