García-Moreno Zinchenko

AI / Agent Engineer · Guadalajara, MX · EST + CET overlap

The tools will change. The work still has to make sense.

Multi-agent workflows, grounded RAG, evaluation harnesses, model routing, and observability for LLM systems that need to work beyond demos.

See the evidence ↓ Try the working model

Open to remote AI / Agent Engineer roles across US and European time zones.

Plate 01 Capabilities · 1:1

The systems I can help a team ship.

Engineering strengths framed around the work an AI product team needs: orchestration, grounding, evaluation, and operation.

Orchestration

Agentic workflows

I build multi-agent pipelines and self-hosted runtimes with tool use, MCP/A2A interfaces, channel adapters, and operational workflows.

  • Multi-agent orchestration
  • MCP and A2A
  • Operational workflows

Grounding

Grounded AI systems

I design RAG systems that retrieve source-backed context, verify citations, and expose evidence to both users and agents.

  • Hybrid retrieval
  • Verified citations
  • Source-backed context

Evaluation

Evaluation and reliability

I use evaluation harnesses, red-teaming, and test-first workflows to make AI-assisted work reviewable against explicit criteria.

  • Evaluation harnesses
  • Red-teaming
  • Test-first delivery

Operation

Observability and operations

I instrument traces, token cost, latency, and errors, and design model-routing and provider-fallback paths for graceful degradation.

  • LLM observability
  • Model routing
  • Provider fallbacks

Plate 02 Selected work · Public repositories

Three repositories. Three operating problems.

These public repositories show approaches to grounded retrieval, reviewable agentic engineering, and self-hosted agent operations. Open a plate, then run its system.

Problem
Codebase answers need source references that can be checked instead of plausible but ungrounded citations.
System
A verified context engine that indexes repositories, retrieves hybrid code context, and verifies citations against indexed snapshots.
Evidence
  • Python package and CLI
  • REST and MCP surfaces
  • Verified citations and context packs
Repository
Open the repository ↗

Plate 02 · A · System drawing

00Press run, or select a block, to follow a question through the system.

Drawing of the public design. Block names describe roles, not module names.

Principles07

How I make AI work reviewable.

The standard is not a persuasive demo. It is a system whose sources, behavior, and failure paths a team can inspect.

Seven stations on a scale model, one illustrative on-call incident. The run is modeled on Triage, a multi-agent SRE workflow that retrieves repository evidence, produces a human-approved TriageCard, creates a Linear ticket, and verifies fix evidence before closure, with Linear, Slack, email and Langfuse integrations. The alert, files and timings below are an illustrative example, not real metrics.

  1. 00 Intent

    Write it the way you'd ask someone.

    No commands, no forms. Each letter drops onto the request desk. Press Delegate and the agent separates what you want, with whom, by when, and what it must not do without you.

    In the model Type a request or pick an example, then press Delegate. Tap a letter to lift it.

    Example request · Meeting

    Prepare the sales closing meeting and send the agenda on Thursday

    Tasks
    1. Prepare the sales closing meeting
    2. Send the agenda on Thursday
    With whom
    Sales
    When
    Thursday
    Not without you
    Send, pay or delete anything

  2. 01 Evidence

    Ground every answer in inspectable source.

    When the alert fires, an agent is dropped into the model, walks to the repository building, and brings back each finding as a labeled citation block you can open.

    In the model Tap the mat or press Dispatch agent. Tap a citation block to open it. Drop three more and they build a house.

    Illustrative example

    Alert checkout-api · p95 latency above 2 s for 10 min

    1. [1] services/checkout/db/pool.ts · L12–19 pool size changed from 20 to 5
    2. [2] deploy #418 diff · merged 14:02 touches pool.ts
    3. [3] runbooks/checkout-latency.md · § Pool exhaustion matching symptom

  3. 02 Boundaries

    Keep model and tool boundaries explicit.

    Every tool is a separate piece with a written scope: put one on the bench and its sheet says what the agent may and may not do with it, while the deploy key stays outside the tape.

    In the model Drag a tool onto the lit pad, or tap it. Try the deploy key.

    Illustrative scopes

    Repository
    May read files and diffs at a pinned commit. May not write, push or open pull requests.
    Linear
    May create one ticket from an approved TriageCard. May not close it without verified fix evidence.
    Slack
    May post to the incident channel. May not message people directly.
    Email
    May send the summary to the on-call list. May not write to other addresses.
    Deploy
    Not granted. Rollbacks stay with people.

  4. 03 Review

    Use human review where ambiguity is expensive.

    The drafted TriageCard waits at a barrier, and nothing is filed or sent until a person holds the button to approve it or sends it back with a change.

    In the model Press and hold the orange button, or hold Approve. Space and Enter work too.

    Illustrative TriageCard

    Severity
    SEV2
    Cause
    Connection pool cut from 20 to 5 in deploy #418 [1][2]
    Action
    Create a Linear ticket for the checkout owners, post to #inc-checkout
    Gate
    Waits for a person. Nothing has been filed or sent.

  5. 04 Evaluation

    Replace confidence with evaluation evidence.

    Before the ticket may close, every claim rides a conveyor through a named check that either passes with evidence or fails in plain sight.

    In the model Tap a box on the belt to open its check. Send failures down their path.

    Illustrative closure checks

    • Citations: every claim cites a file, diff or runbook
    • Cause: matches the deploy #418 diff
    • Fix: PR linked on the ticket and touches the cited file
    • Latency: p95 under 2 s for 30 min after the fix only 12 min of data
    • Test: a regression test covers the pool setting none found

  6. 05 Failure paths

    Design failure paths before the happy path ships.

    The routes for missing evidence, tool timeouts and failed checks are taped to the floor before the run starts, so each failure goes somewhere known.

    In the model Tap a taped route or pick a failure. The agent follows the route drawn in advance.

    Illustrative routes

    Missing evidence
    No TriageCard from guesses. Ask #inc-checkout on Slack for context.
    Tool timeout
    Retry twice with backoff, then hand the partial trace to on-call.
    Failed check
    The ticket stays open and its owner gets the failing evidence.

  7. 06 Trace

    Treat agent operation as engineering, not prompting.

    Every step of the run leaves a span in a trace, the kind Triage sends to Langfuse, so behavior is replayed and debugged from evidence instead of tuned by prompt.

    In the model Drag sideways to rotate the whole model, or use the rotate buttons.

    Illustrative trace

    1. alert.receivedcheckout-api p95 above 2 s
    2. retrieve.repo3 citations, read-only
    3. review.humanTriageCard approved
    4. linear.create_issueafter approval
    5. eval.fix_evidence3 of 5 pass, 2 routed
    6. failure.routeticket held open

The scale model is an interactive WebGL scene; without JavaScript or WebGL this section shows the same seven stations as text. All sample data is an illustrative example, not production metrics.

Plate 04 Experience · Drawn to scale

Professional trajectory.

Recent roles span agent engineering, software systems, conversational AI, and automation. One millimetre is one week; move along the ruler to read any month.

Move along the ruler

– Present

Agentic Engineering Agency

AI / Agent Engineer

I operate the technical stack for an early-stage AI engineering studio, leading client builds and internal tooling across agent pipelines, RAG, evaluation, harness engineering, and observability.

– Present

Voblakye

Software & AI Engineer

Ongoing software and AI engineering work across internal initiatives. I also developed the curricula and instructional materials for the AI courses I currently deliver to a corporate client’s employees.

–

TendencIA

AI & Automation Developer

Built WhatsApp and Telegram bots, customer-support chatbots, and internal automations using third-party LLM APIs. For AIRE, validated and integrated AI-backed flows and owned most of the Flutter and Supabase frontend.

Plate 05 Contact

Let’s build AI systems that hold up beyond the demo.

Based in Guadalajara, Mexico, with EST and CET overlap. If you are hiring for agent orchestration, grounded RAG, evaluation, reliability, or AI engineering infrastructure, I would like to hear about the role.

Also open to selected agentic-systems consulting and build collaborations.

Brief · cut it out to send

Leave me a brief.

Drag the knife along the dashed line

LinkedInSebastián García-MorenoOpen profile GitHubSebastiangmzOpen repositories CurriculumCV · PDFOpen the CV