AI Dev & Fullstack · Healthtech

I don't just train models — I solve business problems and ship them to production.

10 years building clinical software —FHIR, HL7, LIS, PACS/HIS— and talking face to face with the people who use it.

Years in health tech: 10
Freelance clients: 8+

About

A domain engineer, not just a model engineer

I got into healthcare through IT support and ended up writing the integrations. That route taught me something a course can't: I know what happens when a result arrives late.

Today I work at Wiener Lab on a LIS handling tens of thousands of requests a day, and in parallel I run freelance projects end to end: discovery, architecture, development, deploy, and the uncomfortable call when something breaks. I'm studying AI at Universidad Nacional de Rosario.

Direct client work

8+ freelance clients with no project manager in between. Turning a fuzzy business problem into a scope you can defend is part of the job, not an extra.

Technology with a purpose

I care about AI you can audit: guardrails, explicit limits, and systems that fail visibly. The chatbot on this site is built on that idea.

Real clinical domain

FHIR, HL7, LIS, PACS/HIS, medical imaging. Ten years learning why a lab result is not just another row: traceability, privacy and clinical consequences.

AI projects

Computer vision, natural language processing and machine learning

All held to the same standard: each thing is explained by what it actually does, or it's said to not exist yet. The vision projects show their steps with real images from every stage; this site's assistant opens its whole pipeline in the lab.

1–1/7

Open lab
The assistant that answers on this portfolio
RAG with measured guardrails and an agent that knows when to search again.
It only answers about me, and that isn't a line in the prompt: it's a similarity threshold measured against a question set. The lab shows it deciding, fact by fact, and explains the graph behind it.
- Python
- FastAPI
- LangChain
- +2
1. guard
2. retrieve
3. agent
4. tools
5. stream
Open lab
The agent architecture, opened up
The LangGraph graph behind every answer.
Four nodes: two deterministic guardrails, the agent with its three tools, and the stream. Explained node by node, with the decisions holding it up — why the model doesn't decide scope, and why the iteration ceiling isn't an instruction.
- LangGraph
- ToolNode
- SSE
- +1
Step by step
Custom OCR for summarizing books
I bracket the paragraph and the book transcribes itself.
My reading bottleneck was retyping by hand the paragraphs I wanted to keep. Now I photograph the page, and whatever I marked in the margin comes out transcribed and ready to summarize.
- Python
- FastAPI
- OpenCV
- +2
Step by step
Quality control on a production line
Count and classify pills on the belt, from a single photo.
A camera over the conveyor belt and every pill is located, classified by type and counted. No trained model: colour, shape and morphology.
- Python
- FastAPI
- OpenCV
- +1
Step by step
Licence plate reading at an access point
Finding the plate without looking for a white rectangle.
From a photo of the vehicle come the cropped plate and its seven characters, separated and ready for recognition.
- Python
- FastAPI
- OpenCV
- +1
Step by step
Automatic reading of a dice roll
The hard part isn't reading the face: it's knowing when to look.
From the video of a roll come each die's value and the total, read at the first instant all five stopped moving.
- Python
- FastAPI
- OpenCV
- +1
Under construction
Machine learning
Under construction, and I'd rather say so.
I prefer an empty slot to a notebook with an accuracy I can't defend. Once there's a trained model, evaluated on data it never saw and with the error explained, it goes here.
- Python
- scikit-learn
- PyTorch

Freelance

Systems running inside a real business

Working directly with the owner: discovery, development, going live and support. Each business's data stays out: what's shown are videos of the system.

Click-Bar
Order and pay from your phone, no waiting in line at the bar.
A web ordering app for bars: the customer orders and pays from their phone, the bar team works from a real-time order queue and the customer gets a notice when the order is ready. An admin panel shows the top drink, the average ticket and the shift.
- FastAPI
- PostgreSQL
- Next.js
- Mercado Pago
The problem and the process
The problem
On a busy night the bar gets overwhelmed: lines to order and to pay, tickets lost between the waiter and the bartender, and nobody knows in real time what is pending or what is selling.
Process
1. 01Discovery and contract, with scope, flows and wireframes agreed up front
2. 02Customer app: menu, order and Mercado Pago payment from the phone
3. 03Bar screen with a real-time order queue —waiting, preparing, ready— and a notice to the customer when their order is ready
4. 04Admin panel with the venue's numbers, plus manuals and handover
Bakery management
The order is entered once; everything else follows from it.
A system that follows the bakery's processes: the order is entered once, and production, the day's dispatch and delivery with collection come from it, with the PDF ready to send over WhatsApp. The owners gained time and the peace of mind that everything is traceable.
- Next.js
- TypeScript
- PostgreSQL
- Drizzle ORM
The problem and the process
The problem
A bakery in Rosario with over 90 years of history kept its processes centralized in the owners' hands, in Excel and on paper. With no record, there was no way to follow metrics, processes or accounting.
Process
1. 01Meetings with the owners to uncover each process one by one and see what could be improved
2. 02Turning each process into a solution built to speed things up, not just to put them on a screen
3. 03Presenting each solution and iterating until it matched the real process
4. 04Working modules more useful than the paper and Excel they started from, with traceability for every process
IF Distribuciones
From WhatsApp and prices in someone's head to an app with business analytics.
A sales system for the distributor: orders entered in seconds, automatic invoices with an up-to-date running account, the delivery route and a dashboard with real profit, revenue and what clients owe.
- Next.js 15
- PostgreSQL
The problem and the process
The problem
IF Distribuciones is a young frozen-food distribution company: it buys goods and resells them to bars and restaurants. Everything went through WhatsApp, with prices in people's heads and papers floating around.
Process
1. 01Market study and a look at how the distributor worked, end to end
2. 02First version of the app on top of the mapped processes
3. 03Continuous improvement from the client's feedback
4. 04A production version that solved those problems and added business analytics

Experience

Ten years in one domain, three ways of seeing it

Full Stack Developer
@ Wiener lab.
Aug 2022 — Present · Rosario, Argentina
- Backend for a LIS processing 50K+ lab requests per day; response time down 35%.
- Led the integration of AI features and LangChain agents to automate lab workflows: 40% faster diagnostic result delivery.
- Microservices with Docker and Redis for real-time processing, including medical equipment.
- SQL Server databases holding 2M+ patient records with complex relational schemas.
- HL7 integrations between HIS and PACS.
- CI/CD with Docker: deploys from 4 hours down to 30 minutes.
Backend Developer
@ Freelance
Dec 2020 — Present · Remote
- 8+ clients across e-commerce and logistics, working directly with them.
- Recommendation system with LangChain agents, LLMs and vector databases: 85% user satisfaction.
- Async processing with Celery and RabbitMQ: 10K+ daily jobs.
IT Support Specialist
@ Grupo Oroño
Mar 2016 — Aug 2022 · Rosario, Argentina
- Institution-wide software migration for 500+ users across several healthcare facilities.
- 60% reduction in medical report generation and delivery time.
- HL7 integrations between diagnostic imaging and hospital information systems.

Education

University Technical Degree in Artificial Intelligence

Universidad Nacional de Rosario · 2024 — Present

Full Stack Web Development (545 h)

Polo Tecnológico de Rosario & UNR · 2021

Data Analytics

Coderhouse · 2022

Project Management

EducacionIT · 2022

Skills

What I build with

Languages

Python
TypeScript
JavaScript
Java
SQL
HTML/CSS

ML / AI

LLM integration
LangChain / agents
RAG and vector databases
ONNX Runtime (web)
Scikit-learn
TensorFlow
Model deployment

Automation

LLM agents and workflows
Celery + RabbitMQ
API and webhook integrations
Scheduled jobs (cron)

Backend

FastAPI
Django
Flask
Spring Boot
PostgreSQL
SQL Server
MongoDB
Redis

DevOps & platform

Docker
Ansible
CI/CD
Vercel
OpenTelemetry
Jaeger
KeyCloak
Git

Health domain

HL7
LIS
PACS / HIS
Medical imaging
Result traceability

Frontend

Next.js
React
Angular
Tailwind

Where I'm going

AI Dev in healthcare, with AI you can audit

I want to stay in healthtech, but on the model side: AI systems that support clinical and operational decisions and that can be explained, measured and audited. I'm less interested in the biggest model than in the system that knows when to abstain.

AI applied to diagnostics, laboratory and imaging, with metrics before demos.
Guardrails, evaluation and observability as part of the product, not a last-minute patch.
Teams that talk to end users; that's where the domain is learned, not in the ticket.

AI Dev & Fullstack · Healthtech

A domain engineer, not just a model engineer

01Direct client work

02Technology with a purpose

03Real clinical domain

Computer vision, natural language processing and machine learning

The assistant that answers on this portfolio