Changelog
A running record of what's been built, improved, and hardened. Updated as we build.
Implemented cryptographic tenant isolation across all storage and query layers. Each organization's documents, graph nodes, and AI sessions are scoped at the database level — no shared query access between tenants at any layer of the stack.
Shipped the core workspace experience. Users can manage project documents, interact with the extracted relational graph on the canvas, and query the AI reasoning engine simultaneously from a single interface. Canvas supports infinite scroll, node grouping, multi-select, undo/redo, and zoom/pan.
Users can now upload their firm's internal checklists, underwriting standards, and standard operating procedures directly into a project. Uploaded rules are ingested into the graph ontology and become an active compliance layer that runs against every document in the project.
Shipped the core reasoning architecture: vector embeddings stored as properties on graph nodes enable semantic similarity search, followed by multi-hop graph traversals to enforce constraints and dependencies. Every answer is grounded in source document chunks linked to graph nodes — not generated from model weights alone.
All user actions — document uploads, query submissions, entity extractions, access events, terms-of-service agreements — are now written to an immutable database-level audit log. Logs cannot be modified or deleted by users or administrators.
Every AI-generated answer now includes a full citation chain: source document, page number, exact paragraph, and visual bounding box. Answers are not generated independently — they are assembled from verified source text and presented with the evidence attached.
Shipped the backend processing pipeline for PDFs, DOCs, and scanned files. Difficult formats (scanned vs. embedded PDFs) are handled with user-guided adjustment — users define what makes a file important and what specific fields of logic should be extracted, rather than relying on generic automated parsing.
Implemented property graph architecture with shared ontology. Entities extracted from project documents (Grantors, Grantees, legal descriptions, instrument numbers, signature blocks) are mapped to graph nodes with defined relationships. This enables the engine to detect mismatches mathematically rather than through keyword matching.
TitleTrace is currently in private beta. The changelog reflects work completed during the build phase. Public release entries will follow.