Manifold Grace

ENGINEERING & PRODUCT DELIVERY · 9 NOVEMBER 2026

Building a personal AI system: architecture and delivery roadmap

Try the interactive public prototypes · Existing AI research laboratory · Research

The intended product is a modular assistant environment rather than one unrestricted system. Separate capabilities should share a private retrieval layer, a permission broker and an action history. Each component must be independently testable and replaceable.

1. Component architecture

ServiceInputs and outputsImplementation requirements
Identity and consentUser sign-in, scoped permissions, connected service tokensAuthentication, least privilege, credential rotation, encrypted secrets, revocation
Private file vaultUploaded PDFs, notes, papers, metadata, retention preferencesEncrypted file storage, access enforcement, file type controls, deletion policy
Search and retrievalPassage indexing, document IDs, source pages, citationsHybrid lexical/vector retrieval, citation coverage tests, source-aware versioning
Model gatewayQuestions, approved context, structured proposed outputsServer-side API keys, provider abstraction, rate limits, budget caps, model evaluations
Agent workflow engineExplicit task plans and allowed operationsAllowlisted tools, tool-level permissions, injection checks, isolated execution
Action ledgerProposal, approval, execution status, external receipt, rollback attemptAuditable events, idempotency, error recovery, clear irreversible-action notices
ConnectorsMail, calendar, content platforms, business systemsOfficial APIs, per-action scopes, usage limits, platform terms, consent
MonitoringErrors, effectiveness, security incidents, evaluationsMinimised logs, anomaly alerting, security review and routine maintenance

2. Phased delivery

StageDemonstrable outcomeExit criteria
0 — Public interactionBrowser-only demos: evidence matching, triage, tutoring, content queue and repliesAccessible interface; no secret credentials; clear prototype labelling
1 — Grounded research assistantPrivate document upload and question answering with quoted passages and citationsCitation tests, permission isolation, deletion tests, measured error cases
2 — Business process pilotHuman-approved intake and draft workflows with simulated and then limited real integrationsNo unexpected external actions; rollback and failure state verified
3 — Learning and communicationsPersonalised explanations and editable one- or two-option message draftsContent quality tests, author approval, contextual correctness
4 — Licensed content schedulingMedia library, queue, preview, approval and selected platform APIsPublishing permissions, duplicate checks, platform compliance, stop/pause
5 — Proactive personal assistantReminders and suggested actions from user-approved recordsOpt-in triggers, clear reasons and sources, approvals for purchases, bookings or submissions

3. Action permissions and reversibility

Classify operations as read-only, reversible local changes, consequential external changes or legally/financially consequential actions. Let an agent automatically gather permitted information and prepare a plan, but require explicit approval before sending correspondence, publishing to third-party accounts, paying, deleting records, booking paid appointments or filing official forms. Keep a visible undo button only for actions actually reversible; an email that was delivered cannot be reliably recalled.

4. Evaluation and scientific value

Each specialist assistant should have reproducible evaluation data: source-grounding precision, coverage, task completion, unsupported assertion rate, action safety, cost per successful task, latency, accessibility and privacy tests. Compare against a non-AI baseline and record cases where a human is more reliable. Publish sanitised benchmarks and research methods, not private personal data.

5. Service and commercial models

Candidate services: in-person AI configuration demonstrations; fixed-scope research retrieval setups; private workflow analysis; organisational training; maintenance agreements where there is an ongoing, measurable service. Test willingness to pay before adding recurring infrastructure costs. A prototype is not a commercial guarantee, and formal commitments should reflect actual delivery capability.

6. What is not live yet

There is currently no connected language-model backend, account-based sign-in, encrypted private vault, autonomous mail responder, official content scheduling integration or booking/purchase service attached to this demonstration. GitHub Pages publishes static content and cannot run server-side application code by itself. A separate protected backend and operating budget are required.

7. Security and standards references

GitHub Pages hosting limitations · OWASP risks for LLM applications · NIST AI Risk Management Framework · ICO AI and data protection resources

Development architecture and phased plan · Not a claim of deployed private services or production readiness.