M MANABENDU Contact
Portfolio Backend & ML Engineer
Manabendu Karfa Open to internships

I build backend systems that scale and models that learn and reason.

B.E. AI & ML at Cambridge Institute of Technology, building backend services with Java and Spring Boot alongside deep-learning models and shipping both as production APIs and apps, not notebooks.

+91 83107 47563 · Bangalore, India
About

Manabendu at a glance

2024 — 2028

B.E. Artificial Intelligence & Machine Learning

Cambridge Institute of Technology, Bangalore

CGPA 8.5
2022 — 2024

12th Standard — PCME

St. Joseph's Pre-University College, Bangalore

96%
Stack

Skills that have me covered

Full stack across backend engineering and the ML lifecycle
Models Services Infra
Backend & Frameworks
Java
Spring Boot
FastAPI
REST APIs
Kafka
Hibernate
SQLAlchemy
ML & Deep Learning
Python
Scikit-learn
Neural Networks
TensorFlow
PyTorch
XGBoost
CNNs
AI Algorithms
Genetic Algorithms
Optimization
Multi-Agent Systems
Reinforcement Learning
Generative & Agentic AI
LangChain
LangGraph
Hugging Face
Ollama
RAG
Faiss
ChromaDB
Databases
MySQL
PostgreSQL
SQLite
MongoDB
Tools
Linux
Docker
Git & GitHub
Postman
Jupyter
Available for backend & AI/ML internships Get in touch →
Experience

Experience & contributions

Jan 2026 – May 2026

Software Developer Intern

TwinTallies Private Limited — Punjab, India (Remote)
  • Built internal dashboards and automation tools using Java and Spring Boot, turning real-world business requirements into efficient, scalable backend services.
  • Designed and shipped REST APIs and data workflows, collaborating remotely with the engineering team over a 5-month internship.
Nov 2025 – Jan 2026

Open Source Developer (Contributor)

Social Winter of Code (SWOC) — Remote
  • Selected as a contributor to the Winter of Code Social program, an open-source initiative connecting students and developers with real-world projects.
  • Contributed to multiple open-source projects over three months, picking up new technologies while collaborating with mentors and a distributed developer community.
Projects

Projects built end to end

YellowDoc.ai — Enterprise AI platform

Full-stack platform · FastAPI backend, Streamlit web + native Android client
FastAPI IBM Docling PaddleOCR Groq LLM FAISS OpenPyXL Kotlin
  • Built a production-grade document processing platform with a FastAPI backend that turns invoices, receipts, tax statements, and bank statements into multi-sheet, machine-readable Excel workbooks.
  • Designed an extraction pipeline combining IBM Docling for layout and table-structure analysis with PaddleOCR for scanned and low-quality pages, producing a consistent structured contract for downstream processing.
  • Engineered a deterministic LLM structuring layer over LangChain and Groq: system prompts constrain output to a validated JSON schema, preserving exact numbers, spelling, and row ordering.
  • Synthesized the final workbook with OpenPyXL — sheets, header formatting, and numeric formats generated straight from the validated JSON payload.
  • Added a natural-language document Q&A layer with embedding-based retrieval, returning answers with cited sources and confidence flags.
  • Shipped two clients against one FastAPI service — a Streamlit dashboard and a native Android app in Kotlin with Jetpack Compose Material 3 — backed by MySQL and SQLAlchemy for persistence.
YellowDoc.ai landing page showing a scanned invoice on the left and the generated spreadsheet on the right
YellowDoc.ai document Q&A answering a question about invoice tax totals with cited key points

MONCO — Brain Tumor Detection using Deep Learning

Live on Streamlit Cloud · VGG16 classifier served through a REST API
TensorFlow VGG16 FastAPI Streamlit Ollama
  • Built an end-to-end deep learning system that classifies brain MRI scans into glioma, meningioma, pituitary tumor, or no tumor using transfer learning on VGG16.
  • Reached 90% test accuracy and 94.82% training accuracy across 1,600 test images, with ROC-AUC up to 1.00 on the pituitary and no-tumor classes.
  • Designed a FastAPI backend exposing a REST /predict endpoint that returns the predicted class, confidence score, and full per-class probability distribution.
  • Integrated a locally hosted Ollama gemma3:4b model behind a dedicated prompt-building and inference layer, so every prediction ships with a plain-language explanation.
  • Developed a modular Streamlit dashboard with live MRI upload, an interactive Plotly probability chart, and an explanation panel — API, chart, and UI logic kept cleanly separate.
MONCO dashboard classifying an uploaded MRI as glioma at 99.33% confidence, with a probability distribution chart
MONCO history view listing previous MRI analyses and their results
Recognition

Leadership & achievements

🏆

2nd Place — GDG Code Hunt 2025

Cambridge Institute Of Technology

🏆

2nd Place — Agritech Domain

HACKTRONICS (Circuit to Cloud), CMRIT College

🎓

President, Maths Club

St. Joseph's Pre-University College, 2022 – 2024