LANVAR
← The shelf

LIA

Retired toy

An AI companion for people with severe memory loss — running on a neural network built from scratch, back before you could just call an API.

✳ Best Health Hack · NewHacks 2021 PythonNeural network (from scratch) 2021
✳ How it actually works
THE GUARDIAN guardian's portal adds today's facts KNOWLEDGE BASE facts stack up LIA answers, out loud she asks, it answers — no middleman THE PATIENT “who visited today?” one loop, all day: ask → recall → answer LISTENING… RECALLING… ANSWERING… Q → RECALL → A ? MAYA DID. MAYA DID.
FIG 1 · the loop, animated — she asks, LIA remembers, the guardian keeps the jar full
guardian's portal the facts of a life knowledge base who, what, where, when our own network hand-built, trained on the guardian's facts the patient asks “who visited today?” no third party needed dignity, restored ← the answer comes back straight to them NewHacks 2021 · Best Health Hack reborn three years later as NURO ↗
FIG 2 · the machinery — guardian’s knowledge → our own network → patient’s answers
The story

Inspired by a relative living with memory loss, and built with Abhigyan Praveen, Gianluca Piccirillo, and Matthew Cabral at the University of Toronto’s NewHacks 2021.

This was 2021 — a year before ChatGPT. There was no API to call, so the network was ours: hand-built and trained on the patient’s own knowledge base. A guardian fills a portal with the facts of a life, the model learns them, and the patient can ask directly — who visited, what day it is, where things are — and get an answer without a human intermediary on call. Building the thing yourself teaches you where the answers actually come from.

LIA won Best Health Hack, and the idea refused to die: three years later it was reborn as NURO — same goal, but with real conversation recording and semantic memory over embeddings instead of a network trained by hand. Every retired toy on this shelf is a seed.

Next on the shelf Green Switch →