Joan Pellegrino
$ status --liveAvailable for new projects

Joan Pellegrino

AI Engineer &Forward Deployed Engineer

Turning complex integrations into scalable AI products.

Multi-Agent Systems (Google ADK, LangGraph) · FinTech Integrations & Data Pipelines

Explore ProjectsDownload CV
GitHubLinkedInEmail

Ask a Question

Ask about my projects, stack or experience

Tap to talk

Try saying: “What experience do you have with LLM systems?”

This site doesn't save your audio. Your voice is processed in real time by Google Gemini. Basic session metrics are recorded.

AGENTS / LLMGoogle ADK · LangGraph · Gemini
CLOUDAWS · GCP
CODEPython

$ ls ./projects

Featured Projects & Architecture

AI and data systems, each with its architecture.

System Architecture

flowchart TD
User["User Events (Slack / Zendesk / CLI)"] --> Router

subgraph EKS["Agent Services on AWS EKS"]
    Router["AgentRouter (gemini-2.0-flash)"]
    Router -->|"DATA_QUERY"| SQL["JobsSQLAgent"]
    Router -->|"DOCUMENTATION"| Docs["DocsRAGAgent"]
    Router -->|"DIAGNOSTIC"| Diag["SupportDiagAgent"]
    Router -->|"REEXECUTION"| Reexec["ReexecutionAgent"]
end

SQL --> DB[("PostgreSQL via MCP")]
Docs --> KB["Markdown Knowledge Base"]
Diag --> DB
Diag --> CW[("AWS CloudWatch Logs")]
EKS -.-> DDB[("DynamoDB (State & Memory)")]

Reexec --> API[("Backend API")]
API --> Batch[("AWS Batch")]

classDef default fill:#0f172a,stroke:#a855f7,stroke-width:1px,color:#e2e8f0
classDef router fill:#881337,stroke:#f43f5e,stroke-width:1px,color:#fecdd3
classDef specialist fill:#1e1b4b,stroke:#818cf8,stroke-width:1px,color:#e0e7ff
classDef storage fill:#022c22,stroke:#10b981,stroke-width:1px,color:#d1fae5

class Router router
class SQL,Docs,Diag,Reexec specialist
class DB,CW,API,Batch,KB,DDB storage

Multi-Agent Router & Specialists Pattern

The system relies on a scalable Router + Specialists design deployed on AWS EKS. A central Custom Agent (Router) leverages LLMs to classify the intent of an incoming event (e.g., from a Support Ticketing System or Chat interface) and securely delegates execution to specialized downstream agents. All specialists share the same isolated session context and memory backend.

Data Query Agent

Resolves operational queries on relational databases securely via an isolated MCP Toolbox using curated SQL toolsets.

Documentation RAG Agent

Executes full-context document analysis over Markdown knowledge bases, ensuring exact requirement extraction without the fuzziness of vector search.

Diagnostic Agent

Analyzes errors and actively troubleshoots pipeline failures by querying Cloud Execution Logs and relational metadata.

Re-execution Agent

An orchestrator agent that digests ticketing events, validates requirements, triggers jobs via backend API (running on AWS Batch), and safely updates the ticket status upon completion.

State & Persistence Layer

Chat sessions and state transitions are robustly maintained using DynamoDB with automated TTL policies, persisting session memory and execution traces reliably.

REAL-TIME VOICE
Live demo

Voice AI Agent for Portfolio

Real-time conversational voice agent embedded in this site. Visitors can talk to an AI powered by Google Gemini Live API over WebSockets with ephemeral token security that answers questions about my projects, stack and experience.

Gemini Live APIWeb Audio APIWebSocketsAstro
RESULT

Live on this site: talk to it from the hero orb · 260+ real voice sessions

Try it liveView architecture
AGENTIC TOOLING · CLI
Private

Smart-Parser Claude Plugin

Declarative agentic engine for Claude Code CLI with specialized Python skills that accelerates customer data onboarding by synthesizing compliant banking parsers and ETL connectors, running security/schema audits, and automating Git Pull Requests.

Claude APIPythonAgentic SkillsAST ParserGit Automation
RESULT

Drastically reduces client onboarding time-to-value & automates Git PR cycles

System Architecture

flowchart LR
A["Claude Code CLI"] -->|"Loads via Settings"| B("Local Plugin Marketplace")
B --> C{"Agentic Skills Layer"}
C -->|"new-script"| D["Code Generation Engine"]
C -->|"validate"| E["Automated Reviewer"]
C -->|"git-deliver"| F["Workflow Automation"]
D --> G[("Internal Enterprise Repo")]
E --> G
F --> H["Enterprise Git Provider"]

classDef default fill:#0f172a,stroke:#a855f7,stroke-width:1px,color:#e2e8f0
classDef core fill:#881337,stroke:#f43f5e,stroke-width:1px,color:#fecdd3
classDef base fill:#022c22,stroke:#10b981,stroke-width:1px,color:#d1fae5

class C core
class G,H base

LLM Developer Tools Architecture

This project extends the LLM CLI capabilities via a custom local marketplace plugin. It embeds domain-specific intelligence for an Enterprise ETL Engine, enabling the agent to autonomously navigate complex internal contracts, generate specialized data parsers, and execute standard code delivery workflows.

Code Generation

Synthesizes atomic data transformations, API managers, and complex parsers adhering strictly to internal codebase standards and templates.

Automated Review

Proactively inspects existing ETL scripts, flagging critical security or structural deviations with categorized severity levels.

Workflow Automation

Manages the end-to-end Git lifecycle, discovering local repositories, creating conventional feature branches, and submitting templated Pull Requests.

Skill Implementation

The plugin is built using a highly structured declarative format with metadata routing. Each capability is isolated in its own skill module backed by prompt engineering techniques that inject dynamic reference documentation into the LLM context.

CLOUD & DATA
Private

Distributed API & S3 Ingestion Engine

High-throughput ETL and ingestion architecture connecting heterogeneous financial sources (payment REST APIs, SFTP, AWS S3, GCS) to Snowflake, BigQuery, and PostgreSQL with fault-tolerant serverless pipelines and SQS DLQ.

PythonPandasREST APIsAWS S3AWS LambdaPostgreSQLSnowflake
RESULT

Millions of daily financial transactions with a 99.9% availability SLA

System Architecture

flowchart TD
    subgraph Sources["Heterogeneous Data Sources"]
        API["RESTful APIs / Webhooks"]
        SFTP["SFTP Server"]
        ExtS3["External S3 / GCS"]
    end

    subgraph Ingestion["Ingestion & Buffer"]
        LambdaIngest["Ingestion Lambda Worker"]
        RawS3[("AWS S3 Raw Lake (Bronze)")]
    end

    subgraph Processing["Serverless Transformation"]
        ETLLambda["Transform Lambda (Pandas Engine)"]
        DLQ[("Amazon SQS (DLQ)")]
    end

    subgraph Destinations["Dual Analytical Storage"]
        Snowflake[("Snowflake Data Warehouse")]
        Postgres[("PostgreSQL Operational DB")]
    end

    API --> LambdaIngest
    SFTP --> LambdaIngest
    ExtS3 --> LambdaIngest

    LambdaIngest --> RawS3
    RawS3 --> ETLLambda

    ETLLambda -->|"Schema Deviation / Error"| DLQ
    ETLLambda -->|"Parquet Curated Batches"| Snowflake
    ETLLambda -->|"Relational Operational Data"| Postgres

    classDef default fill:#0f172a,stroke:#a855f7,stroke-width:1px,color:#e2e8f0
    classDef source fill:#1e1b4b,stroke:#818cf8,stroke-width:1px,color:#e0e7ff
    classDef worker fill:#881337,stroke:#f43f5e,stroke-width:1px,color:#fecdd3
    classDef storage fill:#022c22,stroke:#10b981,stroke-width:1px,color:#d1fae5
    classDef alert fill:#450a0a,stroke:#ef4444,stroke-width:1px,color:#fee2e2

    class API,SFTP,ExtS3 source
    class LambdaIngest,ETLLambda worker
    class RawS3,Snowflake,Postgres storage
    class DLQ alert

Distributed Ingestion & ETL Engine Architecture

High-throughput, event-driven data pipeline designed to ingest, validate, and normalize records from disparate enterprise channels. Operates with an isolated buffer layer in AWS S3 and an automated schema enforcement engine to stream structured datasets directly into analytical warehouses and transactional stores.

Heterogeneous Ingestion

Orchestrates automated connectors across RESTful APIs, SFTP endpoints, and third-party cloud buckets, validating payload headers and raw checksums.

Serverless Transformation

Vectorized Pandas data transformations running in AWS Lambda, executing schema type coercions, deduplication, and snappy Parquet compression.

Dual Storage Engine

Concurrent multi-sink delivery feeding Snowflake for petabyte-scale analytics and PostgreSQL for low-latency operational reporting.

Fault Tolerance & Dead-Letter Queue (DLQ)

Any structural deviation or transient connection timeout triggers automated exponential backoff. Irrecoverable payloads are routed to an isolated Amazon SQS Dead-Letter Queue accompanied by CloudWatch alerts, guaranteeing zero data loss and straightforward replayability.

Technologies & Professional Journey

Stack & Experience

01.AI & GenAI Engineering

11 tools
Google ADKLangGraphLangChainMCP (Model Context Protocol)Multi-Agent OrchestrationRAGText-to-SQLGemini APIAnthropic Claude APIOpenAI APIVector Databases

02.Cloud & Infrastructure

10 tools
AWS EKSAWS LambdaAWS BatchAmazon CloudWatchDynamoDBAmazon S3Amazon SQS (DLQ)API GatewayGCP (BigQuery, GCS)Serverless & Event-Driven

03.Data & Languages

10 tools
PythonPySparkApache KafkaSQLPostgreSQLSnowflakeBigQueryETL/ELT PipelinesRESTful APIsFinancial Formats (MT940, Parquet)

04.Tools & DevOps

7 tools
GitDockerCI/CDAWS GlueSFTPAstroTailwind CSS

Senior AI & Integrations Engineer

Simetrik

commit 2020..Sept 2026

Operated at the intersection of AI Engineering, Forward Deployed Integrations, and Product Delivery — leading Multi-Agent Systems, RAG, MCP Text-to-SQL, and high-scale FinTech data pipelines:

  • Designed and deployed a stateful multi-agent system on AWS EKS using Google ADK and LangGraph (Router + Specialists pattern) to autonomously resolve operational queries, diagnose CloudWatch logs, and orchestrate re-executions in AWS Batch with persistent memory in DynamoDB.
  • Built a conversational Text-to-SQL agent connected to PostgreSQL via MCP (Model Context Protocol), implementing strict schema validation, read-only security controls, and contextual explanations for business teams.
  • Developed a semantic extraction and RAG pipeline over technical documentation and unstructured contracts, enabling non-technical teams to query knowledge bases in natural language and reducing operational reliance on specialists.
  • Created the Smart-Parser Claude Plugin with Python agentic skills to automatically generate, validate, and audit data parsers and complex ETL scripts, accelerating the delivery and integration cycle in Git.
  • Architected and implemented enterprise ETL/ELT pipelines across REST APIs, SFTP, AWS S3, GCS, BigQuery, PostgreSQL, and Snowflake, processing millions of daily financial transactions with 99.9% uptime (SLA).
  • Implemented fault-tolerant, asynchronous synchronization using Apache Kafka, webhooks, AWS Lambda, API Gateway, and Amazon SQS (DLQ), alongside Python and PySpark validation engines to reconcile financial files (MT940, structured PDFs, CSV, and Parquet).
PythonGoogle ADKLangGraphAWS EKSMCPRAGText-to-SQLApache KafkaPySparkSnowflakePostgreSQLDynamoDBAWS Batch
✦ About

I engineer intelligent systems.

AI Engineer & Forward Deployed Engineer with nearly 6 years of experience turning complex enterprise and FinTech integrations into scalable AI products and high-volume data pipelines.

I operate at the intersection of deep engineering and product delivery — bridging stateful multi-agent orchestration with deterministic financial cores so LLMs deliver measurable business value without compromising mission-critical accuracy:

  • I architect and deploy stateful Multi-Agent systems (Router + Specialists) using Google ADK and LangGraph on AWS EKS with DynamoDB memory — enabling self-service operations across Slack and Zendesk.
  • I build secure Text-to-SQL conversational agents connected to PostgreSQL via MCP (Model Context Protocol), enforcing strict Pydantic schema validation and read-only guardrails so product and finance teams can self-serve insights.
  • I accelerate client onboarding (time-to-value) through semantic RAG pipelines, custom Claude Code plugins (Smart-Parser), and real-time voice products powered by Gemini Live API over WebSockets.
  • I bring nearly 6 years of FinTech integration & data engineering at Simetrik, architecting event-driven ETL/ELT pipelines across payment APIs, Python, PySpark, Snowflake, BigQuery, and Kafka — processing millions of daily transactions at a 99.9% SLA.
View Resume

// engineering evolution

Enterprise Systems
FinTech Integrations & APIs
Cloud & Data Pipelines
Multi-Agent Systems (ADK, LangGraph)
AI & Forward Deployed Engineering

Previously at

Simetrik

Location

Colombia

Education

Computer Science — University of Córdoba

commit 2015..2019

Hackathon

FRONTIER AI Builders Hackathon (Oct 2026) - 'The Black Box' challenge, team entry: replaced a legacy system, passing 99 of 100 hidden tests; ranked 8th of 38 teams.

"The most reliable AI systems are built on solid engineering foundations — connected to real data, designed for production, and grounded in the systems they automate."
Joan Pellegrino · AI Engineer & Forward Deployed Engineer

Let's connect

Get in Touch

$init --collaborate

Let's build something intelligent

Open to opportunities in AI Engineering, Forward Deployed Engineering, and Multi-Agent Systems.

Get in Touch
or find me at

joanpellegrino04@gmail.com