thebluntcoder/thebluntcoder · GitHub

 

"The best model is the one that runs — and improves — itself."

I'm a Senior Data Scientist based in Gurgaon, India, with 7 years of experience building production ML systems in Fintech. I specialise in credit risk intelligence, bureau-based modelling, alternate data, and increasingly, GenAI-powered financial applications.

My work spans the full credit lifecycle — from raw bureau ingestion to automated underwriting decisions — across personal loans, business loans, credit cards, BNPL, LAP, and two-wheeler finance. I've worked with every layer of Indian lending data: CIBIL tradelines, DPD buckets, enquiry patterns, device signals, bank statement cash flows, and GST intelligence.

What sets me apart: I don't just build models. I build end-to-end decision pipelines — config-driven, confidence-tiered, production-hardened from day one. I've interviewed 200+ candidates for data roles, managed analytics teams of 4–5, and am now building my own fintech product on the side.

I build end-to-end. I ship. I keep it running.

Indifi is a fintech lender focused on SME and business loans. I work on the analytics and model development team, building systems that feed directly into automated underwriting.

The most technically complex system I've shipped. A 9-module Python pipeline that estimates income for sole proprietors using only CIBIL tradeline and enquiry data — no ITR, no bank statements required.

The problem: Self-employed SME borrowers often can't produce salary slips or IT returns. Yet their repayment behaviour, credit utilisation, and enquiry patterns on bureau encode rich income signals — if you know how to read them.

The architecture:

Module 1 → Ingestion & Standardisation     Raw CIBIL data → clean schema
Module 2 → Data Cleaning & Imputation      NaN handling, type enforcement, outlier treatment
Module 3 → Customer Segmentation           Salaried vs. Self-employed classifier
Module 4 → Pre-Filter Engine               Config-driven exclusion rules (loan types, DPD thresholds, account age)
Module 5 → Proxy Income Engine             6 estimation methods across tradeline & enquiry signals
Module 6 → Weighted Median Ensemble        Combines all proxies with method-level confidence weights
Module 7 → Confidence Tier Assignment      HIGH / MEDIUM / LOW / CONFLICT classification
Module 8 → Output Generator               UTF-8 CSV with smart filename pattern
Module 9 → Validation & QA                Distribution checks, segment-wise sanity tests, pivot comparisons

The 6 proxy income methods in Module 5:

• EMI back-calculation — reverse-engineer income from active EMI obligations and standard LTI norms
• Credit limit utilisation reversal — infer income from sanctioned credit limits across lenders
• Enquiry velocity signals — loan-seeking behaviour patterns as income proxy
• Repayment behaviour scoring — consistent on-time payments signal stable income
• Tradeline age & mix weighting — seasoned, diversified credit portfolios indicate income stability
• Loan account type composition — product mix (HL, LAP, BL, PL, CC, TW) encodes income tier

Confidence tiers drive downstream decisioning:

• HIGH → feeds directly into auto-approval logic
• MEDIUM → triggers additional bureau-based validation
• LOW → routes to manual underwriter review
• CONFLICT → methods disagree significantly; flagged for investigation

Key production decisions:

• isinstance(a, str) guards in all lambda functions — (a or "") fails silently on NaN floats; learned this the hard way
• Config-driven pre-filters mean risk policy changes require zero code changes
• Output feeds automated underwriting, pre-approved offer targeting, and fraud detection
• Validation: pivot-table comparison against existing income measures, segmented by tier × segment × income band

Extracts, analyses, and scores Financial Information documents using vision models + Claude AI for automated underwriting. Goal: reduce manual document review to near-zero for standard cases by converting unstructured FI PDFs into structured credit signals.

Built a classification model using demographic, GST, and business turnover data to predict bureau pull necessity. Achieved GINI of 0.54, enabling smarter bureau-pull decisions that reduced per-application bureau costs meaningfully without compromising risk differentiation.

Identified underpriced customer segments via model-driven risk stratification → reduced interest rates for targeted low-risk cohorts → +7% disbursement growth with no increase in portfolio default risk. Also analysed premium customer car loan performance → 0.5% reduction in default risk across personal and business loan books.

FinBox builds alternate data infrastructure for lenders — helping banks and NBFCs underwrite the New-to-Credit (NTC) population that traditional bureaus can't score. I joined as a Data Scientist, grew into managing analytics for specific product verticals, conducted 200+ interviews for DS, DA, and DQ roles, and led a team of 4–5 analysts.

Underwriting NTC customers using device intelligence when there's no bureau history.

DeviceConnect extracts signals from Android devices — transactional SMS, installed apps, location patterns, call logs — and converts them into credit features.

• Engineered ~1,500 features from apps, SMS, location, and call log data
• Built a credit scoring model on alternate device data → AUC 70%
• Feature categories: financial app usage patterns, transactional SMS parsing (salary credits, EMI debits, utility payments), location stability signals, call behaviour features
• +4% AUC boost over baseline from feature engineering alone
• Served personal loan decisioning for NTC borrowers — people with zero bureau history
• Improved SMS extraction coverage from 16% → 25% via regex pipelines across English and Vietnamese SMS (FinBox operated internationally)

Parsing PDF bank statements and Account Aggregator data for income and cash flow analysis.

• Built a credit scoring model on Indian bank statement data → AUC 65%
• Engineered intelligent logic for income calculation — separating salary credits from transfers, reversals, and loan disbursements
• Built revolving transaction detection — identifying circular money flows that inflate apparent cash balances
• Worked with both PDF extraction pipelines and AA (Account Aggregator) framework data
• Output fed lender underwriting processes for personal and business loan decisions

Spearheaded development of a new data product around GST data for business lending intelligence — turnover estimation, filing regularity scoring, and sector-based risk segmentation.

Project Engineer on a USA-based automotive client's infotainment system — testing the phone/connectivity component using regression, sanity, and ad-hoc methodologies. Where I learned rigour, documentation, and what it means to ship software that actually works.

An AI-powered fintech product for the Indian credit card market. Currently in stealth.

"Our lawyers said we can’t tell you more yet, but our engineers said it’s going to be legendary. Watch this space."

Stack: FastAPI · PostgreSQL (Supabase) · Redis · Elasticsearch · Claude AI · Celery · Next.js

More details when we launch. If you're in Indian fintech and find this interesting? — Let's talk.

Full-stack AI agent that plans and orchestrates complete evening experiences on Swiggy's MCP platform (Food · Instamart · Dineout).

You describe your event — occasion, guests, location, budget, dietary preferences. The AI generates a complete plan: restaurant booking, food delivery picks, grocery cart, stitched into a minute-by-minute timeline. One tap to approve; the agent places every order.

Stack: FastAPI · PostgreSQL (Supabase) · Redis · Elasticsearch · Claude AI · Celery · Next.js

Technical highlights:

• asyncio.gather() fires all 3 Swiggy MCP servers in parallel — 2.5× faster than serial calls (~300ms vs ~750ms)
• Claude claude-sonnet-4-20250514 streams the plan via SSE with structured section markers
• Redis-cached offer engine with live validation at checkout
• Follow-up chat UI with full conversation history for plan refinement
• All restaurant names and prices grounded in live MCP data — Claude never hallucinates food details

Stack: FastAPI · PostgreSQL · Redis · Claude API · Next.js 14 · Celery · Alembic · Docker

Fully automated AI news delivery — built end-to-end, runs itself.

• Fetches and curates top AI, tech, and stock news daily
• Summarises articles via LLM + Vector DB into crisp, digestible formats
• Delivers stock market insights alongside news
• Supports interactive triggers — users query topics on demand
• Zero manual intervention once deployed

If you have to touch it every day, it's not done.

7 years of working directly with Indian lending data gives context that no course teaches.

Bureau & CIBIL

• CIBIL score bands (300–900), NTC vs thin file vs established profiles, score as a lagging indicator
• Tradeline types: HL, LAP, PL, BL, CC, TW, gold loan, microfinance — each signals differently for risk
• DPD buckets: 0 DPD vs 1-29 vs 30+ vs 60+ vs 90+ — how vintage and recency of delinquency interact
• Enquiry analysis: hard pull vs soft pull, enquiry velocity, lender type patterns, self vs lender enquiry
• Bureau pull strategy: when to pull, which bureau, how to reduce cost without losing predictive signal

Loan Products

Personal Loan · Business Loan · Credit Card · Loan Against Property (LAP) · Two-Wheeler · BNPL · Microfinance — risk modelling and underwriting across all of them

Alternate Data

• Device intelligence: app usage, SMS parsing, location stability, call patterns for NTC underwriting
• Bank statement parsing: income identification, revolving transaction detection, EMI obligation extraction
• GST data: turnover estimation, filing regularity scoring, sector risk signals
• Account Aggregator (AA) framework: structured cash flow data vs messy PDF extraction trade-offs

Underwriting & Risk

• Income estimation: EMI back-calculation, LTI norms, credit limit reversal, tradeline-based proxies
• Scorecards vs ML models — when each is appropriate in a regulated Indian lending context
• Cut-off strategy, approval rate vs risk trade-offs, vintage analysis, champion-challenger design
• Pre-approved offer targeting vs reactive underwriting — how the targeting changes the risk pool

🃏  CardMax AI (stealth)       AI-powered credit card intelligence for Indian cardholders
🔬  Bureau-only income         Estimating self-employed income without ITR or bank statements
🧾  Document AI                Vision models + Claude AI for FI PDF understanding & scoring
🧩  RAG Systems                Retrieval-Augmented Generation for financial intelligence
🤖  AI Agents                  Autonomous task-orchestrating pipelines for credit workflows
🔁  Agentic backends           FastAPI + Claude API + MCP for multi-step financial agents

 

 

• 🏦 Indian Credit & Bureau Data — CIBIL tradelines, DPD analysis, enquiry signals, NTC underwriting, bureau pull strategy
• 💳 Loan Products — PL, BL, LAP, CC, BNPL, two-wheeler — risk modelling across all of them
• 📱 Alternate Data — device intelligence, bank statement parsing, GST data, Account Aggregator
• 🧾 Document AI — FI PDF analysis, OCR, vision models for financial documents
• 🤖 GenAI & LLMs — Claude API, RAG systems, agentic backends, MCP integrations
• 🔄 Production ML Pipelines — Airflow, config-driven architecture, modular design, feature engineering at scale
• 🐍 Python for Fintech — pandas, FastAPI, async patterns, regex for financial SMS and text parsing
• 👥 Building DS Teams — hiring, mentoring, structuring analytics for fintech products

uttkarsh = {
    "pronouns"         : "he/him",
    "based_in"         : "Gurgaon, India 🇮🇳",
    "day_job"          : "Senior Data Scientist @ Indifi — SME credit & bureau modelling",
    "night_job"        : "Founder @ Stealth Startup — building in stealth 🃏",
    "open_to"          : ["Full-time Roles 💼", "Freelance 🧑‍💻", "Consulting 🤝"],
    "hobbies"          : ["🏏 Watching Cricket", "🏸 Playing Badminton", "🎮 PS5"],
    "guilty_pleasure"  : "Endlessly scrolling YouTube 📺",
    "interviews_done"  : "200+ (DS, DA, DQ roles) — yes, I remember the good ones",
    "iq_score"         : "136 — certified by MyIQ (top 1%) 🧠",
    "current_obsession": "Making credit work for people bureaus can't score 🧾",
    "philosophy"       : "Automate everything. Build once. Run forever. 🚀",
}

Open to full-time roles, freelance, and consulting in ML, AI, and Fintech. Also always happy to talk Indian credit data, bureau intelligence, or alternate data modelling.