BF Clerk

Guide

How to use BF Clerk POC (v0.1, build 2026-05-16). Screenshots track the portal theme — toggle the sun/moon in the header to flip the guide images too.

What BF Clerk does

BF Clerk pre-fills Malaysian financial-services forms (subscription, switch, redemption, suitability, application) using:

  • Qwen3-VL vision for form-field schema extraction (admin) and name-card OCR (RM)
  • DeepSeek deepseek-chat for field mapping, risk scoring, and Section 5 recommendation drafting
  • pdf-lib coordinate-overlay fill — no AcroForm requirement, works on any PDF

Flow A — Admin template intake

One-time setup per form template, in this portal.

1Open Templates and start an upload

Open Templates and click Upload new template in the top right.

Templates list page with the Upload new template button highlighted
1Click Upload new template to open the modal

2Drop the PDF, name it, pick a category

In the modal, drop a blank form PDF (or click the drop zone to choose), give it a descriptive name, and pick a category. On submit the portal renders each PDF page to a high-resolution PNG and uploads them to storage in the background.

Upload new template modal with drop zone, name and category fields, and Upload button
1Drop the blank PDF here, or click to choose
2Type a descriptive template name
3Pick a category (Onboarding, Subscription, etc.)
4Click Upload — the portal will render pages then redirect

3The AI Wizard auto-fires

On the template detail page, the wizard fires automatically for fresh drafts: Qwen3-VL reads each page and returns a per-field schema with proposed bbox positions, multi-choice cluster IDs, and name-card source hints. Typical run is 30–90s. To re-run later with different settings, use the buttons in the header.

Template detail header with Run AI Wizard, Re-run analysis, Mark as active, Delete buttons
1Run AI Wizard — auto-fires once on a fresh draft; click here to re-run
2Re-run analysis — text-only fallback (deepseek-chat) if the wizard misbehaves

4Review the schema in the Wizard panels

The Wizard schema review section organises the extracted fields with tight grouping: cross-page bundles at the top (name-card-derivable + duplicate-page repeats), then per-page sections containing multi-choice clusters and standalone fields. Per-page counts include any fields that were lifted into the cross-page bundles, with a breadcrumb (e.g. "3 in name-card above") so the math adds up against the Coord Mapper.

Wizard schema review panels showing name-card bundle and per-page sections
1Name-card bundle — fields the RM bot will auto-fill from a name-card OCR at session start
2Page header — total fields on this page, with breadcrumb for any pulled into bundles above
3Multi-choice cluster (collapsed) — click to expand and see each option
4Standalone fields — anything not in a cluster or cross-page bundle
5Page 2 section follows the same structure

5Verify and nudge in the Coordinate mapper

Each field has a numbered badge pre-placed on the PDF preview at Qwen's proposed position. To nudge: click a field on the left, then click the PDF where the badge should sit. To verify: hit the green check on each row, or click the per-page Verify N button to mark all mapped-but-unverified fields at once. Nudging a verified field resets it back to unverified — the new position needs a fresh look.

Coordinate mapper with field list, page navigator, and PDF preview with numbered badges
1Counters — X/Y mapped and X/Y verified; the verified pill turns emerald when all are reviewed
2Verify N button — bulk-mark all mapped-but-unverified fields on this page
3Per-row verify check — click to mark this field verified; nudging resets it
4Pre-placed badges on the PDF — click a field, then click here to reposition it
5Save coordinates — persists all changes (also fires when you Mark as active)

6Mark the template as active

Once the schema looks right and you're happy with the coordinates, click Mark as active in the header. The template is then selectable by RMs in the Telegram bot. You can flip back to draft any time to keep iterating.

Template detail header with Mark as active button highlighted
1Mark as active — flips the template's status from draft to active
2Delete — only available while the template isn't in active use

Need to rename a field, change its type, or toggle required without re-running the wizard? Open the Advanced editor disclosure below the Wizard panels for a flat-table editor.

Flow B — RM form filling (Telegram)

In progress The Telegram bot integration is being wired up. Screenshots will be added once the bot is live. Planned flow:

  1. RM sends a name-card photo to @BFClerkBot.
  2. Bot OCRs the card (Qwen3-VL) and asks "Which form?" with inline buttons listing active templates.
  3. RM picks a form. Bot computes missing fields against the template's schema and asks for them one at a time. Multi-choice fields use inline-button keyboards; free-text fields take typed replies.
  4. When complete, DeepSeek maps fields to coordinates, scores risk (1–5), and drafts the Section 5 recommendation.
  5. pdf-lib fills the PDF via coordinate overlay. Bot sends the filled PDF back to the RM, then asks "Email to client at <OCR'd email>? YES / NO".
  6. RM taps YES → email_sender delivers to the client; session marked completed.

Stack

  • Frontend: TanStack Start on Lovable (React + Tailwind)
  • Backend: TanStack server functions on Cloudflare Workers via Lovable Cloud
  • Data: Lovable Cloud's managed Supabase (Singapore region)
  • Vision: Qwen3-VL-235B-A22B-Instruct via Alibaba Cloud Model Studio (Singapore)
  • Text: DeepSeek deepseek-chat (non-thinking)
  • PDF: pdfjs-dist 5.4.296 (browser render) + pdf-lib (server fill)

What's next

  • Telegram bot wiring (server fns 5–11)
  • Portal-driven manual fill (alternative to Telegram)
  • Email-confirmation step before client send
  • End-to-end demo with the Kenanga Suitability Assessment form