Resume — Jesús Eduardo Rodríguez Saucedo

Profile

AI Engineer with 10+ years building production-grade software and real-time AI systems, focused on agentic AI, multi-agent orchestration, and LLM-powered applications. I design systems where multiple specialized agents collaborate under a supervisory architecture — with structured outputs, tool integration, and auditable execution — and bring the same rigor to evaluating them: controlled baselines, repeated trials, ablations, and blind labeling to separate genuine model behavior from execution and prompt-scaffolding artifacts. Comfortable owning a system end-to-end, from asynchronous backend architecture to the evaluation harness that proves it works.

Core expertise

Agentic AI & Multi-Agent Systems

Multi-agent orchestrationsupervisor architecturesspecialized agentstool/function callingstructured outputsagent workflowsmulti-provider LLM systems

AI Engineering & Architecture

PythonFastAPIasyncioREST APIsWebSocketsdistributed & event-driven systemsreal-time AIAPI integrations

LLM Evaluation & Safety

Prompt engineeringreproducible evaluation designred teamingbehavioral robustnesshallucination detectionrisk classification

LLMOps & Observability

LangfuseDatadogstructured loggingexecution tracingevaluation pipelinesexperiment provenance

Cloud & Infrastructure

Google Cloud (Cloud Run, Cloud Build)Firebase/FirestoreDockerPostgreSQLRedisSupabase

Delivery

Git/GitHubGitHub ActionsCI/CDTypeScriptReactNext.js

Professional experience

Lead AI Engineer — VocalisAI

2025 – Present

Independent / Google Cloud

Finalist — Google Gemini Live Agent Challenge 2026

Architected a production-oriented multi-agent voice intelligence platform: one supervisor agent orchestrating six specialized agents (reception, qualification, emergency triage, billing, follow-up, outbound) on Gemini Live, FastAPI, WebSockets, Twilio, and Google Cloud.
Built real-time evaluation for hallucinations, boundary violations, and false-urgency behavior, with corrective intervention and hard-veto mechanisms for high-risk model actions.
Designed low-latency asynchronous execution pipelines for real-time voice interactions, with structured validation and observability built into the agentic pipeline.

AI Systems Architect — Multi-LLM Evaluation

2024 – Present

Independent

Finalist — Google Cloud AI + Datadog Hackathon

Designed multi-provider LLM orchestration and evaluation architectures coordinating several providers through structured, comparative workflows.
Implemented structured outputs, stage-level observability, execution checkpoints, and auditable experiment artifacts.
Built evaluation pipelines supporting repeated trials, baseline comparison, component ablations, and behavioral analysis, with infrastructure to separate model behavior from execution and parser artifacts.

AI Systems Engineer — LeadHunter

2025 – Present

Independent

Built an AI-powered lead qualification pipeline: Normalize → Deduplicate → LLM Scoring → Drafting → CRM, integrating PostgreSQL/Supabase, HubSpot APIs, and Langfuse observability.
Implemented structured validation and deterministic processing stages around probabilistic LLM components, turning unstructured lead data into validated CRM-ready records.

Lead Developer & Solutions Architect

2020 – 2024

World Vision Telecom

Engineered production-grade real-time communication systems supporting 500K+ interactions, using Python, FastAPI, Redis, and WebSockets on cloud infrastructure.
Built asynchronous processing pipelines for high-volume communication workloads, with emphasis on reliability, scalability, and operational resilience.
Led architecture and implementation decisions across backend services and real-time application infrastructure, integrating external APIs and communication platforms into production workflows.

Software Engineer

2014 – 2020

Earlier Experience

Software engineering, systems architecture, web platforms, automation, distributed applications, API integrations, and production infrastructure.

Applied LLM evaluation & research

Independent, reproducibility-focused evaluation work applying the same evidentiary standards used above to LLM behavior itself — the discipline behind the safety mechanisms shipped in VocalisAI.

Deployed LLM Eval Kit

2026

Open-source methodology

Built an open, reproducible methodology and toolkit for evaluating LLM safety behavior as deployed in product chat interfaces (not raw APIs): frozen rubrics, real blind labeling, and platform metadata captured as first-class results.
Implemented inter-rater concordance tooling (Cohen's kappa) and a stimulus-freezing workflow (SHA-256 hashing) to keep evaluation runs auditable and comparable over time.

github.com/zoharmx/deployed-llm-eval-kit

Behavioral Safety & Robustness Evaluation

2026

Independent Research

Designed a controlled cross-model evaluation of LLM safety behavior using frozen stimuli, repeated trials, captured outputs/metadata, and a pre-defined labeling rubric.
Built reproducible evaluation artifacts and explicit controls to distinguish model behavior from execution and prompt-scaffolding effects; established coordinated-disclosure procedures for security-relevant findings.

Deliberative LLM Evaluation

2026

Independent Research

Designed a falsifiable 25-case pilot comparing a multi-stage deliberative LLM architecture against direct model responses across five dilemma classes.
Implemented baseline comparison, operational risk-coverage metrics, component ablations, and signed, resumable checkpoints; audited the pipeline, identified parser-related execution artifacts, and recomputed results with documented provenance.

BLINDAJE — AI Agent Security Auditing

2026

Independent practice

Independent security-audit practice for deployed conversational agents: a 35-vector, 12-failure-family catalog covering prompt injection (direct and indirect), tool abuse, cross-user data leakage, and business-logic exploitation.
Every finding is evidence-linked to the exact conversation turn that produced it, with pass/fail criteria fixed before testing and a mandatory human-signed verdict; sample report available on request.

eduardorodriguez.site/blindaje

Operación BLINDAJE

This methodology, available as a service

BLINDAJE is the commercial arm of this research: a security audit for conversational AI agents already in production, run against a catalog of 35 attack vectors across 12 failure families. Pass/fail criteria are fixed before testing, every finding is linked to the exact conversation turn that produced it, and no report ships without a human-signed verdict.

See the service (Spanish)Download a sample report (PDF)

Technical stack

Languages: Python·TypeScript·JavaScript·SQL
AI / LLM: OpenAI·Gemini·multi-LLM orchestration·AI agents·tool/function calling·structured outputs·LLM evaluation
Frameworks & Backend: FastAPI·asyncio·WebSockets·REST APIs
Observability / LLMOps: Langfuse·Datadog·structured logging·evaluation pipelines
Cloud / Infrastructure: Google Cloud (Cloud Run, Cloud Build)·Firebase/Firestore·Docker
Data: PostgreSQL·Supabase·Redis
Frontend: React·TypeScript·Next.js
DevOps: Git·GitHub·GitHub Actions·CI/CD

Recognition

Google Cloud AI + Datadog Hackathon — Finalist

Google Gemini Live Agent Challenge 2026 — Finalist

Looking for someone to own an agentic system end-to-end?

From asynchronous backend architecture to the evaluation harness that proves it works.

Get in touch See the work