Written by Talha Asif
CTO & AI Engineer
An AI assistant for auditors helps engagement teams gather prepared-by-client (PBC) checklists, retrieve workpapers, scan for mismatches, and draft rationales—only from engagement evidence. VisionsCraft built Audit Genie with a strict source-grounding filter so unverified claims do not reach the UI.
Why generic chat is unsafe on an engagement
Partners cannot take a citation they cannot find in the file. Spreadsheets do not show what is overdue, grounded, or invented. A useful assistant must refuse to ship a line item that is not anchored in the standards library or uploaded workpapers.
PBC gather as a grounded workflow
The PBC agent retrieves library-only standards, asks Claude for structured JSON, grounds every line item, and rejects unverified rows into a fallback loop before the React checklist renders. That is slower than a chatbot list—and that is the point.
Six-layer stack (Audit Genie)
- React/TypeScript SPA: PBC, timeline, remediation matrix, audit chat
- FastAPI gateway: JWT, CORS, async audit-log middleware
- Routers: auth, engagements, documents, workpapers, agents, chat, Playwright pulls
- Agents: PBC gather/population/procedure/review, bulk analysis, remediation—plus the grounding filter
- RAG: parse, embed, hybrid retrieve, library cache, ChromaDB
- Data and models: PostgreSQL, local files, AIRouter with Claude primary and Ollama/RunPod fallback
Case study: Audit Genie. Related services: RAG and knowledge systems.
Chat and reports
Document upload embeds into ChromaDB and can trigger bulk integrity analysis. Chat refine streams partner-ready explanations with standards retrieval. Report generate drafts workpapers from retrieved evidence—not from model memory.
Honest limits
This does not replace partner judgment or sign-off. It reduces hunting and unanchored drafting. We do not publish fake hours-saved metrics for unnamed firms.
FAQ
What is an AI assistant for auditors?
Software that helps engagement teams gather PBC checklists, retrieve workpapers, scan for mismatches, and draft rationales—only from engagement evidence, not freeform chat.
How does Audit Genie stop hallucinated citations?
A strict source-grounding filter keeps items tied to the standards library or uploaded workpapers and rejects unverified rows into a fallback loop before the UI shows them.
Which models power Audit Genie?
Anthropic Claude is primary via an AIRouter. On provider error the stack can fall back to local Ollama or RunPod GPU so an engagement is not blocked by one outage.
What is PBC in this product?
Prepared-by-client checklists. A gather agent structures library requirements as JSON, grounds each line, then renders an interactive checklist partners can track.
See the related build: Audit Genie. Explore RAG and knowledge systems or book a consultation with VisionsCraft.