Applied AI · Company operations / Live beta
An AI assistant that understands the company around the work.
Company-aware AI with permission-scoped knowledge, department assistants and reviewed actions.
The problem
Company knowledge and work are spread across departments, documents and connected tools. Useful AI needs that context, but it must also respect who is allowed to see it and who can authorize an action.
What I owned
The company workspace, role and department model, knowledge ingestion, assistant experience and integration/action boundaries. I worked across Angular interfaces, PostgreSQL policies, retrieval and server-side orchestration.
The difficult constraint
Retrieval relevance is only one part of correctness. A relevant document must still be authorized, a generated item must cite permitted evidence, and a prepared action must remain subject to current policy.
The architecture
Private documents → validated ingestion and chunks → semantic embeddings and pgvector → permission-filtered evidence → company-selected model → citation-backed response. Persistent context supports follow-up work. Exact approval and execution receipts sit outside the model.
Why these decisions
Enforce tenant isolation and knowledge permissions in the data/runtime boundary. Validate structured model output before it becomes work. Keep capability policy, approval and audit explicit so a fluent answer cannot silently become an external action.
What is live
The public beta and adaptive intelligence checkpoints are deployed. Repository release notes document API, Angular and SQL boundary coverage. Some signed-in acceptance, controlled model-quality evaluation and real provider write scenarios remain open; an available adapter does not imply a connected integration.
What this taught me
The most valuable AI work is often ordinary systems engineering: retaining provenance, keeping authorization current, managing failure states and making the next action reviewable.
Evidence & scope +
- Tasko package manifest: Angular 22, TypeScript 6, Node 22, Supabase 2.115 and Vitest 4.
- Implementation status, September 2026: deployed adaptive intelligence, 28 local SQL suites and documented live acceptance gaps.
- Assistant, integration and database source inspected; screenshot is an anonymous public product view.