Projects
Thesis Project — UNC Córdoba, 2025
Clinical Decision Support System for Fracture Detection in X-Rays
End-to-end CDSS for fracture detection in X-ray images. 97.18% accuracy, 86.6% sensitivity, AUC-ROC 0.973. Now scaling: extending detection from the wrist to the whole body, and taking TraumaVision through entrepreneurship programs such as Emprende U 2026.
Pipeline
DICOM → CLAHE → YOLOv8m → Detection → Report
AI-First Digital Agency · Argentina
Founded and operate an AI-native digital agency. Built a 38-agent coordinated pipeline using Claude API — Sonnet 4.6, Haiku 4.5, and Opus 4.7 with aggressive prompt caching, extended thinking, tool use, and MCP server integrations (Playwright, Google Drive, Gmail, Calendar, Canva, Vercel). 4 products live in production. Argentina, 2026.
2026
Founded
4 LIVE
Sites in Prod
AI-FIRST
Approach
Industrial Hackathon — RTS Group
1st Place · 2025World's first industrial hackathon, organized by RTS Group. 150 multidisciplinary teams of 5 people competing in real industrial engineering challenges. Team Algarroba LID took first place with a proposal for automating human-in-the-loop agents for sensor control in a nuclear power plant.
1ST PLACE
Result
ALGARROBA LID
Team
Case Study · B2B SaaS · Built via Eleven · 2026
A working hospital-equipment platform, and the website it never had.
01 / Problem
KIBBO is a management platform (CMMS) for hospital medical equipment. The product worked, but kibbo.io still ran an old WordPress site that didn't explain it and throttled search crawlers. A biomedical engineer had nowhere to understand KIBBO, see it running, or check who already uses it.
02 / My role
Through Eleven, I led design, front-end development, SEO/GEO and the English edition. The build ran on our AI-agent pipeline, with my review and sign-off at every checkpoint.
03 / Key decisions
Write to the technician, not the buyer.
Third person, no hype, every claim shown on a real product screen, because the reader is the one who will run it every day.
Keep KIBBO and LEEX apart.
The hardware maker's figures are credited to LEEX, so KIBBO stands as its own product and never claims what it can't back.
Give mobile its own system.
"Expediente" reads each page like a case file instead of squeezing the desktop scroll into a phone.
Make English a real edition.
Own slugs, reciprocal hreflang and an untouched Spanish canonical; user quotes stay verbatim in Spanish with a labelled translation, so no one is misquoted.
04 / Result
Live on kibbo.io in Spanish and English, replacing the legacy WordPress, with GA4 measuring from launch.
ES · EN
Live on kibbo.io
11
Pages in English
23
URLs in the sitemap
SEO / GEO
Case Study · Internal B2B System · LEEX · 2026
Finds public tenders across 33 sources and helps sales decide which to bid.
01 / Problem
LEEX, an Argentine medical equipment company, sells mainly through public tenders: demand isn't generated, it's detected across dozens of government portals, each with its own format and deadlines. Seeing a tender late means never having it; seeing it without knowing if you qualify means losing it. The previous tool was a single-user script feeding a spreadsheet.
02 / My role
I led it end to end: scoped it with sales and management, designed the architecture, and built the engine and web app with a pipeline of AI coding agents under my direction.
* Client data and amounts blurred
03 / Key decisions
Code before AI.
A deterministic dictionary-plus-vector match filters first and a Claude model judges only what survives, because that keeps it cheap, testable and explainable.
A person signs every verdict.
The model drafts a requirement-by-requirement case citing the tender and the product sheet, and sales decides, because a missed tender costs far more than a review.
Three sweep states, not two.
"Nothing there", "couldn't see" and "saw far less than usual", because the expensive failure is the silent one.
Public first.
The engine reads each portal's public view and logs in only as a documented last resort, which cut credential exposure and let the build start without waiting on access.
04 / Result
Running internally on real data: the first full pass went from sweep to triage to human-signed verdicts. Next step: onboarding the sales team.
33
Procurement sources swept
19
Sources the old tool never reached
28
Product categories matched
10k+
Automated tests
Projects
Thesis Project — UNC Córdoba, 2025
Clinical Decision Support System for Fracture Detection in X-Rays
End-to-end CDSS for fracture detection in X-ray images. 97.18% accuracy, 86.6% sensitivity, AUC-ROC 0.973. Now scaling: extending detection from the wrist to the whole body, and taking TraumaVision through entrepreneurship programs such as Emprende U 2026.
Pipeline
DICOM → CLAHE → YOLOv8m → Detection → Report
AI-First Digital Agency · Argentina
Founded and operate an AI-native digital agency. Built a 38-agent coordinated pipeline using Claude API — Sonnet 4.6, Haiku 4.5, and Opus 4.7 with aggressive prompt caching, extended thinking, tool use, and MCP server integrations (Playwright, Google Drive, Gmail, Calendar, Canva, Vercel). 4 products live in production. Argentina, 2026.
2026
Founded
4 LIVE
Sites in Prod
AI-FIRST
Approach
Industrial Hackathon — RTS Group
1st Place · 2025World's first industrial hackathon, organized by RTS Group. 150 multidisciplinary teams of 5 people competing in real industrial engineering challenges. Team Algarroba LID took first place with a proposal for automating human-in-the-loop agents for sensor control in a nuclear power plant.
1ST PLACE
Result
ALGARROBA LID
Team
RTS GROUP · 2025
Event
Case Study · B2B SaaS · Built via Eleven · 2026
A working hospital-equipment platform, and the website it never had.
01 / Problem
KIBBO is a management platform (CMMS) for hospital medical equipment. The product worked, but kibbo.io still ran an old WordPress site that didn't explain it and throttled search crawlers. A biomedical engineer had nowhere to understand KIBBO, see it running, or check who already uses it.
02 / My role
Through Eleven, I led design, front-end development, SEO/GEO and the English edition. The build ran on our AI-agent pipeline, with my review and sign-off at every checkpoint.
03 / Key decisions
Write to the technician, not the buyer.
Third person, no hype, every claim shown on a real product screen, because the reader is the one who will run it every day.
Keep KIBBO and LEEX apart.
The hardware maker's figures are credited to LEEX, so KIBBO stands as its own product and never claims what it can't back.
Give mobile its own system.
"Expediente" reads each page like a case file instead of squeezing the desktop scroll into a phone.
Make English a real edition.
Own slugs, reciprocal hreflang and an untouched Spanish canonical; user quotes stay verbatim in Spanish with a labelled translation, so no one is misquoted.
04 / Result
View site →Live on kibbo.io in Spanish and English, replacing the legacy WordPress, with GA4 measuring from launch.
ES · EN
Live on kibbo.io
11
Pages in English
23
URLs in the sitemap
SEO / GEO
Case Study · Internal B2B System · LEEX · 2026
In development — launching soonFinds public tenders across 33 sources and helps sales decide which to bid.
01 / Problem
LEEX, an Argentine medical equipment company, sells mainly through public tenders: demand isn't generated, it's detected across dozens of government portals, each with its own format and deadlines. Seeing a tender late means never having it; seeing it without knowing if you qualify means losing it. The previous tool was a single-user script feeding a spreadsheet.
02 / My role
I led it end to end: scoped it with sales and management, designed the architecture, and built the engine and web app with a pipeline of AI coding agents under my direction.
* Client data and amounts blurred
03 / Key decisions
Code before AI.
A deterministic dictionary-plus-vector match filters first and a Claude model judges only what survives, because that keeps it cheap, testable and explainable.
A person signs every verdict.
The model drafts a requirement-by-requirement case citing the tender and the product sheet, and sales decides, because a missed tender costs far more than a review.
Three sweep states, not two.
"Nothing there", "couldn't see" and "saw far less than usual", because the expensive failure is the silent one.
Public first.
The engine reads each portal's public view and logs in only as a documented last resort, which cut credential exposure and let the build start without waiting on access.
04 / Result
Running internally on real data: the first full pass went from sweep to triage to human-signed verdicts. Next step: onboarding the sales team.
33
Procurement sources swept
19
Sources the old tool never reached
28
Product categories matched
10k+
Automated tests
Stack
Motion & Video
Post-production by AI agents: Claude Code cuts, animates, builds the 3D and renders every piece in Remotion. The MRI explainer and the model comparison are 100% AI-made; the three product reels start from footage shot on an iPhone.
Click to play with sound.Tap to play with sound.
Stack
Tech Stack
Recall 86.6% · mAP50 91.9%
English
B2 — LTC Eastbourne + California
Spanish
Native
