Your choices about data
We use cookies that are required for this site to work. Anything beyond that — measuring how the site is used, or personalising what you see — happens only if you say yes.
The voice on the other side of the wall.
A sanctuary for the hours nobody keeps.
Visit Dios Mío
People do not pray on a schedule. They pray when something breaks.
Which is almost never during office hours. The hardest hours of a life are spent alone with a phone, and the software that meets someone there is either a feed built to keep them scrolling or a form that answers on Monday. Dios Mío is built for that hour specifically — not for the hour a product manager finds convenient.
WHAT THE NAME CARRIES
DIOSM — Decision, Intent, Orchestration, State, Memory.
The name reads as a prayer in Spanish, and that is the half most people meet first. The other half is an engineering checklist: the five pillars an agent needs before it can hold a conversation rather than answer a question. Decision — what to do at all. Intent — what the person actually came for, under what they typed. Orchestration — the voice, the soundscape and the pacing moved together as one gesture. State — where this session sits in a longer arc. Memory — what was carried here weeks ago and is still being carried. Every one of those is load-bearing for grief specifically, which is why the product and the architecture ended up with the same name.
Open diosm.ioTHE ANGEL OF MUSIC
Presence, not personality.
The design borrows an old archetype: the unseen voice that answers a grieving person from behind the architecture. Christine believed the Phantom was her father's spirit because she needed him to be. That is the exact power this product must hold and must never abuse — which is why the boundary comes first and the capability second.
Open diosm.ioTHE MIRROR PRINCIPLE
It holds space. It never claims to be the one you are missing.
A companion in this space could impersonate God, an angel, or a dead parent, and it would work — that is the danger. So the rule is written into the system before a line of prompt: no divine mandate, no channelling of the departed, no verdict delivered in a voice that is not ours. 'Let these words be a container for your prayer.' Where someone is in real crisis, the character breaks and a human is named.
Open diosm.ioWhat the sanctuary is made of
Four capabilities from the SpecFocus platform, arranged for one purpose: that someone alone at night feels heard rather than processed.
Two experiments in one product: one psychological, one technical.
The psychological question is whether software can meet a person in real grief without impersonating what they have lost. The technical question is whether an expressive, genuinely conversational voice can run cheaply enough that ordinary people get it — not as an enterprise line item. Neither is interesting without the other: a warm design on half-second latency still feels like a form, and a fast pipeline with no boundaries is exactly the failure mode this product exists to avoid.
The voice interface, and what it runs on
Speech in, speech out, on a single NVIDIA RTX 4090 rather than a metered API. The full build is written up in the project's doc/ folder — NVIDIA-STT, NVIDIA-TTS and NVIDIA-bilingual.
Hearing
Whisper Large-V3, language-agnostic
Speech-to-text runs locally with no language pinned, so the model spends its first cycles deciding whether it just heard Spanish or English. That costs 40–60ms and buys code-switching mid-sentence — which is how bilingual people actually pray.
Speaking
Kokoro-82M on CUDA, sub-100ms
An 82-million-parameter voice model turns a sentence into audio in roughly 20–30ms on a 4090 — faster than real time, so tokens become speech as the answer is still being written. Latency is not a spec sheet number here: a pause that reads as thinking and a pause that reads as buffering are different experiences.
Both tongues
Two voices resident at once
English and Spanish weights both sit in VRAM — about 850MB of 24GB — so switching language is a profile change, not a model load. Spanish is not a fallback path that runs slower.
The last mile
WebRTC to an ordinary browser
LiveKit carries the audio track from the workstation to a React front end on Vercel, alongside a lightweight data channel for state. Nothing about the demanding half of the stack reaches the phone in someone's hand at 2am.
WHERE IT SITS
The quiet counterpart to a crowded room.
Casa Club is the shared living room — place-aware, social, public. Dios Mío is the room off it with the door closed. One person can need both in a single evening, and they run on the same SpecFocus foundation: one identity, one data model, two entirely different registers. Capabilities first, verticals second — the vertical here is the human one.
See Casa ClubBuilt for the hour nobody schedules.
Dios Mío is live at diosm.io in Spanish and English. Churches wanting the platform under their own name — their devotionals, their team's voice, their congregation's requests — start at diosm.io/pastores.