ShowcaseRAG & Automation
Audit Genie
Audit Genie is VisionsCraft's AI assistant for accountants and auditors: a React SPA for PBC checklists and workpapers, FastAPI orchestration, Claude-primary multi-LLM routing, ChromaDB hybrid RAG, and a strict source-grounding filter so every answer cites engagement evidence—or it does not ship.
The problem
Engagement teams lose days chasing prepared-by-client (PBC) items, hunting workpapers for adjustments, and drafting rationales without a reliable evidence trail. Generic chatbots invent citations. Spreadsheets cannot show partners what is overdue, grounded, or unverified.
The solution
An end-to-end Client Auditor stack: PBC gather and refine agents, document ingest with embeddings, bulk coverage/mismatch scanning, remediation options, chat-with-timeline, and report generation—all behind JWT security, async audit logs, and a grounding filter that rejects unanchored claims.
Architecture spans six layers: (1) React/TypeScript client with PBC checklist, timeline, remediation matrix, and audit chat; (2) FastAPI gateway with JWT, CORS, async audit-log middleware, and global exception handling; (3) routers for auth, engagements, documents, workpapers, agent orchestration, chat/timeline, and Hermes Playwright scraping; (4) specialized agents—PBC gather, population, procedure, review, bulk analysis, and remediation—with a strict source-grounding filter and multi-tier fallback loop; (5) RAG engine with document parsing, embeddings, hybrid retrieval, library cache, and ChromaDB; (6) PostgreSQL, local file storage, and AIRouter multi-LLM (Anthropic Claude primary, local Ollama/RunPod GPU fallback). Flow: PBC gather retrieves library-only standards, asks Claude for structured JSON, grounds every line item, and rejects unverified rows into a fallback loop before the UI renders an interactive checklist. Document upload embeds into ChromaDB and can auto-trigger bulk forensic integrity analysis. Chat refine streams partner-ready plain-English explanations with standards retrieval. Report generate drafts workpapers from retrieved evidence—not from memory.
Implementation
SPA for PBC, workpapers, and chat. FastAPI orchestrator with JWT and async AuditLog middleware. Agent suite for PBC gather/refine, bulk analysis, remediation, and report drafting. Hybrid retriever over ChromaDB with library-grounded mode. AIRouter with Claude primary and Ollama/RunPod fallback. Hermes Playwright scraper for portal pulls. Engagement and document status tracked in PostgreSQL with grounded vs unverified flags on every AI output.
Tech stack
- React
- TypeScript
- FastAPI
- Claude
- ChromaDB
- PostgreSQL
- Playwright
Results
- Interactive PBC checklists grounded in standards libraries—not freeform lists
- Document Q&A and bulk integrity checks tied to uploaded workpaper evidence
- Unverified AI claims rejected by a strict source-grounding filter
- Partner-ready rationales streamed with standards retrieval
- Full request audit trail for every orchestrated agent call
FAQ
What is PBC in Audit Genie?
Prepared-by-client checklists. The PBC agent gathers industry library requirements, structures them as JSON, grounds each item against the library, and only then renders the interactive checklist UI.
How do you stop hallucinated audit answers?
A strict source-grounding filter keeps grounded items and rejects the rest into a fallback loop. Chat and reports retrieve evidence from ChromaDB before the LLM drafts language.
Which models power Audit Genie?
Anthropic Claude is the primary LLM via an AIRouter. On error, the stack can fall back to local Ollama or RunPod GPU inference so engagements are not blocked by a single provider outage.
Who is the product for?
Accounting and audit teams that need engagement discipline—PBC tracking, workpaper RAG, remediation options, and citeable AI—for Client Auditor workflows.