Michael Charo

Lead Software Engineer · Frisco, Texas

Michael Charo

AT&T Chief Data Office · Platform architecture and applied AI

I lead the engineering of production platforms and the AI agents that run on them. Over ten years at AT&T I have delivered the foundations other teams build on: access control, encryption, observability and release automation.

Currently technical lead for an enterprise fraud-prevention platform serving 400 to 500 fraud agents.

Platform architectureAI agents and MCPSecurity and access controlObservabilityTechnical leadership
View selected work LinkedIn GitHub Contact
$20M+savings credited to the fraud platform I lead
75%shorter release window with AI pre-flight checks
3 teamsadopted the encryption service I built
150+engineers at my talk on AI-assisted development standards

Selected work

Case studies

Internal AT&T systems are described by what they do rather than by name, and diagrams are redrawn. The talk and hackathon projects are public, with code.

Foundations for an enterprise fraud-prevention platform

AT&T · 2024–present
Problem

A platform used by 400 to 500 fraud agents needed security and operations that could keep up with new teams, new roles and frequent releases without being rebuilt each time.

Approach

Designed end-to-end authentication and role-based access across the React front end and FastAPI back end for 15 roles. Built a shared encryption microservice behind a simple REST API. Made OpenTelemetry tracing the team standard with reusable front-end and back-end instrumentation.

Result

The platform is credited with $20M+ in savings, a 30% cut in fraud losses and calls resolved about 100 seconds faster. The encryption service was adopted by 3 teams, and tracing cut mean time to recovery by 60%.

FastAPIReactPostgreSQLOAuth 2.0 / OIDCRBACOpenTelemetry

AI pre-flight review for production releases

AT&T · 2025
Problem

Releases took a full 8-hour window, much of it spent manually checking diffs for configuration and database changes that would break production.

Approach

As Git and release lead, I put an AI agent in front of each release. It reviews the diff for new variables, secrets, table references and PostgreSQL index differences before deployment starts.

Result

The deployment window dropped 75%, from 8 hours to 2, and deployment-related incidents fell 85%.

AI agentsCI/CDRelease engineeringPostgreSQL

An AI governance intake agent where code makes the decision

AT&T · 2026 · Spec
Problem

Teams asking to use AI filled out a long governance form, and reviewers worked out by hand whether a request could reuse an approved solution or take a faster review path.

Approach

I wrote an implementation-ready spec for an intake app. People describe their system in chat or draw it. The LLM only holds the conversation and extracts structure. Deterministic code matches technologies to a catalog, applies the track rules and keeps the audit trail.

Result

Every request ends in a Reuse, Fast Track or Standard decision that can be traced back to a rule, and the system suggests legitimate changes that qualify for a faster track.

LangGraphFastAPINext.jsTypeScriptReact FlowAgent design

6.5 minutes, 1 GPU, 151 intents

2026 Data Science Summit · Talk
Problem

Many teams send every chat message to a frontier model just to decide where it should go. That costs money and time on every request.

Approach

Fine-tuned Qwen3-1.7B with 4-bit QLoRA on the public CLINC150 dataset, on one consumer GPU, and benchmarked it against zero-shot frontier models on a sealed 5,500-row test set. Added a confidence threshold so the model escalates instead of guessing.

Result

96.8% accuracy, 2.2 points above a zero-shot frontier model, at 1.2 ms per request. Training took 6.5 minutes. The threshold raised out-of-scope detection from 69% to 92%. A benchmark, not a production deployment.

Qwen3-1.7BQLoRAFine-tuningLLM evaluationIntent routing
Code on GitLab

ATLAS: guardrails for an agent you can't modify

TM Forum Innovate Americas 2026 · Hackathon
Problem

The Trustworthy AI & Data challenge: prove an AI agent stays inside its limits on an AWS agent platform the team was not allowed to change.

Approach

With a team of AT&T engineers, built in three days an inline guard that refuses over-cap spending, a control-test tool that judges the platform's own records against frozen, versioned thresholds, and a closed loop that withdraws the agent's spend cap after a breach.

Result

A working demo judges drove live, returning PASS, BREACH or NO EVIDENCE for each control, with the full architecture written up and published.

AWS BedrockAgentCoreCedar policiesStreamlitAI governance
Code on GitHub
$1.7M/yr

A computer-vision check of buried-wire installation photos, shipped 60 days from kickoff and validated by field-operations finance across about 19,000 dispatches.

90–140×

A core API endpoint cut from 30–40 seconds to a third of a second with one MongoDB aggregation, connection pooling and Gunicorn.

0 critical

findings in the security and records audit I led for an enterprise conversational AI platform, including encrypting 680K stored documents with no data loss.

Experience

Career timeline

Shown as a distributed trace, the way I read production systems: AT&T is the root span and each role is a child span.

Span
2014201620182020202220242026
Charge Nurse / SupervisorNix Behavioral Health
1y 6m
B.S. Computer ScienceUTSA
2016
IT ContractorIntech Southwest
1y 4m
AT&Troot span
10y 5m
Associate Applications DeveloperTechnology Development Program
2y 2m
Senior Specialist Data Scientist
2y
Professional Software Engineer
3y 7m
Lead Software EngineerFraud Prevention
2y 8m
M.S. Artificial IntelligenceUdacity Institute of AI & Technology
in progress

Lead Software Engineer

AT&T · Mar 2024 – present

  • Technical lead for an enterprise fraud-prevention platform, owning architecture, security, observability and releases.
  • Built agent proofs of concept (an MCP server on FastAPI, a browser-automation agent, a LangGraph desktop agent) that were adopted into the company's internal agent platform.
  • Co-designed company-wide standards for AI-assisted development and presented them live to 150+ engineers. Helped dozens of developers adopt GitHub Copilot.
  • Designed a plug-in task engine on the existing PostgreSQL store, so new task types ship with zero additional storage.
  • Weekend and nightly on-call; kept the platform within SLO through a high-profile device-launch weekend.

Professional Software Engineer

AT&T · Aug 2020 – Mar 2024

  • Led the security and records audit of an enterprise conversational AI platform to zero critical findings.
  • Designed three-environment blue/green deployments with instant rollback, keeping the platform above 99.9% availability.
  • Merged 135+ pull requests from 17 engineers in one quarter as UI code owner, coaching on design and performance.
  • Ended hourly forced sign-outs on two AI platforms with silent token refresh, and shared a sign-in component neighboring teams adopted.
  • Extended a computer-vision field-quality platform ($3M cash avoidance) and owned an employee self-service support platform that grew visitors 4.5× and automated 15,140 password resets.

Senior Specialist Data Scientist

AT&T · Aug 2018 – Aug 2020

  • Built and shipped the computer-vision wire-installation check in 60 days, saving $1.7M a year with statistically significant results.
  • Lead developer of an internal chatbot-builder platform and SDK adopted by 6 business units.

Associate Applications Developer

AT&T · Jun 2016 – Aug 2018

  • Made a VoLTE network-monitoring query about 90% faster and added per-call and per-session analysis.
  • Built a data emulator that let development start before data access was granted, plus reusable REST services; a proof of concept built with operations reached production.

Speaking

Talks and recognition

  • 6.5 Minutes, 1 GPU, 151 Intents2026 Data Science Summit
  • TM Forum Innovate Americas hackathon, Trustworthy AI & Data2026 · team ATLAS
  • Archetypes and workflows for AI-assisted coding2025 · company GenAI forum, 150+ engineers
  • Zero-touch automation and data-driven development2021 · company technical symposium
  • Top 10, company security capture-the-flag2024

Toolbox

Skills

AI and agents
MCP, LangGraph, LangChain, LoRA / QLoRA fine-tuning, LLM evaluation, RAG, computer vision
Backend
Python, FastAPI, Flask, Node.js, REST, GraphQL
Frontend
TypeScript, React, Next.js
Data
PostgreSQL, MongoDB, Snowflake, Databricks / Spark
Cloud and ops
Azure, AWS, Docker, CI/CD, blue/green, OpenTelemetry, SLOs
Security
OAuth 2.0 / OIDC, JWT, RBAC, encryption services, SAST

Education

Education and certifications

  • M.S. Artificial IntelligenceUdacity Institute of AI & Technology · in progress
  • B.S. Computer Science, Software EngineeringUniversity of Texas at San Antonio · 2016
  • B.S. NursingTexas A&M University–Corpus Christi · 2012
  • Databricks Certified Associate Developer for Apache Spark 3.02022 · plus Azure Fundamentals and Udacity nanodegrees in generative AI, AI programming and full-stack web

Contact

Get in touch

I welcome conversations about principal engineering, AI agent and forward-deployed roles. Email is the fastest way to reach me.

Emailmichaelcharojr@gmail.com
LinkedInlinkedin.com/in/michael-charo
GitHubgithub.com/Mikey778