AI Assistant
An AI-driven assistant that keeps personal relationships, finances, and work in three focused modes, then brings in the right specialist perspective—from software architecture and accounting to cooking and creative work—without mixing contexts.
- Status
- Product concept
- Role
- Product & AI architecture
- Focus
- Application
Product capabilities
What it helps people do
- Personal Relations mode for thoughtful communication, important moments, shared plans, and relationship context
- Finance mode for budgeting, expense understanding, financial planning, and decision support
- Work mode for planning, problem-solving, documentation, prioritization, and professional decisions
- Current specialist roles: Software Architect, Chartered Accountant, Chef, and Artist
- Intent-aware routing that selects the right mode, specialist, tools, and knowledge source for each request
- Separate memory and permissions for personal, financial, and work contexts
- Human-in-the-loop confirmations before sensitive recommendations or external actions
- Expandable specialist marketplace for adding new professional roles over time
Three assistance modes
One interface. Clear context boundaries.
People choose the context before they begin, making the assistant’s memory, tools, permissions, and responsibilities easier to understand.
Personal Relations
Helps prepare thoughtful messages, remember meaningful moments, plan shared activities, and reflect on conversations without pretending to replace human judgement.
Finance
Turns transactions, budgets, goals, and financial documents into understandable summaries, options, reminders, and questions to discuss with a qualified professional.
Work
Supports planning, research, architecture decisions, documentation, meeting preparation, prioritisation, and follow-through while keeping company context separate.
Current specialists
Focused expertise for the task at hand.
Each role has its own knowledge, tools, output format, and safety rules instead of relying on one generic system prompt.
Software Architect
Design reviews, trade-off analysis, system decomposition, API and data modelling, reliability, security, and technical decision records.
Chartered Accountant
Financial categorisation, cash-flow interpretation, document checklists, compliance-oriented questions, and explainable calculations with clear professional disclaimers.
Chef
Meal ideas based on ingredients, dietary preferences, time and budget, with practical substitutions, preparation plans, and shopping lists.
Artist
Creative briefs, concept exploration, style references, critique, iteration prompts, and project planning across visual and expressive work.
System architecture
Context-aware. Tool-enabled. Safe by design.
A deterministic orchestration layer surrounds the language model, controls context and tools, and records enough evidence to evaluate every specialist.
Conversation gateway
A React or mobile client sends a request with an explicitly selected mode. FastAPI handles identity, sessions, streaming responses, rate limits, and request validation.
Mode and intent router
A small classifier and LangGraph workflow determine the permitted context, specialist role, retrieval sources, tools, and response policy for the request.
Specialist agent registry
Each specialist is a versioned configuration containing instructions, allowed tools, knowledge collections, output schemas, confidence rules, and safety boundaries.
Separated memory
Personal, finance, and work memories use distinct namespaces, encryption keys, retention controls, and access policies so information cannot leak across modes.
RAG and tool layer
Vector retrieval grounds answers in user-approved documents. Tool calling handles calculations, calendars, tasks, recipes, code analysis, and other deterministic actions.
Trust and evaluation
Citations, uncertainty, confirmation gates, audit logs, prompt and retrieval tests, specialist scorecards, and user feedback make behaviour measurable and reviewable.
Skills required
AI engineering grounded in product trust.
AI product design
Define jobs-to-be-done, decide when AI should assist or abstain, design confirmation moments, and make uncertainty understandable.
LLM orchestration
Prompt and context engineering, structured outputs, LangGraph state machines, tool calling, retries, fallbacks, and model routing.
RAG and data
Document ingestion, chunking, embeddings, metadata filters, hybrid retrieval, reranking, citations, and evaluation datasets.
Security and privacy
Tenant isolation, consent, encryption, secrets, fine-grained access, data deletion, auditability, and protection against prompt injection.
Backend engineering
FastAPI services, asynchronous jobs, event processing, PostgreSQL, vector storage, caching, observability, and reliable integrations.
AI evaluation
Golden test cases, groundedness and tool-use checks, hallucination tracking, safety tests, cost and latency measurement, and feedback analysis.
Delivery approach
Start narrow. Earn trust. Expand carefully.
- 01
MVP: launch Work mode with the Software Architect specialist, explicit mode selection, short-lived sessions, structured answers, and no autonomous external actions.
- 02
Grounding: add user-approved document ingestion, source citations, separated memory, retrieval evaluation, feedback capture, and privacy controls.
- 03
Specialists: introduce Chef and Artist with narrow tools and test suites; add Finance only after calculation checks, disclaimers, audit logs, and expert-reviewed boundaries are ready.
- 04
Personalisation: add opt-in long-term preferences, user-editable memories, specialist hand-offs, proactive suggestions, and configurable notification rules.
- 05
Scale: route requests across models by complexity, cache safe retrieval results, move long tasks to queues, measure cost per successful outcome, and add specialists from a governed registry.
Product guardrails
Useful assistance without invisible overreach.
- Personal, financial, and workplace information must never cross mode boundaries unless the user explicitly approves a specific hand-off.
- Finance guidance must show assumptions, calculations, sources, uncertainty, and when a qualified professional is required.
- The assistant should request confirmation before sending messages, changing calendars, moving money, publishing content, or performing any consequential action.
- Long-term memory must be opt-in, visible, editable, exportable, and deletable—not an invisible transcript archive.
- Specialist quality needs domain-specific evaluation; a fluent response is not evidence that the recommendation is correct.
REQUEST A DEMO
Want to see AI Assistant in context?
Share what you would like to explore.
Sign in to like or comment.