Technology
vivome is not a RAG chatbot. It's technology that turns records into structure
A semantic event ontology that turns a diary line into standardized event data, on-device-first AI, and a structure where we can't see your data rather than a promise that we won't. Here is how vivome's AI is built, without exaggeration — every technical term is explained in plain words where it appears.
10
Codebook domains
Semantic code system
82%+
Tagging golden-set F1
Continuous regression gate
110+
On-device insight engines
Deterministic · only 1 uses an LLM
6,800+
Automated tests
Jest, passing continuously
01 · Semantic event ontology
One diary line becomes analyzable event data
Most AI health apps store records as free text and, when needed, retrieve documents to feed a language model (RAG). vivome works one step earlier. The moment a record is saved, the original sentence is preserved as-is while it's also decomposed into standardized event units, and patterns like temporal relations, co-occurrence, and personal baselines are read on top. The same ten thousand records, in vivome, are structured data that can be searched, compared, and analyzed from day one — even when one sentence mixes several events, each event and the relation between them is stored separately.
The original, preserved
Your sentences, photos, and icon taps are stored as-is, without judgment. Even when the way it's interpreted changes, the original never does.
Standardized event units
Sentences are decomposed into events using semantic codes from a 10-domain codebook: emotion, body, symptoms, intake, activity, and more.
Pattern interpretation
What rose together, and how it differs from your usual, is computed on top of event data, and spoken in pattern language, not diagnoses.
Semantic codes — names both humans and machines read
Codes carry meaning (headache, walking) rather than numbers (024), so both people and AI can spot an error at a glance.
“I was anxious, so I took a walk”
Event 1 — signal
Emotion · anxiety
Event 2 — response
Activity · walking
Relation — anxiety motivated the walk (event 1 → event 2)
One sentence is stored as two events plus one relation — which is why questions like “what do I usually do when anxious?” can later be answered with data.
In plain words
- RAG (retrieval-augmented generation)
- In plain terms — A technique where the AI searches for relevant documents and pastes them in as reference before writing an answer.
- In vivome — vivome doesn't rely on it — instead of searching documents later, it structures your records the moment they're saved.
- Ontology
- In plain terms — A predefined map of concepts: what something is, what properties it has, and how it connects to other things.
- In vivome — vivome's ontology is its 10-domain codebook: emotion, body, symptoms, intake, activity, and more.
- Event code
- In plain terms — A standard name tag attached to each piece of a record. Different wording, same meaning — same tag.
- In vivome — “My head is throbbing” and “have a headache” land on the same headache code, so they aggregate together later.
02 · On-device-first AI
Your record is protected on your device, first
Tagging in vivome happens on-device, first. The moment you record, an on-device rule-based engine classifies it instantly and the save is complete — it works fully offline, and its accuracy is checked continuously against a human-labeled test set. On-device AI can help understand richer phrasing when needed, but what's already saved is never deleted or overwritten. There is no path for an AI judgment to corrupt your original record.
Instant, on-device save
Classified and saved instantly, with no AI call needed. Recording works the same in airplane mode.
AI only refines how it reads
On-device AI helps make sense of metaphor and nuance, but your original text and saved record never change.
Insights are deterministic first
110+ pattern-analysis engines compute deterministically on-device. Only one of them uses a language model.
Accuracy is guarded by an automatic gate
Every code change is scored against a human-labeled golden set; changes that drop F1 below 82% don't pass.
In plain words
- Rule-based engine
- In plain terms — A program that finds expressions in a sentence using predefined language rules. Unlike AI, it never guesses, so the result is always the same.
- In vivome — vivome's on-device engine that classifies and saves your record the instant you write it.
- On-device
- In plain terms — Processing happens inside your phone, not on a company server (the cloud).
- In vivome — Tagging and 110+ insight computations run on the device by default — which is why they work offline.
- LLM (large language model)
- In plain terms — An AI model, like ChatGPT, that understands and generates sentences.
- In vivome — In vivome it plays a supporting role only: enriching classification and phrasing results that were already computed.
- F1 score · golden set
- In plain terms — F1 is an accuracy score that checks both misses and false tags; a golden set is an answer key labeled by humans.
- In vivome — vivome runs a continuous regression gate that blocks any code change scoring below F1 82% on the golden set.
03 · Privacy by structure
Not a promise that we won't look — a structure where we can't
Health and emotional records are among the most private data there is. Many apps appeal to trust with an operating policy — “we won't look” — but promises can break under operators, breaches, or acquisitions. vivome grounds trust in technical structure instead. The sensitive-data vault is encrypted on your device with proven open-standard cryptography (XChaCha20-Poly1305, Argon2id, BIP39) and the decryption key exists only on your device, so the server holds ciphertext only — the company cannot open it. External AI integration ships off by default (fail-closed): no call happens before explicit consent, and even after consent, sensitive references such as names are masked.
Zero-knowledge vault
Sensitive data is locked on-device with open-standard cryptography. The server stores ciphertext only; the company holds no key.
Consent gate — off by default
Sending anything to cloud AI opens only with explicit consent (fail-closed). Before consent, the default is on-device processing.
Sensitive-reference masking
Even when a consented request is sent, sensitive references such as people's names are replaced with placeholders first.
Sensitive photos stay on your device
Sensitive photo classes such as faces and body images are handled under one overriding principle: they don't leave your device.
What the company sees
- · Anonymous, aggregated statistical signals
- · (If you opt in) the minimum input needed for analysis
- · Account and usage logs needed to run the service
What the company can't see
- · Sensitive data stored in the zero-knowledge vault
- · Sensitive photo classes prioritized for on-device processing
- · Most daily records, tagged on-device and never sent
In plain words
- Zero-knowledge
- In plain terms — Data is locked on your device and only you hold the key — so even the company running the server can't know the contents.
- In vivome — vivome's sensitive-data vault works this way: the server keeps only unreadable ciphertext.
- XChaCha20-Poly1305 · Argon2id · BIP39
- In plain terms — Not homemade cryptography — three open standards long proven in secure messengers, password protection, and crypto wallets.
- In vivome — They lock the vault (XChaCha20-Poly1305), derive the key from your password (Argon2id), and back the 12-word recovery kit (BIP39).
- Fail-closed
- In plain terms — A design whose default is locked until explicitly allowed — the opposite of ‘open first, opt out later.’
- In vivome — External AI integration ships off, so before consent no cloud call happens at all.
Honest scope: we do not claim full end-to-end encryption of every record. What exists today is a zero-knowledge vault for sensitive data, on-device-first processing, a consent gate, and sensitive-reference masking.
04 · Anti-hallucination architecture
Hallucination is blocked by structure, not by prompting
In vivome, the language model's role is narrow — classifying sentences into codes, and phrasing statistics that have already been computed. Facts like dates, frequencies, and patterns are computed from actual records by deterministic on-device engines, not by the model. The same holds when answering insight questions: code computes the statistics, and the language model only turns verified numbers into sentences. Any code or number that can't be confirmed against real records is filtered out before it reaches the screen. Two defaults hold throughout: never invent values that aren't in the records, and stay silent when the sample is too small.
What doesn't exist is filtered out
Any code or number the model produces that can't be confirmed to exist is discarded on the spot.
Results are checked again
The model's suggestions reach the screen only after a separate check confirms them.
Zero invented values · minimum sample
Numbers absent from the records are never fabricated; underpowered patterns are not asserted. When there's no record, the answer says so.
Non-diagnostic pattern language
The output is observation, not diagnosis. It stops at “these rose together more than usual,” never “you have X.”
The sentence we don't write (a diagnostic verdict)
“You have an autonomic nervous disorder.”
The sentence we write (a pattern observation)
“Over the last 7 days, records of feeling cold, muscle tension, and breathing tightness rose together more than usual. Checking them alongside sleep, meal timing, and stress records will show the recurring pattern more clearly.”
In plain words
- Hallucination
- In plain terms — When AI confidently makes up facts that don't exist — dates and counts included.
- In vivome — vivome computes dates and counts by aggregating records, not by the model — removing the path to invention itself.
- Deterministic engine
- In plain terms — A computation with no guessing: the same input always produces the same output.
- In vivome — vivome's 110+ insight engines work this way; only one uses a language model.
05 · Identity-transition continuity
Link your email later — the vault never breaks
vivome lets you start recording immediately, with no signup required. Linking an email is something you do later, only when you need it — to view your records on another device, or to back them up. The catch: the sensitive photos you've already saved by then are encrypted in a way even the company can't open, and the decryption key lives only on your device, tied to the identity you had before you linked anything. The moment that earlier identity disappears, the key and your saved data risk falling out of sync. vivome is designed to carry that connection through safely, and when it hits something it can't resolve automatically, it stops and surfaces it instead of guessing.
When in doubt, it stops
When something can't be safely auto-resolved, nothing is forced through — the process stops for a person to check.
The old account isn't deleted until everything checks out
The old account is deleted only after everything has been confirmed safely moved. Even if the app crashes mid-way, there's no moment where data disappears.
The encryption key never leaves your device
The decryption key always stays on your device — it moves to your new account without ever being sent to the server.
Verified in the real production environment
Both the success path and the refusal path were actually run against the real production server, using isolated test accounts.
In plain words
- Anonymous-first onboarding
- In plain terms — Letting you use the app's features right away, without creating an email or password. An account gets linked later, only when you need one.
- In vivome — vivome lets you start recording the moment you open the app; you link an email later, if you ever want to view your data on another device.
- Device secure storage (Keychain/Keystore)
- In plain terms — A locked compartment your phone's operating system provides, which even the app itself can't casually read out. The same kind of storage password managers rely on.
- In vivome — The zero-knowledge vault's decryption key lives here — it is never transmitted to the server.
- Refuse-before-mutate design
- In plain terms — A design where, on hitting something ambiguous or risky, the system stops instead of making its own call and pushing through.
- In vivome — When two accounts' encryption keys conflict in a way that can't be safely auto-merged, the account link is refused outright.
Honest scope: not every ownership conflict is resolved automatically. In the rare case that can't be safely auto-merged (e.g. two accounts each already had their own vault under a different key), the account link is refused and flagged for a person to check — better than silently losing data.
Frequently asked questions
Q. So is this a chatbot like ChatGPT?
No. vivome's core is data structuring — the step before any chatbot. The language model's role is narrow: classifying sentences into codes and phrasing already-computed statistics. Computing statistics, detecting patterns, and comparing baselines are never done by the language model.
Q. How is this different from health apps like Apple Health?
Those apps are strong at storing metrics like steps and heart rate. vivome structures natural-language diaries, photos, and icon taps into events, and analyzes how emotion, symptoms, intake, and sleep move together over time. Storing metrics and structuring meaning are different layers of technology.
Q. When foundation models get better, won't this technology become unnecessary?
As models get stronger, the layers a model vendor absorbs (model choice, call orchestration) split from the layers it can't. A domain codebook and its semantic structure, verification standards like a golden set that defines what a correct tag is, and consented long-term record data can't be built from outside, however good the model. vivome has invested in those unabsorbable layers from the start — and works better on top of better models.
Q. Is this medical AI? Does it diagnose?
No. vivome is not a medical device and makes no diagnostic or treatment judgments. Output is limited to pattern-observation language — “these records rose together more than usual” — enforced both in the system prompt and in deterministic post-processing. Signals that need medical judgment come with guidance to consult a clinician.
“Models can be rented; structure cannot. vivome's technology lives in the structure that handles your records.”
See what vivome AI does →vivome is not a medical service and does not replace diagnosis or treatment. All analytical output is limited to observation and pattern language.