B.Tech CSE (AI & ML) · KIIT University · CGPA 9.31
Backend-heavy full-stack engineer with a focus on ML systems — from pipeline architecture through production inference. I build things that ship: deployed APIs, live ML endpoints, and multi-role applications with real users.
📍 Kolkata, India · sohamlodh06@gmail.com · LinkedIn · Portfolio
End-to-end ML pipelines (feature engineering → model selection → FastAPI inference) and production full-stack systems (MERN, Next.js, TypeScript). I care about correctness under load, clean API contracts, and things that work in production, not just on localhost.
Logistic Regression · Optuna · SMOTETomek · SHAP · FastAPI · React
An 18-stage ML pipeline on 50,000 records across three relational data sources. Key engineering decisions: stratified splitting before EDA to prevent leakage, VIF-based multicollinearity elimination (5 features dropped), WoE/IV feature selection (10 retained, IV up to 2.38), and SMOTETomek resampling for 10.6% class imbalance. Bayesian optimisation via Optuna (50 trials, TPE sampler) outperformed a tuned XGBoost baseline.
| Metric | Result |
|---|---|
| AUC-ROC | 0.9834 |
| Macro F1 | 0.9776 |
| KS Statistic | 85.91 |
Inference served via FastAPI + Uvicorn on Render. Frontend on Vercel with SHAP explainability and a logit-space credit scoring algorithm mapping default probability to a 300–900 institutional risk scale.
XGBoost · RandomizedSearchCV · FastAPI · React · Vercel
End-to-end regression pipeline on 9,905 records. Designed a composite total_risk_score feature from compound medical history fields to replace high-cardinality categoricals — materially improved model signal. Validated 18 features via VIF (all < 2.5); hybrid ordinal + one-hot encoding pipeline; tuned XGBoost via RandomizedSearchCV.
| Metric | Result |
|---|---|
| R² Score | 0.9944 |
| vs. Linear Regression baseline | 0.956 |
| vs. Tuned Random Forest | 0.978 |
MERN · JWT · Cloudinary · Vercel
Live (Patient) · Live (Admin/Doctor) · GitHub
Production MERN platform across two independently deployed frontends backed by a centralised Express REST API on MongoDB Atlas. Role-based JWT authentication with separate login flows and permission-scoped dashboards for three account types. Appointment scheduling with slot-conflict detection; hardened with input validation, bcrypt hashing, and rate limiting.
Technical Member — Web Development · Coding Ninjas KIIT Chapter (Aug 2025 – Present)
- Architected a full-stack event management platform (Next.js, Hono, Prisma, MongoDB) with a real-time leaderboard handling concurrent score streams from multiple judges; engineered the write pipeline to prevent race conditions under parallel load.
- Built a secure quiz platform with server-side session tracking, tab-switch detection, timed execution, and randomised question delivery.
- Developed a Project Management System with four role levels enforcing access-controlled workflows via middleware; authored API contracts and permission-logic docs for downstream contributors.
Open Source Contributor · GirlScript Summer of Code (Jul – Oct 2025)
- Implemented authentication modules: OTP login, JWT sessions, Google OAuth, and password recovery — with unit tests across success and failure paths for all four flows.
- Diagnosed and resolved production-level defects by reproducing failure conditions from logs; fixes reviewed and merged into main.
KIIT University — B.Tech, CSE (AI & ML) · 2024–2028 · CGPA 9.31
Relevant Coursework: Data Structures & Algorithms, Operating Systems, DBMS, OOP
- Classify Images with TensorFlow on Google Cloud — Google (2025)
- DSA in Python — NPTEL (2025)
- Postman API Student Expert — Postman (2025)