Research and source retrieval assistant
Compare a question with source excerpts and return inspectable matches. Future: grounded retrieval, citations, contradictory evidence and version tracking.
Annotated project explanation library · Business automation · Email drafting · Cybersecurity and backups
SOFTWARE ENGINEERING · AI PRODUCT RESEARCH · APPLIED PORTFOLIO
A programme for building specialised assistants that support everyday life, education, research and enterprise. The long-term architecture is a private, permission-based personal AI workspace with distinct agents and a human-reviewable action history. This public site shows selected prototypes and documented designs, not access to anyone's private accounts.
Demonstration status: The interactive examples below execute deterministic JavaScript in your browser using sample or locally entered text. They are functional prototypes of workflows, not trained language models, connected mailboxes, unattended scheduling services or private document vaults. Entries are not sent to a server by these demos. Do not enter confidential records, identity details or passwords.
The Societal Resilience & AI Public Adviser connects the research assistant programme with geopolitical evidence, ethics of political advisers, economic adaptation, community resilience and enterprise questions. Its first interaction captures and compares two cited news-source descriptions locally; it does not retrieve live articles, run a trained AI model, verify truth or predict outcomes.
Every tool in the 63-tool register now has an explicit risk field. The 20 documented cases cover data exposure, software supply chains, third-party compromises and governance failures. These are historical incident lessons, not current security ratings of each tool.
Before connecting new provider accounts, compare 63 shortlisted AI tools and services for local use, free cloud quotas, source extraction, data management, evaluation and security. The register records account requirements, cost limitations and required tests; no new subscriptions or keys were established by publication.
Explore Salesforce, CRM Analytics, Einstein Discovery and 23 integration options alongside open-source automation and analytics infrastructure. The published register is an assessment, not a connected Salesforce tenant.
Compare a question with source excerpts and return inspectable matches. Future: grounded retrieval, citations, contradictory evidence and version tracking.
Test an approval-aware intake workflow. Future: accounts, document processing, permissions, audited integrations and reversible actions.
Future: a conversational model grounded in approved sources, with uncertainty reporting and a choice of response styles.
Check understanding of AI concepts with explanations and hints. Future: personalised learning sequences with teacher and learner controls.
Generate a year-long publication queue for review. Future: licensed media ingest, channel-specific approval and official scheduling APIs.
Generate one or two editable draft variants. Future: verified account access, contextual retrieval, recipient policy and approval gates.
Future: encrypted records, permissioned search across papers and files, reminders, consent and secure deletion workflows.
Explore useful recurring services, research tooling and responsible resource-management topics, including water systems.
Search a small, explicitly synthetic research library. Matching is lexical; a match does not verify the source or establish a factual conclusion.
Simulate intake triage with explicit rules. It creates a review recommendation; it cannot trigger an external action.
A short formative check, not a qualification or diagnostic assessment.
Question: What makes a source-grounded AI answer more trustworthy?
Preview how a collection of approved content could be staggered over 12 months. This demonstration does not post, connect to social accounts or generate copyrighted media.
| Layer | Responsibilities | Safeguard |
|---|---|---|
| Public portfolio | Static demonstrations, source documentation, evaluation cases and capability summaries | No personal records or credentials |
| Private identity and vault | Authenticated user, encrypted file storage, metadata, search index and contextual retrieval | Access scopes, audit events, encryption and verified deletion policies |
| Agent orchestration | Task interpretation, retrieval, planning, tool calls and results | Allowlisted tools, minimal permissions, injection defences and cost budgets |
| Actions and approvals | Mail drafts, calendar proposals, content schedules, forms and document workflows | Approval before consequential sends, purchases, legal submissions or account changes |
| Observability | Source citations, change history, test results, error handling and rollback paths | Reversal where possible; sent messages and outside-world effects cannot always be undone |
Research programme · Existing AI research laboratory · Software engineering · Implementation roadmap · Public source repository
A secure private vault would ingest user-approved files, index passages with source and page metadata, and answer only from documents the user is authorised to access. Reminders could flag deadlines or expiries, but a new passport or official appointment would require an eligible, connected service and explicit confirmation. An action dashboard should show proposed, approved, executed and failed tasks with undo options for reversible operations.
Potential offerings include in-person demonstrations, AI literacy sessions, workflow assessments, document search prototypes, research-method support and maintenance subscriptions. Demonstrate value through defined outcomes, transparent prices, maintenance costs and customer consent before selling ongoing access. Research into essential resources, including water treatment and desalination, can be a separate evidence-led sustainability track rather than an assumption that scarcity is deliberately manufactured.
Commercial boundary: portfolio demos and unvalidated ideas are not representations of production-ready AI, guaranteed income or public deployment of private systems. Borrowing to fund speculative automation is not a prerequisite; validate demand and costs first.
Each implementation should publish a problem statement, limited use case, data permissions, model or rules used, reproducible sample tests, accuracy and failure rates, security threats, latency, cost per task, human review requirement, accessible demo and a dated version history. Keep private evaluation datasets and third-party records outside the public repository.
Next engineering sequence: source-grounded research assistant → private document retrieval → approval-aware workflow agent → education assistant → content queue → messaging integrations. A model provider or locally hosted model can later replace deterministic components; the model must be evaluated before any consequential automation is enabled.
Public research and prototype record · 9 November 2026. Current prototypes are browser-only and do not involve model inference.