~$whoami
Software Engineer
Juan Ignacio Castore Bergioli
Building
Buenos Aires, Argentina
scroll
About
Software Engineer and UTDT graduate (GPA 9.13) specializing in advanced AI architectures: Agentic Workflows, RAG systems, and end-to-end automation solutions. Proven track record leading complex technical integrations in high-pressure startup environments. Bilingual Spanish/English (C1), with foundational German (A2).
Experience
AI Engineer — Operational Agent Owner
FEX (OpsTech Startup — AI-Assisted Legacy ERP Integration)
BackendAIInfraLeadershipProduct
- ›Lead of FEX's Operational Agent: the AI agent that handles each client's operations directly against their own ERP. My concrete operational responsibility: no incoming receipt is lost, and uptime stays >99% for the 200+ organizations operating daily. I define product roadmap and features; I report directly to the CTO.
- ›Stack I maintain: backend Python + FastAPI + SQLAlchemy 2 + Alembic on Cloud Run over Postgres; frontend Next.js 15 + React 19 + Shadcn + Clerk + React Hook Form/Zod; reading-service Python on Cloud Run with AWS Textract + OpenAI for structured invoice extraction, queued via Cloud Tasks. End-to-end logging and monitoring to sustain the uptime SLA.
- ›FEX integrates 6 Argentine ERPs in production today (Bejerman, Onvio, Carena, Zetti, Softland, Macrosistemas). I led the Bejerman integration (Thomson Reuters legacy ERP), resolving interoperability with undocumented on-premise SOAP/XML systems; I contribute code to several adapters in addition to owning the product.
- ›Led the integration with ARCA (formerly AFIP, the Argentine tax authority): validation of CUIT, CAE, sales point and amount of each invoice against the agency before loading it into the ERP — a critical compliance step for operating in Argentina.
- ›Every feature on the Operational Agent starts with an explicit business case: resource cost vs. expected impact on the company AND on the end user. Scope is right-sized accordingly — from 2-day fixes to month-long builds — to avoid spending resources on solutions that don't move the metrics that matter.
- ›Closed-loop measurement discipline: every shipped feature is instrumented at the user level, and post-ship metrics feed back into the next prioritization round (what to double down on, what to deprecate, what to build next). The data loop is central to the decision process, not an add-on.
- ›Designed and implemented the invoice auto-loading flow that eliminated manual form-filling for 40% of processed invoices, dropping user time to zero in those cases.[40% de facturas procesadas sin intervención del usuario]
- ›Designed the multi-role invoice approval workflow (inspired by GitHub PRs and SAP release strategies) for enterprise expense control and evolved it to Phase 2 self-service: each organization configures its own routing rules from an editable UI, no code changes. Stack ~88% ready for pilot.
- ›Led a 100% AI-native development culture, making architecture decisions and accelerating delivery cycles through AI tooling.
- ›Picked by the CTO as the engineering reference to evaluate Claude Code: I mastered it in ~1 month, concluded it was worth adopting, and led its rollout while training the team. I scaled usage beyond engineering —operations, support, sales and leadership, including the CEO— to 100% of the company (~15 people): with Claude connected to the repos, the whole organization shares context and resolves technical and product questions autonomously.[Adopción del 100% de la empresa (~15 personas, todas las áreas)]
- ›Within the first 2 months of the engineering team's adoption, the practices and workflows I proposed lifted development throughput to 8.8× the prior baseline —a metric we track week over week— against an internal target of 15×.[Ritmo de desarrollo 8.8× vs. baseline (meta 15×)]
- ›Implemented learning-documentation workflows into a central company knowledge repository, bringing to an entire organization the knowledge-management vision I had originally conceived at Kora.
Co-Founder & CTO
Orion Studio Automation
BackendAIInfraLeadershipdata
- ›Co-founded Orion, an automation consultancy focused on the insurance sector. Single paying client to date (insurance broker, recurring pilot), with active conversations with Ar-Vida and Paraná Seguros insurers to scale the solution across their full broker networks.
- ›Designed and implemented a 5-tier entity resolution pipeline (canonical normalization, token subset/overlap analysis, Levenshtein fuzzy matching, LLM semantic validation) matching ~4,280 contacts against ~250 weekly policy renewals, with automated delivery via WhatsApp. Target match rate >90%.
- ›Self-hosted cloud architecture: Hetzner VPS CX23 (Ubuntu 24.04) with Docker Compose (n8n + PostgreSQL 16 + Redis + Evolution API / WhatsApp gateway), Cloudflare zero-trust tunnel, automated daily backups to Google Drive. Migrated from a prior on-premise deployment on a repurposed personal machine.
- ›As CTO, I own infrastructure, major implementations, and server operations; I keep the platform stable enough that my co-founder (CEO) can directly edit specific n8n nodes and handle client needs without blocking on me.
- ›First-class multi-tenant architecture: the same stack serves multiple brokers/insurers with logically separated data (Paraná Seguros + Myrna/AR VIDA today), ready to scale to full networks via insurer partnerships.
- ›FastAPI BFF (`orion-api`) exposing CSV/contact upload to the frontend and orchestrating n8n calls; auto-deploy to the VPS via GitHub Actions over SSH on every merge to main.
- ›FASE 2 in development: automatic capture of incoming WhatsApp payment receipts, stored in hierarchical filesystem YYYY/MM + Postgres, ready for posting to the insurer's portal.
Co-Founder & CTO
BackendFrontendProductLeadership
- ›Led full-stack development of a gastronomy ordering platform with scalable architecture (React, Node.js, Firebase).
- ›Designed and implemented RESTful APIs and integrations with Mercado Pago and Fudo.
- ›Managed the technical roadmap, real-user validation, and public product launch.
- ›Automated internal processes using Python and serverless architecture, improving operational efficiency.
- ›Built the 'kitchen ticket printer' as a Node.js service running on Raspberry Pi: real-time Firestore listener (`onSnapshot`) that prints each new order on a thermal printer and marks it as printed, ensuring no order is lost in the kitchen.
Teaching Assistant
Universidad Torcuato Di Tella
Teaching
- ›Teach practical classes and evaluate assessments in Algorithm Design, Computer Systems, and Introduction to Programming.
- ›Developed the ability to explain complex technical concepts clearly to diverse audiences — directly applicable to technical consulting and communicating with non-technical stakeholders.
Projects
befeproduct
- ›Independently built and deployed a marketplace connecting users with local service professionals in San Juan (plumbers, electricians, psychologists, etc.) with direct WhatsApp contact and event tracking.
- ›Stack: Next.js 15 + React 19 + Tailwind CSS 4 + Supabase (PostgreSQL + Auth + Storage), Server Components for SEO, RLS and Server Actions, auto-deploy to Vercel from GitHub.
- ›Live features: public directory with full-text search, self-service professional registration, department-level geolocation, WhatsApp contact with lead tracking.
aibeproductlead
- ›Co-founded Kora in 2025: SaaS for digital onboarding automation. Today it runs with a dedicated team and my role is advisory — my active focus is on FEX, Orion and UTDT.
- ›Designed and implemented the RAG system with a dynamic knowledge base: plain-text files as user-editable memory, with an index system for efficient contextual access.
- ›Developed the user-editable 'Agent Memory' feature — direct context management and real-time customization of AI behavior.
Conversational AI Agent for Soyana (Undergraduate Thesis)
2025aibeproduct
- ›Developed an autonomous commercial conversational agent using AI (Gemini, Whisper) to automate order intake, orchestrated with n8n on GCP (139 nodes, 11 use cases).
- ›Specialized multi-agent architecture: Router Agent, Conversational Agent, Order Validator — each with a focused task to optimize debugging and prompt performance.
- ›Automated 11 critical use cases including order intake with disambiguation, volume discounts, order history queries, and human handoff.
- ›Measurable result: customer lookup time reduced from 4 min to 30 sec (8× improvement); monthly catalog and weekly reminders fully automated.
mialquiler.app — Property Management SaaS
2025befeproduct
- ›Independently built a full-stack SaaS application (PostgreSQL/Supabase) to automate administrative management between landlords and tenants.
- ›Validated with a real user (multi-property landlord) who confirmed willingness to pay and provided feature prioritization feedback.
Orbital Trajectory Optimization (Engineering Modeling)
2025dataother
- ›Solo project: parametric optimization pipeline for LEO orbits via massive grid search (10K+ candidates) with physical constraint validation (200–2000 km band, peak-to-peak radial variation, angular monotonicity).
- ›Stack: Python (NumPy, Pandas, SciPy.optimize). Implemented Newton-Raphson on Rosenbrock + orbital stability analysis + atmospheric density simulation.
Audio Deep Learning — Genre Classifier + Autoencoder
2024aidata
- ›End-to-end deep learning pipeline for audio in PyTorch (3-person group project, TDVI course): 2D CNN over mel-spectrograms with BatchNorm/Dropout/data augmentation and experiment tracking in Weights & Biases.
- ›Companion project: 1D convolutional autoencoder for unsupervised audio representation (110K → 27K dims), symmetric encoder-decoder with stride-2 convolutions.
University Machine Learning Competition
2024aidata
- ›2nd place. Built an ML pipeline with CatBoost as the final model, including feature engineering, exploratory analysis, and hyperparameter tuning.
Profitability Analysis — Wind Turbine in Patagonia (Monte Carlo)
2024dataother
- ›Consulting case for the course Computational Applications for Business (UTDT): Monte Carlo simulation to evaluate the profitability of installing wind turbines in Argentine Patagonia, modeling wind variability, energy price variability, and turbine failure risk as stochastic variables.
Schedule Optimization — Retiro–Tigre Train Line
2024dataother
- ›Algorithm Design coursework (UTDT — the course where I'm now a Teaching Assistant): formulated and solved a scheduling problem for the Retiro–Tigre train line, maximizing demand met over operating cost, modeling tradeoffs between frequency, capacity, and peak usage.
Facility Location — Combinatorial Optimization
2024dataother
- ›Optimization coursework in ACN (UTDT): optimal assignment of 30 clients to 10 telephone exchanges in greater Buenos Aires using a given distance matrix — classic facility location problem solved with constructive heuristics.
bondiCounter — CABA Bus Tracking via QR
2024befeproduct
- ›Independent project with real traction: real-time tracking of buses in Buenos Aires (React + Leaflet + Node.js + public city transport APIs). Distributed physical QR codes on many bus stops so passengers could scan and see where their bus was.
- ›Got genuine scans from real passengers but the model wasn't profitable — full lifecycle learning: idea → MVP → physical distribution → real usage → kill due to unit economics.
Coffee Market — Dynamic Pricing (Strategic Management)
2025productleaddata
- ›Capstone project for the Strategic Management elective (business minor, UTDT): Business Model Canvas for a coffee chain with stock-market-style dynamic pricing. Acted as CTO of the 5-person team: defined pricing strategy and the technical pipeline.
Deep Q-Learning Agent for Connect 4
2025aidata
- ›Group project for the Artificial Intelligence & Neuroscience course (UTDT): Deep Q-Learning agent in PyTorch trained to play Connect 4, with experience replay, target networks and epsilon-greedy exploration. Result: DQN agent consistently beats random baseline (>70% win rate).
ADO-1 — Autonomous Digital Organism (Personal Research)
2026 – Actualidadaiinfra
- ›Ongoing personal research: architecture of an autonomous digital organism (ADO-1) with an OODA loop (Observe-Orient-Decide-Act), four-tier cognition (S1 executor, S2 strategist, S3 supervisor, S4 override) and active guardrails against malicious patterns. Isolated in a Docker sandbox with `network_mode: none`, executes controlled actions on the host via PowerShell.
- ›Stack: Python + Docker + OpenAI API (with Ollama local fallback). Next step: release as open source.
Personal Investment Portfolio — AI-Assisted
2025 – Actualidaddataother
- ›Active management of personal investment portfolio with a custom-built dashboard (Streamlit + yfinance + Plotly + pandas): evolution analysis, market cap comparison, ARS/USD/crypto normalization. Built in intensive collaboration with AI — a natural extension of the AI-native practice from work into day-to-day financial decisions.
Skills
Languages
Python (experto)
Go
JavaScript / Node.js
Java (Spring Boot)
C / C++
R
Assembler
SQL
AI & Data
Agentic Workflows
RAG Architectures (Retrieval-Augmented Generation)
Machine Learning (CatBoost, XGBoost, regresión, clustering)
Deep Learning (PyTorch, TensorFlow, redes neuronales)
NLP (Hugging Face transformers, spaCy)
scikit-learn
LLM Integration (OpenAI, Gemini, Whisper)
AWS Textract (extracción estructurada de documentos)
Apache Spark (MapReduce, pipelines escalables)
Simulación Monte Carlo
Weights & Biases (experiment tracking)
Cloud & Infrastructure
Linux (Pop!_OS daily driver, Ubuntu, Debian — uso avanzado; Kali Linux en materia de Ciberseguridad de UTDT, labs tipo CTF atacando/defendiendo servidores)
Google Cloud Platform (GCP — Cloud Run, Cloud Tasks, Functions)
AWS (Textract para extracción de documentos)
Hetzner (VPS, on-premise self-hosted)
Firebase / Supabase
Docker / Docker Compose (uso intensivo en producción)
Kubernetes (formación académica, materias de infra en UTDT)
Raspberry Pi (servidor on-premise para tesis y proyectos reales)
Arquitectura Lambda / Serverless
Cloudflare zero-trust tunnels
CI/CD (GitHub Actions, auto-deploy via SSH a VPS)
Observabilidad y Monitoring
Microservicios
Backend & APIs
Node.js / Express
Java Spring Boot
Firebase Functions
APIs RESTful (diseño y desarrollo)
Integración con sistemas legados (SOAP/XML)
n8n (orquestación de workflows)
Frontend
React
Svelte
Next.js
HTML / CSS
Vite
Databases
PostgreSQL
MySQL
MongoDB
Firebase Firestore
Redis
Cassandra
Álgebra relacional
Product & Methodology
Desarrollo 100% AI-native
Agile / Scrum
Product Discovery
Consultoría técnica
Education
BSc in Digital Technologies (Major in Computer Science, Minor in Business)
Universidad Torcuato Di Tella
GPA: 9.13 / 10. Completed 39 courses (7 electives in AI, Infra, and Data Engineering — 7 beyond the 32 required to graduate)
Certified Tech Developer (Backend)
Digital House
Full scholarship by Mercado Libre and Globant
Electromechanical Engineering (3 years completed)
Universidad Nacional de San Juan
Transferred to Digital Technologies after discovering a passion for computing. Strong foundation in basic sciences and systems thinking retained from the engineering track.
Saint Paul — Bilingual English Program
Saint Paul (San Juan, Argentina)
Bilingual foundation underpinning current C1 proficiency.
Achievements
★2nd place in university Machine Learning competition (final model: CatBoost)
★Full scholarship from Mercado Libre and Globant for Certified Tech Developer (2021–2023)
★GPA 9.13 — BSc in Digital Technologies, UTDT
★GitHub Pull Shark achievement · 71 repositories · GitHub Pro · founder of the Orion-Studio-Automation organization
Languages
Spanish
Native
English
C1 – Fluent
German
A2 – Basic