Finding the high-value businesses in 50 lakh KYB records.
How we built an automated discovery and scoring engine for Axis Direct — enriching raw corporate filings with digital-footprint intelligence and scoring every viable business on-premises.
- CLIENT
- Axis Direct
- INDUSTRY
- Banking & finance
- MODEL
- On-prem Qwen on vLLM
- STACK
- pgvector · Zod · Next.js
Five million records. No idea who matters.
Axis Direct held a registry of 50 lakh (5 million) KYB records but had no automated way to identify and prioritise high-value commercial accounts.
Static regulatory filings gave fragmented administrative data that didn’t reflect real commercial scale, brand equity or digital traction — leaving relationship managers to find prime prospects across millions of uncurated entities by hand.
A funnel from 50 lakh to a ranked shortlist.
Deterministic pre-filtering
Rules on corporate structure, paid-up capital, vintage, status and GST turnover prune dormant and micro-entities to ~5 lakh candidates.
Vector semantic pruning
1024-dim embeddings with pgvector HNSW compare every profile to golden high-value exemplars, isolating the top ~1 lakh.
Digital footprint resolution
Automated OSINT finds and verifies web domains, LinkedIn and social presence — headcount, reach, engagement and active status.
Explainable on-prem scoring
A self-hosted Qwen with Zod schemas produces composite scores, Tier 1–4, key strengths and audit justifications — reviewed in a Next.js dashboard before CRM export.
Every stage narrows the funnel — and humans sign off before anything reaches the CRM.
- INPUTRaw KYB records5,000,000
- STAGE 1Rule pre-filter~500,000
- STAGE 2Vector similarity~50,000–100,000
- STAGE 3–4OSINT + AI scoringTier 1–4
- STAGE 5Human reviewCRM export
| STAGE | PHASE | FUNNEL VOLUME | IMPLEMENTATION |
|---|---|---|---|
| IN | Raw KYB records | 5,000,000 records | Uncurated commercial registry & statutory filings |
| 1 | Deterministic pre-filtering | ~500,000 candidates | Paid-up capital, vintage, active status, GST turnover |
| 2 | Similarity to golden exemplars | ~50,000–100,000 | text-embedding-3-large @ 1024-dim, pgvector HNSW |
| 3 | Targeted OSINT & social discovery | Enriched profiles | Search orchestration, LinkedIn / Instagram / web resolution |
| 4 | AI scoring & structured output | Tier 1–4 | Self-hosted Qwen on vLLM + Zod; few-shot calibrated on manual scores |
| HITL | Human-in-the-loop review | CRM export | Gate 1: verify AI reasoning · Gate 2: portfolio export |