Ask your claims. Not your calendar.
Claimsmith is an LLM-backed assistant that reads live claim data, drafts diary entries, and cites its sources - with a mandatory human-approval gate before anything writes back. Built as a per-customer sidecar API, not a shared service.
Most "AI for claims" is a chatbot bolted to a knowledge base. Claimsmith is different: it runs against your actual claim schema - reserves, payments, diary entries, notepad, documents - via read-only SQL tools with per-tenant logins. Every answer cites the specific rows it read. Every drafted diary entry lands in a review queue before anything mutates.
The examiner asks "when was the last IME on claim 04829 and who did it," and Claimsmith answers with the specific rows it read. When the examiner asks it to draft a diary entry, the draft appears with an approve/edit/discard control. Nothing goes to the database without a human clicking approve.
Capabilities
-
Natural-language queries
Over reserves, payments, diary entries, claim metadata, and - with Phase 1.5 - scanned documents via semantic search.
-
Human-approval gate
Drafts diary entries but never writes without an explicit reviewer approval. No silent mutations.
-
Cited answers
Every response cites the specific claim rows and document sections it read. Traceable, not vibes.
-
Per-tenant isolation
Each customer gets their own SQL login (read-only), Qdrant namespace, and API instance. No shared context.
-
Audit the auditor
Every conversation is logged. Compliance officers can review who asked what, with per-officer read logging.
-
Semantic document search
Qdrant embeddings over Notepad and scanned documents so relevant history surfaces even without exact keyword matches.
Read-only tools. Human-approval gate. Auditable by design.
Twenty-six SQL-backed tools are exposed to the model: 24 are preload-only (context-hydrating, never LLM-callable directly), 2 are LLM-callable. The model can read anything the tenant's SQL login can read. It can write nothing. Every diary draft is a proposal, not a commit.
The examiner desk, faster and calmer.
Not a general-purpose chatbot. Claimsmith knows your schema, your reserves, your diary conventions - because it reads them directly. Answers reconcile with the reports your team already runs.
Faster case reviews.
"Summarize the last 90 days of activity on claim 04829." Answers with cited rows, and drafts the next diary entry for review.
Never re-read a claim from scratch.
Load a claim, ask what's happened since the last diary, and get a chronological summary - grounded in the actual data, not a fine-tune's memory.
See every question asked.
Full conversation audit trail, with per-officer read logging so the audit itself is auditable. Prove what the model was asked and what it was told.