Carlos Garcia Reyes
Get in touch

Carlos Adrian Garcia Reyes · Lead Software Architect & AI Platform Leader

I architect AI platforms that hold up in production.

10+ years taking enterprise systems from zero to production and owning them once live. I lead the platform, the AI strategy, and the engineers behind them, so teams ship faster and AI stays secure, governed, and worth what it costs.

Email meDownload resume

Open to Staff or Principal Architect, Head of Platform, and AI Platform leadership roles. Remote, U.S. time zones. Select consulting engagements.

Now
Lead Software Architect and AI Platform Engineer, GlobalLogic / 95 Percent Group
Scale
200K+ daily active users across 300+ U.S. school districts
Team
~20-engineer platform organization across U.S. and Mexico
Before
Kavak and Clip, two $1B+ venture-backed startups

Impact

What changed after I took ownership

  • Production release cycle

    Platform tooling and delivery pipelines that removed hand-offs, with DORA metrics reported to the CTO and VP.

    Every 5 weeks
    Every 3 days

  • New report delivery

    A reporting agent system built on a governed catalog of business concepts.

    2–3 week sprint
    3-day catalog change

  • District roster processing

    A focused Go rewrite of the slow path, which had been capping how fast districts could onboard.

    Hours for one district
    ~20 districts in minutes

  • AI costs

    Model right-sizing, temperature tuning, and prompt optimization, backed by token-cost agents, per-role limits, and monitoring.

    60% lower

  • Core microservices refactored

    AI-accelerated refactoring, scoped through client discovery workshops I led.

    23 in 45 days

Selected work

The decisions behind the numbers

95 Percent Group, via GlobalLogic · 2026 · Proposal, semantic layer, query engine, agent, and cost controls

Ask-a-question reporting

Support agents, administrators, and executives needed numbers to do their jobs, and every report was an engineering ticket waiting on a sprint. So most questions never got asked. I proposed and built a system where you type the question in English and get the dashboard back.

Letting a model write its own queries would have demoed well and been quietly wrong. So it doesn’t. Identity and tenant isolation also belong below the agent, not inside it.

Where the model is allowed to act. Select a step.

2. Agent

The model only picks which question to ask. It does not write queries, compute numbers, or see field names.

What the cost analysis showed

  • 78%of the bill was input tokens, because the catalog is re-sent on every round trip.
  • ~50%of spend is what caching that catalog is worth. Changing model version moved it by 2%.
  • 61%less cost on the low reasoning setting, which matched the middle one on accuracy.
60%lower AI costs

95 Percent Group, via GlobalLogic

Enterprise AI adoption, with governance built in

I own the AI governance model: role-based access, usage policies, and acceptable-use standards. PII-handling guardrails sit inside the AI gateway, alongside access controls and rate limits that protect sensitive data and internal systems.

Every developer is equipped with custom MCP tools, modular LLM skills, and deployment pipelines for AI agents. That foundation made it possible to refactor 23 core microservices in 45 days.

3 daysbetween production releases

One95 Literacy Ecosystem, 200K+ daily active users

From a release every 5 weeks to every 3 days

I treated release cadence as a platform problem. Tooling and delivery pipelines removed engineering bottlenecks and bureaucratic hand-offs, taking the cycle to 2 weeks at the midpoint and then to 3 days, tracked in an executive DORA dashboard.

The core architecture, including the authentication and authorization layer, lets new engineers ship business features from day one without touching identity, permissions, or other cross-cutting concerns.

~20districts in a couple of minutes, up from hours for one

One95 Literacy Ecosystem. Design owner and lead engineer

Roster ingestion rewrite

Every district onboarded through this service, and it capped how fast we could grow. I chose not to repair it and re-scoped the problem: keep authentication and the shared layers, and rewrite only what was slow. Fetching split from transformation, and transformation moved to a Go batch service.

The harder part was not technical. The VP and CTO had to accept a language most of the team did not know, and security had to agree the existing auth layer covered the new path. So I built a working MVP and demoed it before asking for the decision.

Hindsight: QA passed for several sprints, then production found third-party timeouts and incomplete records we could not see. I added dashboards, step logging, alerts, and compensating actions. Those should have shipped with the rewrite.

2+stolen cars recovered soon after launch

Kavak, 2022

In-vehicle panic button

Someone being robbed while driving has seconds and one free hand, and every option available assumed they could explain where they were. The car already knew. I spotted that and built a GPS integration: one press sends an alert with the live position straight to the C4 emergency command centre.

It raised customer satisfaction by 10%, became a differentiator in the sales pitch, and earned my promotion from Senior Software Engineer to Lead Engineer.

Today: I would use a model to widen the list of GPS failure modes and accidental-press cases before writing code. I did that by hand at the time.

Leadership

What I bring to your engineering organization

AI platform architecture

Agents and agent harnesses, RAG, evals, embeddings, and MCP tooling, designed around one question: where is the model allowed to act, and what must stay deterministic?

AI governance, security, and cost

Usage policies, role-based access, LLM guardrails, and PII protection, with consumption monitoring that makes spend a design input rather than a surprise.

Platform engineering and delivery

Developers treated as internal customers. High-concurrency distributed systems, with availability and reliability as product requirements, and owned once they are live.

Engineering leadership

A ~20-engineer organization across multiple pods with full ownership: 1:1s, roadmap, and code-review standards. Comfortable in client discovery and in front of a CTO.

Experience

10+ years, from startups to enterprise platforms

Full resume (PDF)

GlobalLogic / 95 Percent Group

Dec 2022 – Present

Lead Software Architect, AI Platform Engineer

System architect and technical lead for the One95 Literacy Ecosystem, owning platform engineering, AI strategy, and delivery end to end.

  • Lead the ~20-engineer cross-team platform organization in a distributed U.S./Mexico team.
  • Own the AI strategy, governance model, and tooling used by every developer.
  • Cut production release cycles from every 5 weeks to every 3 days.

Kavak

Jan 2022 – Dec 2022

Lead Engineer, promoted from Senior Software Engineer

Mexico’s first unicorn (US$8.7B valuation in 2021), during post-funding hypergrowth.

  • Led a 12-person squad, and took the team on-site to turn operational insight into product improvements.
  • Drove the monolith-to-microservices transition with cloud-native Node.js and AWS Lambda services.
  • Shipped the panic button and theft-reporting feature, still live today.

Clip

Aug 2021 – Jan 2022

Infrastructure and Platform Engineer

Mexican fintech unicorn processing payments for hundreds of thousands of merchants.

  • Built the internal RabbitMQ library and core Node.js libraries, services, and SDKs used across product squads.
  • Owned and operated AWS infrastructure (EC2, ECS, Kubernetes, CloudFormation) and its CI/CD pipelines.
Earlier experience, 2017 – 2021
  • ClearMechanicSoftware Engineer · 2021

    Led third-party integrations for internationalized automotive software.

  • SofttekSoftware Engineer · 2020 – 2021

    Internal management tools and security policies for a banking client.

  • Secretaria de FinanzasSoftware Engineer · 2019 – 2020

    Digitized municipal reporting and scheduling on self-managed servers.

  • DritdiSoftware Engineer · 2017 – 2019

    Co-built a medical-education startup, a 3D training app licensed to schools.

Stack and credentials

AI
Agents and harnesses, RAG, evals, fine-tuning, embeddings, vector databases, MCP, prompt engineering
Governance
Usage policies, responsible AI, guardrails, data protection, rate limiting, token and cloud cost optimization
Languages
Node.js, TypeScript, Go, Python, SQL
Cloud
AWS, Azure, Kubernetes, Helm, Terraform, microservices, Kafka, PostgreSQL, MongoDB, Redis
Operations
Metrics, logging, distributed tracing, SLOs, incident response, Argo CD, canary and blue-green deployments
Languages
Spanish (native), English (advanced, C1)
Education
B.S. in Information Technology Engineering, Universidad Tecnológica de Tabasco

Consulting

Not hiring, but stuck on an AI platform problem?

My focus is a leadership role, and I also take on select, scoped engagements with teams moving AI from demo to production.

Ask about an engagement

  • Agent architecture design

    Decide where the model acts, what stays deterministic, and how answers get verified before anyone relies on them.

  • AI governance and security review

    Access model, usage policy, PII guardrails, and gateway design, reviewed against how your teams actually use AI.

  • AI cost reduction

    Find where the tokens go, then apply model right-sizing, caching, and limits without giving up accuracy.

Contact

Tell me what’s hard about the role.

Send the role or the problem, and what’s getting in the way. I work remotely from Mexico City on U.S. time zones.

[email protected]