Selected projects
AI architecture and system design for applications, platforms, and critical operational software. Some in production with hard metrics, others conceptual prototypes that demonstrate how I design with AI.
The common thread across all of them: trust as a property of the system — earned in how architecture is decided, how context is handled, and how AI reaches the user, not added on top of a model that works.
TÜV NORD Argentina · Sistema Cóndor
2023–2025 · Enterprise platform · AI-assisted operations · Laravel · Vue
Key results
80% less administrative load
100% of operations centralized into a single platform
2 years in production
TÜV NORD operated with manual processes scattered across multiple tools. Generating a commercial proposal took days of manual work, with no centralized visibility into certifications or audits.
The problem
Internal users with very different expertise levels interacting with intelligent automation. The system had to decide when to act alone and when to request human validation — without frustrating experts or abandoning less technical users. Trust had to scale with experience, not be uniform across roles.
What I designed
The AI generates pre-proposals but always presents them as editable drafts — a design pattern that preserves user control without losing the efficiency of automation. Assisted flows show what the system did and why, so users can validate with context. Validation rules are configured by role, not by generic permissions.
Key results
80% less administrative load
100% of operations centralized into a single platform
2 years in production
Mi Caja · Government Platform
2019–2023 · Platform at scale · 100k+ active users · Laravel · Vue
Key results
70% reduction in processing times
100,000+ active users in production
Full traceability of inter-institutional documentation
Elimination of unnecessary in-person visits
Managing administrative procedures for retired personnel depended on in-person visits and manual coordination between agencies. Each procedure required physical travel, paperwork, and phone follow-up. Information didn't flow between areas, and no one had real visibility into the status of each case.
The problem
Processes dependent on in-person presence, manual coordination between agencies, and zero visibility into the real status of each case.
What I designed
A digital platform that centralizes procedures and coordinates approval flows between agencies, with a web portal, mobile app, and complete inter-institutional workflow. Security and traceability for critical systems — the foundation I now build robust software with integrated AI on.
Key results
70% reduction in processing times
100,000+ active users in production
Full traceability of inter-institutional documentation
Elimination of unnecessary in-person visits
B2B Industrial Support Tickets
2025 · Conceptual piece · AI agents · context engineering · Laravel · Vue
What the design demonstrates
Dual-agent architecture: separate identities for operator and technician, with structured context handoff
Explicit confidence levels on each suggestion, with traceable source
Escalation patterns to humans with no context loss
50% reduction in resolution time (projected)
75-88% of context pre-filled automatically (projected)
A conceptual piece modeling how a B2B industrial ticketing platform could add AI beyond a chatbot.
The problem
The real bottleneck wasn't ticket volume. It was what happened inside each one: incomplete descriptions, back-and-forth rounds, technicians starting from scratch each time. AI couldn't solve this with a single generic system — different roles needed different interactions, and the confidence threshold to act on AI suggestions was different for each role.
What I designed
Two agent systems with distinct identities — one for the operator under pressure, another for the expert technician — connected by a structured context object. Each suggestion includes an explicit confidence level and its source, so users can decide how much to trust each output. The system knows when to escalate and transfers complete context to the human, with no loss.
What the design demonstrates
Dual-agent architecture: separate identities for operator and technician, with structured context handoff
Explicit confidence levels on each suggestion, with traceable source
Escalation patterns to humans with no context loss
50% reduction in resolution time (projected)
75-88% of context pre-filled automatically (projected)
Eunoia Care · AI Assistant for Caregivers
2025–present · Healthtech · Personal prototype · AI architecture · Multi-agent system
Design results
Significant reduction in caregiver cognitive load
Patient medical information available in seconds
Longitudinal medical history built through use, not as a separate task
Families caring for a chronic patient live flooded with fragmented medical information: lab results, ultrasounds, hospitalizations, instructions from different specialists. Forgetting and fragmentation are the main source of anxiety and care errors.
The problem
Designing an experience where a caregiver — often emotionally invested and without clinical training — could capture information without friction, in any format, and recover it when needed, in the format they need. Trust was the central design constraint: without it, the caregiver would never delegate to the system.
What I designed
The complete product architecture and a multi-agent system with distinct specialties (lab interpretation, longitudinal medical history, natural language conversation), each with its own context, tone, and autonomy limits.
The caregiver captures information in any format — photo, text, audio — and the system converts it into a structured event. The AI responds with data, charts, or honestly indicates when it doesn't have the information. I built a functional prototype with agents running on the real model, validating design decisions before implementation — especially each agent's limits and uncertainty handling in a medical context.
The design prioritized three axes: lowering capture friction, building trust through verification and honesty, and respecting the emotional load of the context.
Design results
Significant reduction in caregiver cognitive load
Patient medical information available in seconds
Longitudinal medical history built through use, not as a separate task
Have an AI project in mind, or a system that needs more rigor in its AI layer?
Tell me what you're building and I'll respond within 24 hours.