Experience and selected projects

Professional experience

1. 

   Sep 2026 – Present

   United States · Remote

   In production

   AI EngineerRampy ↗ (opens in a new tab)

   I lead engineering at Rampy and work closely with the three founders and with specialists in blockchain and integrations. Rampy is an early-stage fintech and blockchain startup whose non-custodial app opens the world of digital assets (swaps, yield vaults, perps and market news) to everyday people through a generative interface that turns complex concepts into simple ones. I own the multilingual AI assistant and the eval suite that gates every release (3,000+ cases, 90%+ pass rate), plus the backend, web and mobile app.

   • Built a multilingual AI assistant (English, Spanish and Portuguese) and the eval suite behind it: 1,000+ scenarios replicated and translated into each language (3,000+ cases), graded in LangSmith by an LLM-as-a-judge validated against human and product review, and used as a release gate. Production agents pass 90%+ on average. A case passes when the answer is correct and grounded, the right tool is used, the structured contract for the UI is valid, guardrails hold and the reply solves the user’s need in plain language.
   • Designed the multi-agent LangGraph graph, with cycles, financial tools, background workers, retries and failure recovery, and migrated the agentic stack from Agno as the flows grew. Claude is the primary model, with DeepInfra and Vertex AI as fallback providers, GraphRAG retrieval over Neo4j and PostgreSQL/pgvector, and Mem0 memory. Removing redundant tool calls and tightening memory and context made responses 2× faster with 40% fewer tokens, verified with load tests and evals.

   • 90%+

     pass rate on a 3,000+ case eval suite

     Global average for production agents on 1,000+ scenarios run in English, Spanish and Portuguese, graded in LangSmith by an LLM-as-a-judge validated against human and product review.

   • 2×

     faster responses in optimized agent workflows

     At least 2×, measured with load tests and evals on comparable optimized agent workflows.

   • 40%

     fewer tokens in optimized agent workflows

     Measured with load tests and evals after removing redundant tool calls and tightening memory and context; not a benchmark of every query.

   • 2×

     faster optimized financial queries

     At least 2×, measured with load tests on comparable optimized vault, market and swap queries in production.

   • ~1 h → 10 min

     delivery pipeline

     Approximate observed before/after for validation, build and deployment, with checks retained.

   Leading engineering since September 2026 with the founders, specialists in blockchain and integrations, and an experienced advisor on specific architecture decisions.

   More contributions (13)

   • Designed the structured-output contract shared by the agents and the app, so answers render as generative UI: native buttons, cards and deep links into swap, vault and perps flows, with live market data. The agent prepares each transaction and the user confirms it in a non-custodial Privy wallet.
   • Shipped trading across 200+ Hyperliquid perpetual markets (crypto, indices, stocks and commodities), swaps across 100+ tokens via LI.FI, and Morpho, Aave and Compound vaults for positions, yields, deposits and withdrawals, with completeness checks and webhook-based status tracking. Designed and implemented the company’s fee model for swaps, vaults and perps, and worked with Privy authentication, Stripe integrations, Relay for cross-chain connections and Alchemy RPC for on-chain queries in production.
   • Evolved the inherited codebase step by step: the backend from no defined architecture to hexagonal (ports and adapters) and then Clean Architecture; the React/Next.js frontend from no defined architecture to a feature-based structure and then Clean Architecture, with a complete design system with reusable component abstractions.
   • Built a consolidated portfolio experience for assets, vaults and 24-hour returns; reduced comparable vault, market and perpetual data retrieval from seconds into the millisecond range through prefetching, parallel provider calls and selective caching.
   • Cut the validation, build and deployment pipeline from about one hour to ten minutes while retaining its checks; optimized application bundles, Docker images and development, staging and production environments on DigitalOcean.
   • Implemented end-to-end encryption (E2EE) for user conversations in the production app, with messages persisted encrypted in the database.
   • Shipped the React Native app, using Android Studio for Android development and testing in emulators and on physical devices. The Android app is live and the iOS version is ready, with its launch pending.
   • Implemented complete mobile event tracking with Google Analytics through Firebase, Clarity Mobile on every screen, and Singular SDK for Meta App Events, coordinating integrations to prevent duplicate events. Designed, built, tested and deployed a campaign landing page for app downloads.
   • Built the enterprise back office to manage application operations and centralize monitoring, telemetry and operational status; I continue to expand its capabilities.
   • Added domain and malicious-instruction controls to agent workflows, checked with automated behavioral tests and human review.
   • Centralized secrets with Infisical, per-user access, automated synchronization and on-demand rotation; worked across deployment, monitoring, backups, databases and server permissions.
   • Standardized spec-driven engineering through an agent skills marketplace and repeatable local setup, connecting requirements, architecture, implementation, automated tests and human UI review.
   • Prepared scalable infrastructure with the design goal of supporting thousands of users, organizing environments, containers and operational controls for product growth.

   System

   Product architecture

   • Python
   • Solidity
   • FastAPI
   • LangGraph
   • LangChain
   • LangSmith
   • Claude API
   • Agno
   • Google Vertex AI
   • DeepInfra
   • GraphRAG
   • Neo4j
   • RAG
   • Mem0
   • pgvector
   • Hexagonal Architecture
   • Clean Architecture
   • TypeScript
   • React
   • Next.js
   • React Native
   • Android Studio
   • Hyperliquid
   • LI.FI
   • Relay
   • Alchemy
   • Privy
   • Stripe
   • End-to-end encryption
   • Microsoft Clarity
   • Firebase Analytics
   • Meta App Events
   • Singular SDK
   • PostgreSQL
   • Redis
   • Celery
   • Docker
   • DigitalOcean
   • Infisical
2. 

   Dec 2025 – Sep 2026

   Argentina · Remote

   In production

   Software EngineerTeamcubation ↗ (opens in a new tab)

   The promotions platform for Payway, a leading Latin American fintech, serving an ecosystem of 350,000+ merchants, 70+ card issuers and 13 countries.

   • Built a GraphRAG assistant and its harness to answer merchants’ questions about promotions and company policies through the merchant portal, connected to the Spring WebFlux BFF. It combined Neo4j graph context with PostgreSQL/pgvector retrieval using LangChain, LangGraph and OpenAI API, and passed 95%+ of a 500+ case eval suite graded in LangSmith by an LLM-as-a-judge validated against human review. Guardrails kept it on the promotions domain and prevented data leaks; the assistant was read-only and never executed commercial actions.
   • Delivered the event-driven pipeline that ingested 20M+ historical promotions into production, followed by thousands of new promotions daily: a serverless flow with Amazon S3, SQS and a Python/FastAPI Lambda using ordered processing strategies, connected to Java/Spring Boot and Node.js/NestJS microservices that loaded their respective PostgreSQL databases. The complete production load was validated with zero data loss.

   • 95%+

     pass rate on a 500+ case eval suite

     Promotion assistant answers graded in LangSmith by an LLM-as-a-judge validated against human review.

   • 20M+

     promotions loaded

     The complete historical load, executed in production, followed by thousands of new promotions daily.

   • Zero

     data loss

     Validated across that same complete production load.

   Key engineering contributor with substantial component ownership within the team. Application development, configuration, observability and diagnostics in Docker and Kubernetes environments.

   More contributions (6)

   • Delivered the whole promotions experience: a Single-SPA React and TypeScript microfrontend, organized by feature and using a design system, embedded as a native section of the merchant portal and connected through a Spring WebFlux BFF to Java/Spring Boot and Node.js/NestJS microservices and providers.
   • Built Java/Spring Boot and TypeScript/Node.js/NestJS microservices to ingest, normalize and process promotions from merchants, financial entities and providers, adapting contracts per integration.
   • Built Python and Node.js tooling to replay bulk loads safely against isolated local databases, reconcile results and surface failed promotions and edge cases before release.
   • Integrated the HTTPS merchant portal with internal HTTP enterprise services through Amazon API Gateway.
   • Structured backend services with layered architecture and clear responsibilities.
   • Built a project-specific agentic engineering environment to preserve context across integration repositories and support consistent feature delivery.

   System

   Portal, ingestion and AI architecture

   • Java
   • Spring Boot
   • Spring WebFlux
   • TypeScript
   • Node.js
   • NestJS
   • Python
   • FastAPI
   • React
   • Single-SPA
   • PostgreSQL
   • MongoDB
   • pgvector
   • LangChain
   • LangGraph
   • LangSmith
   • OpenAI API
   • GraphRAG
   • Neo4j
   • AWS Lambda
   • Amazon API Gateway
   • Amazon S3
   • Amazon SQS
   • Docker
   • Kubernetes
   • OpenTelemetry
   • GitLab CI/CD
3. 

   Jan 2024 – Present

   Argentina · Remote

   In production

   AI Engineer & Full-Stack DeveloperSelf-employed

   Occasional, project-based work for e-commerce and digital-product clients, focused on AI automation and shipping complete products quickly.

   • Automated manual processes that used to take hours or days so they now run in minutes: n8n, ManyChat and Meta Ads workflows, and content agents built with Claude and MCP that research, write and publish.
   • Delivered complete websites, custom software and e-commerce projects in days or weeks, giving clients a reach they couldn’t get before. Across e-commerce and digital products, that work generated USD 10,000+ in client profit.

   • Hours → minutes

     manual processes automated

     Processes that took hours or days now run in minutes through automations and content agents; timings vary by workflow and client.

   • USD 10,000+

     cumulative client profit

     Profit generated for clients by the websites, custom software and e-commerce projects delivered; not personal revenue.

   Independent work for individual clients. These figures add up several client projects, not Pequeverso alone.

   More contributions (2)

   • Delivered Pequeverso’s digital-products storefront with Node.js and Next.js on Hostinger, and built its AI shopping assistant: GraphRAG on LangGraph combining lexical, vector and graph retrieval, grounded in the store catalog, with server-validated prices and links. It answers product questions without executing purchases and is not live yet.
   • Built back offices, stock and point-of-sale extensions, WordPress solutions, APIs and full-stack integrations, with Python and Node.js tooling for automation, bulk processing and data analysis. Used Kotlin and Swift in freelance and personal mobile projects.

   View case studyPequeverso client website (opens in a new tab)

   • Claude API
   • Model Context Protocol
   • LangGraph
   • GraphRAG
   • Python
   • Node.js
   • TypeScript
   • Next.js
   • React
   • Kotlin
   • Swift
   • REST APIs
   • WordPress
   • Hostinger
4. 

   Dec 2024 – Dec 2025

   Bahía Blanca, Argentina · On-site

   In production

   Software EngineerCooperativa Obrera ↗ (opens in a new tab)

   A permissions-management back office for authorized administrative staff, system administrators and managers at Argentina's largest consumer cooperative: 150+ branches, 2.7M+ members and an ecosystem of 50+ internal systems.

   • Designed and built a permissions-management back office integrating 10+ enterprise systems; replaced roughly ten minutes of manual permission lookup with an automated workflow taking under ten seconds, giving administrators a consistent experience across different databases and permission models.
   • Standardized system onboarding with reusable structures, filters and validation, reducing permission-related tickets by at least 80% and eliminating at least 95% of recurring assignment bugs, based on before/after tickets and logs.

   • 10 min → <10 s

     manual to automated permission lookup

     Replaced a roughly ten-minute manual search with an automated back-office workflow, reducing manual steps through the system's architecture and integrations. Separately, critical endpoints averaged below 300 ms in repeatable on-premise tests, supported by SQL tuning, indexing and pagination.

   • ≥80%

     fewer permission tickets

     Compared against tickets and logs before and after standardized onboarding; ≥95% of recurring assignment bugs eliminated.

   • 10+

     systems integrated

     The systems within the permissions back office’s scope, integrated personally, inside an ecosystem of 50+.

   Principal design and implementation responsibility for the permissions-management product and its integrations.

   More contributions (5)

   • Built a responsive React/Next.js and TypeScript frontend, organized by feature and using a design system. It communicated exclusively with a Python/FastAPI BFF, which standardized views and integrated Java/Spring Boot, Node.js/NestJS and PHP services through OpenAPI/Swagger contracts.
   • Connected enterprise systems through Java/Spring Boot and TypeScript/Node.js/NestJS microservices using layered architecture, encapsulating their MySQL or MariaDB databases and permission models; PHP services and legacy applications were also integrated through the BFF.
   • Tuned SQL queries, indexes and pagination; critical endpoints averaged below 300 ms in repeatable tests on enterprise on-premise servers, a separate measurement from the automated permission-lookup workflow.
   • Automated bulk loads and data validation with Python; combined LDAP authentication, role-based access and exportable audit reports with MySQL, MariaDB and Redis. The web development team took over maintenance after my departure.
   • Built a personal cross-system agentic engineering environment to preserve context and work in parallel across independent integration repositories.

   System

   Permissions system architecture

   • Python
   • FastAPI
   • Next.js
   • React
   • TypeScript
   • Java
   • Spring Boot
   • Node.js
   • NestJS
   • PHP
   • OpenAPI
   • SQL
   • Docker
   • LDAP
   • RBAC
   • MySQL
   • MariaDB
   • Redis
   • Prometheus
   • Grafana
   • GitLab CI/CD