Software 17 May 2026 5 min read Title: A lantern, and a thank-you: introducing Felinar By Old Forge Technologies 95 reads Contents There is a particular silence in a house where someone is living with dementia. It is not the absence of noise. It is the absence of conversation that goes anywhere. The same question, gently, for the ninth time before lunch. A story that starts and does not find its end. A long afternoon where the person you love is present in the room and somewhere else entirely. I have been building something for that silence. Its name is **Felinar** - Romanian for *lantern*. This is the first time I have written about it in public, and it feels right that the occasion is also a thank-you. ## What Felinar is Felinar is a voice companion for elderly people living with dementia. You talk to it. That is the entire interface. No buttons to find, no icons to decode, no screen to unlock - because dementia takes those things away first, and usually in that order. Speech, for most of the long middle of the disease, stays. A family member sets it up: the names that matter, the places that matter, the routines of a day - the time the grandchildren usually call, the orchard that comes up in every good story. From then on, the person simply speaks, and something speaks back. It remembers across days. It redirects gently when a memory slips rather than correcting in a way that wounds. It keeps company through the flat hours when the carer cannot be in two places at once. It is built to be a *product*, not a one-off. Families subscribe; the subscription carries the cost of the conversation. That arrangement is deliberate. Dementia moves a person further from a keyboard every month, while the family around them stays able to operate one. ## Why it exists I will not dress this up. Felinar exists because of one house, one woman, and her daughter who carries almost all of it alone. The alpha has exactly one user: My grandmother. Every feature that ships, ships because of how she responded to it on a given afternoon - not because a roadmap said so. That is also why the alpha is in Romanian. Felinar is designed to be multilingual from the foundation up, and English, German and Spanish are on the roadmap behind it. But the first language had to be hers, and Romanian happens to be one of the harder cases - the warmth, the regional cadence, the diminutives an elderly woman expects from someone speaking kindly to her are exactly the things most voice systems, tuned in English, do not reach for. The wager is simple: get it right for the difficult listener and the easier ones come almost for free. The unglamorous parts matter as much as the voice. Every conversation with a vulnerable user is logged and retained server-side - not opt-in, not partial. A product that talks to someone who cannot always consent in the moment owes its family a complete and honest record. That is not a feature. It is a condition of being allowed to do this at all. ## The hard problem If you take one technical thing from this piece, take this: the entire product lives or dies on **turn-taking**. Dementia slows speech. Pauses lengthen. Sentences restart. A companion that jumps in at the first half-second of silence is rude, and the rejection is immediate and permanent. One that waits too long is absent. The usable window between *rude* and *asleep* is narrow, and it is not where general-purpose voice assistants are tuned, because they were built for people whose speech does not break down. Getting that window right - and making the voice itself warm enough that it is welcomed into the room rather than refused - is the work. Everything else is plumbing around it. ## The thank-you Which brings me to ElevenLabs. The current build runs on a general-purpose audio stack because it was the fastest route to something real in front of a real person. It is not what carries Felinar to a thousand families in five languages. Voice quality and cross-lingual warmth are not a feature of this product; they *are* the product, and ElevenLabs is built squarely at that problem. I applied to the **ElevenLabs Startup Grants** programme. The first answer was no - a category misread, an honest one, quickly made right when I explained. The second answer was yes. I want to be plain about this, because grant programmes rarely get public credit for the part that actually matters: a team that will look again when someone makes a reasonable case is doing the thing properly. I am grateful for the credits, and more grateful for the reconsideration. The grant funds a specific, honest experiment: rebuild Felinar's voice path on ElevenLabs, run it side by side with the current stack against the same user, and measure the things that are easy to fudge and shouldn't be - rejection, conversation length, turn-taking failures per session, and a carer's plain answer to *did that sound like a person*. If ElevenLabs is the better voice for someone with dementia, the data will say so. If there is a gap, we will know exactly where it is. Either way it gets written up, in the same plain English as everything else here. <a href="https://elevenlabs.io/startup-grants"><img src="https://eleven-public-cdn.elevenlabs.io/payloadcms/pwsc4vchsqt-ElevenLabsGrants.webp" alt="Supported by the ElevenLabs Startup Grants programme" style="width:250px"></a> A lantern is a small thing. It does not light the whole road. It lights enough of it, for the person carrying it, to take the next step without fear. That is the whole brief. Thank you to everyone helping me build one.
