Research narratives.

Early-childhood therapy for kids with a disability certificate (CUD) lives or dies on what happens at home between sessions — yet that’s exactly where therapists lose visibility and families feel overwhelmed. Orbe bets the gap is operational, not clinical.

A continuity platform where a therapist activates professionally-validated micro-activities, delivers them to families over WhatsApp — no login, no new app to learn — and gets structured progress back, so the next session starts with real data instead of guesswork.

Asynchronous by design: the therapist picks activities from the platform; an async flow dispatches clear instructions through WhatsApp; families respond in the tool they already use; progress signals return to the therapist. Built for accessibility (WCAG 2.1 AA) and near-zero operational load on the family.

• Chose WhatsApp over a dedicated app — for an overwhelmed parent, adoption beats features every time
• Chose asynchronous over real-time — care shouldn’t depend on the therapist being online at the exact moment an activity happens

Won Startup Weekend Rosario 2025. Co-created with licensed professionals across occupational therapy, speech, psychopedagogy and psychology, and now live with a B2B model aimed at cutting treatment drop-off (an 18–25% problem for clinics).

Open alpha with consultorios in Rosario, validating adherence and drop-off metrics against the B2B pricing model — then expand the validated-activity library across therapeutic areas.

Pieza started as a curated marketplace for local artisan objects: one place for pieces that otherwise depended on fairs, algorithmic chance, or word of mouth to be discovered.

Paused. The product and curation held up, but interviews with artisans surfaced the harder truth: they would pay only if Pieza brought new customers and traffic, not just another place to show their work.

A curated storefront for independent makers: selected pieces, editorial context, local checkout, and a professional buying experience. It solved display better than discovery.

• Built the storefront first because the visible problem looked like fragmented discovery and uneven presentation
• Paused after realizing the real value would have to be demand generation, not catalog infrastructure

The metaphor changed from storefront to street. A beautiful window does not sell if nobody walks past it; distribution turned out to be the product.

Return only from a distribution-first angle: how to create attention, trust, and buying intent for local creators before polishing the storefront again.

Recruiters accumulate data across spreadsheets, ATS exports, notes, and transcripts — but when a new search opens, recovering that context costs more than starting fresh. Recollect started as a tool to turn accumulated data into reusable recruiting memory. The promise: never lose a good candidate again.

Pivoting. The recruiting memory system reached a solid architectural direction — durable memory separated from ephemeral outputs, provenance tracking, human confirmation as the gate, qualitative matching, and anti-features as a trust contract. But the architecture proved relevant beyond recruiting, which opened a broader question.

A copilot responds in the moment; a memory improves the next moment. Serenity, the agent, was a persistent surface for search, comparison, and memory recovery — not a decorative chatbot. Candidate Notes became memory only when confirmed by the recruiter. Matches showed evidence, gaps, and risks instead of scores. Provenance tracked every signal to its source.

• Started as Talent Rediscovery — import a Talent Pool, search in natural language, get evidence-backed shortlists
• Shifted from "AI search" to "the system around the AI" — provenance, confirmation, and limits became the product
• Separated durable memory (Candidate Records, Notes, Search Requests, Interview Summaries) from ephemeral outputs (Shortlists, Matches, Briefings)
• Defined anti-features as trust contract: no auto-sending, no auto-merging, no performance prediction

The product outgrew its original thesis. The hardest value was the system around the model — context, limits, confirmation, traceability — not the model call. The memory architecture proved relevant beyond recruiting, which changed the map.

Pivoting. The direction is decided but not detailed here yet. What Recollect proved: useful AI in a product is not putting a model in the middle of a flow — it's building the system that lets the AI remember, explain, limit itself, and work under human supervision.