90%+ compliance accuracy. Every answer cited to a specific page and paragraph. And when the answer isn't in the documents, the AI says so instead of guessing. TechEniac built a production RAG system that mortgage professionals trust with regulatory decisions.
With 10 years of mortgage lending experience across origination, underwriting, compliance, and portfolio management, the founder had seen the same challenge firsthand: loan professionals spending 20–30 minutes searching complex guidelines for simple answers.
He wanted to change that with an AI system that delivers accurate, source-backed answers with precise citations without ever guessing.
Mortgage lending runs on thousands of pages of guidelines that are frequently updated, cross-referenced, and complex to navigate. The founder identified three compounding problems making manual compliance checks unsustainable:
Answering a single compliance question often requires cross-referencing multiple mortgage guidelines, waiting periods, and investor overlays. Loan officers can spend 20–30 minutes per lookup, creating hundreds of hours of repetitive work at scale.
Fannie Mae, FHA, VA, and other agencies regularly update their requirements. Teams must continuously review changes, update internal references, and retrain staff making it difficult to keep everyone aligned.
Generic AI tools can provide confident but inaccurate answers. In mortgage lending, a wrong answer can create serious compliance and financial risk. The system needed to ground every response in verified source documents with clear citations.
Every answer must cite its source document name, section number, page, and paragraph enabling instant verification against the original guideline
When the information doesn't exist in the guidelines, the system refuses to answer rather than fabricating a plausible-sounding response
Handle all guideline formats digital PDFs, scanned PDFs, Word documents, HTML, spreadsheets without losing critical structure (tables, nested lists, hierarchical sections)
Process monthly guideline updates automatically, ensuring current guidance takes priority while maintaining audit trail of historical versions
Achieve 90%+ accuracy on real-world mortgage compliance questions verified against certified underwriter answers
Digital and scanned PDFs are processed while preserving tables, headings, lists, section structures, and page-level references. Structured extraction captures the document’s hierarchy and cross-references, while version tracking ensures updates are mapped, affected content is flagged, and the latest guidelines always take priority.
Content is chunked by document structure rather than token limits, keeping complete rules, tables, and lists intact. Each chunk retains rich metadata including document version, section, page, and hierarchy enabling precise citations and targeted retrieval across specific guidelines or update periods.
The system combines semantic and keyword-based retrieval to understand both natural-language questions and precise guideline references. Qdrant captures contextual meaning, while BM25 identifies exact terms and sections. Reciprocal Rank Fusion intelligently combines both signals, adapting retrieval based on the type of query for more relevant and reliable results.
Gemini generates answers with inline citations for every factual claim, backed by the retrieved source documents. A validation layer verifies each citation before delivery, rejecting and regenerating responses when sources cannot be confirmed. Citations include the document, section, page, and paragraph, making every answer easy to verify.
The system is designed to refuse rather than guess when reliable guidance isn’t available. A dual validation layer checks retrieval relevance and topic alignment before generating an answer. Around 8% of queries are intentionally declined, ensuring uncertain or out-of-scope questions never become confident but potentially misleading answers.
Table Extraction Accuracy from Complex PDFs
Implemented Google Document AI's table extraction with post-processing validation layer. Extracted tables validated against expected column counts and data types (percentages, dollar amounts, time periods). Tables failing validation flagged for manual review rather than ingested with incorrect structure.
Cross-Reference Resolution Across Guideline Sections
Built a cross-reference resolution layer that identifies section references within chunks, retrieves the referenced section, and appends it as supplementary context during generation. The AI sees both the original chunk and the referenced section.
Version Management During Monthly Guideline Updates
Implemented version-aware retrieval system. Every chunk carries version timestamp and document revision identifier. Retriever defaults to most current version but supports explicit historical queries ("What was DTI limit before January 2026 update?"). When new version ingested, superseded chunks flagged (not deleted), preserving audit trail while ensuring current guidance takes priority.
Compliance Accuracy
Measured against 150 real mortgage compliance questions with verified answers from certified underwriters—system's accuracy validated against human expert baseline.
Citation Rate
Every answer includes document name, section, page, and paragraph architecturally enforced, not optional. Loan officers verify answers in seconds against original guidelines instead of spending 20–30 minutes searching.
Confident Refusal Rate
Questions not covered in guidelines receive explicit decline, not fabricated answers. This refusal rate builds trust lenders know when the system says yes, it's grounded in actual guidance.
Table Extraction Accuracy
On first pass, with validation layer catching remaining 6% before ingestion critical for complex guideline matrices and rate grids.
Answer Completeness Improvement
From cross-reference resolution multi-section rules answered completely instead of partially. Complex guideline dependencies now handled correctly.
Reduction in Repeat Query Costs
Semantic caching catches equivalent questions asked in different words same query in different phrasing serves cached response instead of regenerating.
AI / ML
Retrieval
Document Processing
Backend
Frontend
Cloud & DevOps
FinTech / InsurTech- Insurance Claims Processing
ClaimBot An AI-powered insurance claims chatbot that automates FNOL intake across web and voice channels. ClaimBot reduced claims processing time by 78%, integrated with Guidewire and Duck Creek, and generated FCA-compliant audit records for every interaction.
Edtech
Managing multiple brand accounts with a content team that can't keep up? That's the exact problem ContentForge AI was built to solve. Let's talk about yours.
