Model stack
| Surface | Model | Why |
|---|
| Multi-turn coaching (paid tier) | Claude Sonnet 4.6 | Long-context reasoning, tool use, structured outputs. |
| High-volume conversations + nudges (free + paid) | Claude Haiku 4.5 | Fast, low-cost, good enough for short turns and contextual messages. |
| Vision (document OCR, pill bottle reading, win extraction) | Claude Sonnet 4.6 (vision) | Single model handles image + reasoning without a second model hop. |
Multi-model by design. Claude (Sonnet
4.6 + Haiku 4.5) runs on the Anthropic API direct for user-facing
coaching. Gemini 2.5 Flash on Vertex AI handles
secondary classification — for example, sorting user inputs into the
six 42 Map life domains (career, relationships, finances, health,
personal growth, purpose) to feed domain-health analytics. Patet,
our AI money coach, is the first product to ship — live in beta at
patet.plainlydigital.com.
Path forward: as Plainly Digital scales onto Google Cloud, Claude
workloads will consolidate onto
Claude on Vertex AI Model Garden — bringing
observability (Cloud Logging + Cloud Trace), governance (IAM +
Cloud Audit Logs), and billing under a single GCP control plane.
What AI does in each product
Patet pairs 121 financial literacy lessons with Glyphe,
an AI coach that reads the user's real Plaid transaction stream
and uploaded bank statements. Glyphe runs on Claude Sonnet 4.6
for paid-tier conversations (Connected, Coached) and Haiku 4.5
for free-tier and lesson-completion nudges.
AI tasks:
-
Multi-turn coaching that references specific transactions
("you spent $112 on DoorDash this week, up from $40")
- Statement upload — Claude vision + reasoning extracts transactions from PDF/CSV/Excel
- Plaid-driven lesson assignment — model picks the lesson that matches the spending pattern
- Money Personality, Money Roast, Future You Visualizer, Money Wrapped weekly recap
- Paycheck detection + payday auto-coach briefing on every deposit
- Crisis routing (988 / Crisis Text Line) with Unicode-normalized text matching
Why this isn't possible without AI: The product thesis is "money lessons tied to your actual spending." Rule-based systems can match categories but can't generate the empathetic, lesson-aware coaching responses that connect a spending pattern to a curriculum module. The certification credential is also AI-graded — passing the 50-question assessment requires more than pattern matching.
Glyphe / AI Life Advisor — in development
Personality-calibrated AI advisor. Twelve-question persona calibration
builds a persistent context that the model uses across every future
conversation. Year-in-review story format. Tiered limits never gate
crisis routing. Glyphe is the reference architecture for the
portfolio's Google Cloud build-out (Cloud Run + Firebase Hosting).
AI tasks: personality calibration, long-context advisory across life domains, persistent memory, year-in-review story generation. Multi-model design: Claude Sonnet 4.6 (Anthropic API) for the user-facing coaching response + Gemini 2.5 Flash (Vertex AI Model Garden) for post-response topic classification feeding domain-health analytics.
Why this isn't possible without AI: the product is the AI advisor. Remove the model and there is no product.
Vinla — AI health experiments
N-of-1 health experiments with statistical rigor (Welch's t-test +
effect size) wrapped in an AI coach that interprets the results and
suggests next experiments. Persistent client-side AI memory across
sessions. Food vision via Claude.
AI tasks: experiment interpretation, food vision + nutrient extraction, persistent personalization across sessions, crisis routing.
Why this isn't possible without AI: N-of-1 statistics is the easy part. Translating "your blood pressure dropped 4 points when you cut afternoon coffee" into actionable next steps is the hard part — and that's the AI's job.
Winlet — AI win extraction
AI extracts user accomplishments from screenshots (Slack DMs, email,
calendar events) into a Wrapped-style recap. Hype Circles add social
celebration.
AI tasks: vision-based win extraction from arbitrary screenshots, win remixing (tweet / meme / affirmation tone), monthly + yearly Wrapped recap generation.
Why this isn't possible without AI: "screenshot to win" is a vision + reasoning task. Manual entry exists in every productivity app; AI extraction is the differentiator.
ClearDoc — AI document explainer
Point a camera at any confusing document — a bill, a contract, a
benefits letter — and get a plain-English summary, action items, and
key numbers in seconds. Sensitive identifiers (SSN, card numbers)
redacted on-device before any AI call.
AI tasks: document vision + structured extraction, plain-English summarization, action-item + key-date extraction.
Why this isn't possible without AI: document explanation across arbitrary domains (medical, legal, financial) requires multi-domain language understanding. Hard-coded parsers don't generalize.
SitterSheet — AI sitter-guide generator
An AI-guided wizard turns the user's answers (pet routines, house
info, sitter-specific quirks) into a warm, comprehensive sitter
guide. Share with a 30-day expiring link or export to PDF.
AI tasks: guide generation tuned for warmth + clarity, content adaptation per-sitter (full-context vs. quick-reference).
Why this isn't possible without AI: the product is the generation step. A static form would be a different (worse) product.
Safety + governance architecture
- Input guardrails: prompt injection detection,
role-override detection, safety-bypass detection, jailbreak
detection. Per-message Unicode + homoglyph + l33tspeak
normalization.
- Output guardrails: medical / legal / financial
boundary checks. Every response carries an AI disclosure badge.
- Crisis detection: always-on text classification
routes to 988 (suicide + crisis lifeline), Crisis Text Line, or
211.org regardless of subscription tier. Crisis routing is
never gated.
- Audit logging: NIST AI Risk Management Framework
aligned. Every AI call carries an audit record tied to a
per-user identifier (not the underlying PII) for incident
response.
- Cost guard: per-user daily and monthly call
caps, per-tier model selection (Sonnet vs. Haiku), hard
kill-switch on daily AI spend. Prevents runaway cost on a
single bad prompt loop.
- No training on user data. All AI calls go through
Anthropic's zero-retention API path (request not stored beyond
the immediate response). When we migrate to Vertex AI Model
Garden, the equivalent zero-retention contract applies.
- Financial PII at rest: dual-write to AES-256-GCM
encrypted companion columns. Plaid access tokens encrypted with
the same primitive. Backfill scripts maintain encryption parity
across migrations.
Migration path: Anthropic API → Vertex AI Model Garden
Plainly Digital is mid-flight on a Google Cloud consolidation. Today,
compute lives on Render + Vercel; AI calls hit Anthropic's API
directly. The target state:
- API services on Cloud Run (autoscale, min-instances=1 for no cold start)
- Static frontends on Firebase Hosting (already live for plainlydigital.com itself)
- Database on Cloud SQL Postgres (Patet, Vinla; Neon stays for Winlet)
- Auth on Firebase Auth (Vinla; Patet retains self-issued JWTs)
- Secrets in Secret Manager
- CI/CD via Cloud Build (template proven on the Glyphe build and on plainlydigital.com itself)
- AI calls via Claude on Vertex AI Model Garden for unified observability + audit + billing
Patet, our AI money coach, is live in beta and ships first. The rest of
the portfolio follows onto Cloud Run + Firebase Hosting through 2026,
with Claude workloads consolidating onto Vertex AI Model Garden as the
build-out completes.
Contact
AI architecture questions, partnership inquiries, or program
eligibility review: apps@plainlydigital.com.
