Profile risk analysis
DeepSeek-powered risk summaries score borrowers 0–100 with grade (A–D), risk level, factor breakdown, and actionable recommendations cached for 24 hours.
Platform package 03
AI Review Layer
AI risk summaries, Copilot with eight database tools, email and document extraction, inbound lead review, and prompt governance for staff.
Overview
AI that supports lender review — grounded in your borrower data.
The AI Review Layer sits inside the Management Portal and augments staff workflows with structured summaries, document extraction, and a conversational Copilot that can query the database on demand. It is designed to accelerate review work — not replace lender decisions. Every AI output is tied to real borrower records, and prompt behavior can be aligned with your internal credit practice.
On each profile, AI Risk Analysis generates a scored summary with grade, risk factors, and recommendations. In the assistant, staff ask natural-language questions and the model invokes tools to look up borrowers, rank risk, list overdue notices, and — for admins — propose and apply governed edits with a two-step confirmation flow.
Email and document intake
Inbound emails and attachments are parsed into structured lead records. Low-confidence extractions route to a pending-review queue for staff correction before approval.
Prompt and edit controls
Admin prompt configuration aligns summary format and emphasis. Copilot edits require explicit user confirmation and are audit-logged with before/after previews.
Capabilities
What the AI Review Layer delivers to staff.
Summaries, extraction, and conversational access — all scoped to the borrower's real record and your team's permissions.
AI risk summary
On-demand risk analysis on any borrower profile. Score breakdown chart, listed risk factors and strengths, and grade derived from repayment behavior, loan history, and document context.
Copilot assistant
Standalone chat at /assistant with eight function-calling tools. Staff ask questions in plain language; the model looks up borrowers, aggregates stats, and lists overdue notices dynamically.
Auto-resolved context
HKID, application number, loan number, mobile, and name patterns in chat messages auto-load borrower context — so common queries answer in one round-trip without manual lookup.
Email AI extraction
IMAP-connected accounts fetch unread messages, parse body and attachments via AI, deduplicate by message ID, and insert structured InboundLead records with confidence scoring.
Inbound lead review
Low-confidence parses land in pending_review status with admin alerts. Staff correct fields, approve into live records, or reject with audit context preserved.
Prompt administration
Admin prompt controls at /admin/prompts align summary format, risk emphasis, and extraction rules. Research cache and TE inquiry history available for cost and quality monitoring.
Workflows
How AI fits into credit and intake review.
From profile open to risk-informed decision
Credit teams use AI summaries as a starting point — not a final verdict.
- Open a borrower profile and trigger AI Risk Analysis on the summary tab.
- Review score, grade, risk factors, and recommendations alongside raw loan history.
- Ask the Copilot follow-up questions — "show cross-company exposure" or "list overdue notices for this borrower."
- Compare AI output against manual review of application info, DSR, and credit reference modules.
- Export or note findings in case notes for committee or audit trail.
From inbound email to structured lead
Operations teams triage AI-parsed leads before they become live borrower records.
- Scheduled or manual email sync fetches unread messages from connected IMAP accounts.
- AI extracts name, HKID, mobile, loan amount, and purpose into InboundLead fields.
- High-confidence parses can auto-route; low-confidence items queue for staff review.
- Staff correct extraction errors, link Softmedia sources, and approve into CustomerRecord.
- Duplicates are skipped by email message ID; all sync runs return processed/saved/error counts.
Copilot tools and surfaces
Eight tools. Governed edits. Full audit trail.
The AI Review Layer exposes read tools to all authenticated staff (scoped by company) and write tools to admins only. Every chat query and applied edit is logged.
- 01lookup_borrower — search by HKID, name, application number, loan number, or mobile; returns up to 10 deduplicated matches.
- 02get_full_profile — complete merged profile with rule-based risk computation after a single match is identified.
- 03list_top_risk_borrowers — rank borrowers by risk score, active loans, or total loans within company scope.
- 04aggregate_borrower_stats — bucket counts by company, status, risk level, or grade with unique-borrower deduplication.
- 05list_borrowers_filtered — paginated structured lists for blacklisted, high-active-loan, or company-specific cohorts.
- 06list_overdue_notices — live Softmedia overdue and upcoming notices filterable by mode, company, or borrower record.