Open to senior IC roles · Pune / Nashik / Bangalore / Chennai · Remote, Hybrid & In-Office

Pratik Mandge

Software Engineer · Applied AI & Production ML · Regulated Fintech

Building intelligent systems for fintech and credit infrastructure. Deep in loan management systems, machine learning, risk models, and AI-powered platforms — in production, at scale.

View work Download Resume Get in touch →

0.954 EWS AUC

60% LLM cost cut

150K+ loans scored

3+ YoE

About

I'm a software engineer focused on production AI in regulated finance. Three years in, I'm building the kind of systems where the software directly decides whether someone gets a loan — which concentrates the mind about correctness, calibration, and what "working in production" actually means.

My background is mechanical engineering, which I traded for code in 2022. That domain overlap turns out useful: manufacturing, lending, and logistics all share the same operational pathology — good data is hard to get, bad decisions have asymmetric costs, and the people who understand the system best usually aren't the ones building the software.

Right now I'm deepening system design and writing publicly about production ML. The goal is a senior IC role at a company where the engineering quality bar is a genuine constraint.

Currently: Software Engineer · Fintech
Location: Bangalore, India
Stack: Python · FastAPI · Django · React · XGBoost · AWS Bedrock
Education: B.Tech Mechanical Engineering · CGPA 8.55
Languages: English · Hindi · Marathi · German (A1)
Open to: Senior SWE / ML Engineer · ₹35–50L+ · India

Experience

Oct 2024 — Present Current

Software Engineer

Fintech Platform YC W17

Fintech platform processing $2B+ in loans across 1M+ customers. Owning the AI/ML surface: credit-risk modelling, LLM infrastructure, and lending-system core logic across five lender clients.

Early Warning Signal (EWS) — end-to-end XGBoost credit-risk pipeline: DPD feature engineering, pincode risk encoding, class-weighted training, daily batch inference. AUC 0.954 · 94.5% accuracy · 150K+ loans. CEO-endorsed company-wide launch.
FinAgent — AWS Bedrock LLM framework (Nova Pro + Llama 3) with SHA-256 prompt caching. AIExplainer for reports, configs, errors and code. 60% cost reduction · 95% UAT accuracy · 2–5s latency. Adopted by three internal teams.
Report Comparison Engine — Excel/CSV/ZIP reconciliation with Standard + Side-by-Side diff modes. Pandas → Polars migration. 95% reduction in manual validation time.
Core Lending System — refactored repayment waiver/reversal/closure logic, extended NACH with multi-MID support, optimised DuckDB/Django ORM report pipelines (hours → minutes).

PythonXGBoost AWS BedrockDjango PolarsDuckDB Celery

Nov 2023 — Sep 2024

Associate Web Developer

WebTech Developers Pvt Ltd · Pune

Built Deskshare — full-stack e-commerce (React + Django REST + MySQL) with 6-language i18n, IP-based location detection, secure payments. Deployed on Nginx; reduced downtime 15%.
Integrated Azure Speech-to-Text and Language Translation APIs. Built CRUD admin dashboard with RBAC — improved admin efficiency 30%.

React.jsDjango REST MySQLAzure

May 2023 — Oct 2023

Technical Trainer

ABC Trainings · Jalgaon

Trained 250+ students in React, Node.js, MERN, Python, Java, C++. 95% satisfaction rate. Developed hands-on curriculum across stacks.

Dec 2022 — Jul 2023

Software Development Intern

Virtusa Consulting Services · Remote

Full-stack social media app (React + Spring Boot + MySQL) with JWT auth and real-time WebSocket chat.

Projects

Production · Fintech

Early Warning Signal (EWS)

Production XGBoost credit-risk model predicting loan default before the next EMI. Complete pipeline: DPD trajectory features, pincode risk encoding, class-weighted training, daily batch inference, visualization API. Iterated LR → RF → XGBoost.

0.954AUC

94.5%Accuracy

150K+Loans scored

PythonXGBoost Scikit-learnDuckDB Django

Production · Fintech

FinAgent — LLM Integration Framework

AWS Bedrock-based AI framework (Nova Pro, Llama 3) with AIExplainer for financial reports, configs, errors and code. SHA-256 prompt caching for semantic deduplication cuts inference cost by 60%. Adopted by three internal teams.

60%Cost reduction

95%UAT accuracy

2–5sLatency

PythonAWS Bedrock DjangoCelery pgvector

Production · Fintech

Report Comparison Engine

Automated Excel/CSV/ZIP reconciliation with Standard and Side-by-Side diff modes. Migrated Pandas → Polars for large-file performance. 95% reduction in manual validation time.

PythonPolars DjangoCelery

GitHub ↗

Kurtiz — Fashion Store

Full-stack clothing e-commerce: catalog, cart, orders, payments, reviews, notifications.

ReactDjango REST PostgreSQLDockerJWT

Konnect — Social Media App

Full-stack social platform with real-time WebSocket chat, notifications, posts, and JWT auth.

React.jsSpring Boot MySQLWebSocket

Autonomous PR Pipeline

Writing up. Agentic system (Claude + OpenCode) that opens, reviews, and rolls back its own PRs in production.

PythonClaude API CI/CD

Skills

Languages

Python
JavaScript / TypeScript
Java
SQL
C++

Frameworks

Django + DRF
FastAPI
React.js
Node.js / Express
Spring Boot

AI / ML

XGBoost / Scikit-learn
AWS Bedrock + LLM Eng.
RAG / Agents / Evals
Polars / Pandas / DuckDB
pgvector / LangChain

Data & Cloud

PostgreSQL
Redis
AWS (S3, Bedrock, EC2)
Docker / Nginx
Celery / async Python

Certifications & Achievements

⚡

EWS Product Launch

AI-Powered Predictive Risk & Default Prevention — CEO-endorsed. April 2026.

☕

Oracle Certified Java Associate

Oracle Corporation · January 2023

☁️

AWS Cloud Economics Accreditation

Amazon Web Services · January 2023

📊

Python for Data Science

Coursera – IBM · July 2022

Education

B.Tech in Mechanical Engineering

R.C. Patel Institute of Technology, Shirpur

Dec 2020 – Jul 2023 · CGPA 8.55/10

Diploma in Mechanical Engineering

Shri Sai Institute of Technology, Amalner

Jun 2017 – Jun 2020 · 84.62%

Writing

Long-form posts on production ML, LLM engineering, and fintech systems. One every 5–6 weeks — depth over frequency.

Coming soon

Building an Autonomous PR Pipeline with Claude and OpenCode

How we built an agentic system that opens, reviews, and rolls back its own PRs in production.

AgentsLLMCI/CD

Coming soon

How We Cut LLM Costs 60% with SHA-256 Semantic Caching

The cache key is the prompt itself. Here's why that's not naive — and why it worked in production on AWS Bedrock.

LLMCostAWS

Coming soon

Production XGBoost for Credit Risk: What Textbooks Don't Tell You

DPD trajectories, pincode risk encoding, calibration, class imbalance. The real pipeline behind EWS.

MLFintechXGBoost

Contact

Open to senior IC conversations in AI/ML engineering, fintech infrastructure, or applied LLM systems. India-based preferred, open to remote for the right role.

The fastest path: email with one sentence about what you're building and one about why you're reaching out. I respond to everything that isn't a mass-send.

Download Resume (PDF)

✉ pratikmandgebussiness@gmail.com → in linkedin.com/in/pratik-mandge ↗ gh github.com/pratikmandge ↗