N°01

Mervin
Mandanna

I build intelligent vision systems that transform raw pixels into understanding.

01 — About

ResearchProfile

I work at the intersection of computer vision and applied machine learning — currently building deep learning pipelines that turn raw line-scan imagery into automated defect detection for railway inspection systems. Earlier research at IIT Roorkee applied the same static/dynamic ML mindset to a different signal: classifying Android malware through hybrid static, dynamic, and deep-learning analysis.

Alongside applied CV work, I build full-stack systems that put models into use — from a reinforcement-learning pipeline for fleet decision-making to an agentic browser-automation platform. I currently serve as Secretary of ByteXync, my college's technical club, where I help run hackathons and technical programming for the CSE department.

Current Focus

Computer VisionImage SegmentationGeometric VisionDeep LearningApplied Malware ML

Research Interests

Vision TransformersMultimodal AIGenerative AIReinforcement LearningAgentic Systems

Future Direction

Longer-term, I'm drawn to Quantum Machine Learning — exploring how quantum computing could eventually reshape how learning algorithms represent and search high-dimensional problems.

02 — Trajectory

Experience&Education

Intern — Scientist

January 2026 — Present

Patil I-Labs, Patil Rail Infrastructure Pvt. Ltd · Bangalore, India

Building computer vision and deep learning systems for Machine Vision Inspection Systems (MVIS), automating railway wheel inspection and defect detection.

Problem

Manual inspection of railway wheels for defects is slow and inconsistent; line-scan imagery needs to be turned into reliable, automated defect signals.

Impact

Preprocessing and synthetic-data pipelines now underpin downstream defect-detection model development.

  • Designed preprocessing pipelines for line-scan wheel imagery — geometric correction, wheel boundary extraction, and image normalization.
  • Built synthetic data generation and segmentation workflows to improve model development and evaluation for real-world deployment.
OpenCVPyTorchImage SegmentationGeometric Correction

Research Intern — Android Malware Detection

September 2025 — November 2025

Indian Institute of Technology (IIT) Roorkee

Researched Android malware detection via static, dynamic, and hybrid ML/DL analysis; co-authored a paper on a hybrid detection framework.

Problem

Static or dynamic analysis alone each miss classes of Android malware; a hybrid ML/DL approach was needed to improve detection robustness.

Impact

Co-authored a paper advancing hybrid ML/DL approaches to Android malware detection.

  • Built feature extraction pipelines with Androguard, DroidBox, and Cuckoo Sandbox.
  • Trained and compared SVM, Random Forest, and CNN models across static, dynamic, and hybrid feature sets.
AndroguardDroidBoxCuckoo SandboxSVMRandom ForestCNN

Frontend Developer

February 2025 — August 2025

Win Research Center · Bangalore, India

Built and maintained frontend features for interactive research dashboards using React, Next.js, and Tailwind CSS.

Problem

Cross-functional research tooling needed faster, more responsive dashboards for day-to-day use.

Impact

Reduced load times by 30% through state-management and rendering optimizations.

  • Built and maintained frontend features using React.js, Next.js, and Tailwind CSS.
  • Optimized state management and reduced load times to improve day-to-day UX.
ReactNext.jsTailwind CSS

Bachelor of Engineering, Computer Science

Expected May 2027

Dayananda Sagar College of Engineering · Bangalore, India

CGPA: 8.47. Coursework and independent research spanning machine learning, computer vision, and systems.

03 — Research

FeaturedResearchProjects

Two end-to-end systems that pair a modeling approach — reinforcement learning and adaptive AI automation — with a production-shaped full-stack architecture.

August 2025

KMRL

View on GitHub

A multi-objective optimization system that uses reinforcement learning to decide which trainsets are fit for service.

Problem

A metro operator needs to select which trainsets from a fleet are serviceable and ready for active duty, weighing operational status and certification constraints simultaneously — a combinatorial decision that doesn't reduce to a single metric.

Method

  • Modeled trainset selection as a multi-objective optimization problem evaluating 25 trainsets against operational and certification parameters.
  • Applied reinforcement learning to learn serviceability policies and automate the selection of 13 active units.
  • Exposed the decision pipeline through a full-stack architecture for scalable, repeatable analysis.

Challenges

Balancing multiple, sometimes conflicting objectives (operational readiness vs. certification constraints) inside a single reward signal that produces stable, explainable policies.

Architecture

FastAPI backend hosting the RL environment and optimization logic, with a Next.js frontend for interactive decision analysis and visualization.

Results

  • Evaluated 25 trainsets on operational and certification parameters.
  • Automated selection of 13 active units for service.

Tools

Next.jsFastAPIReinforcement LearningPyTorchGymnasiumpandasNumPy
September 2025

Mimiker AI

View on GitHub

An AI-powered browser automation platform that records human workflows and replays them adaptively when the UI changes.

Problem

Traditional browser automation breaks the moment a target UI changes structure, forcing manual rewrites of brittle scripts for repetitive web workflows.

Method

  • Built a record-and-replay engine that captures user workflows directly from browser interactions.
  • Used Gemini AI to interpret UI structure changes and adaptively update automation steps instead of failing outright.
  • Deployed the resulting workflows as reusable, cloud-hosted automations.

Challenges

Making automation resilient to UI drift — detecting when a recorded selector or flow no longer matches the live page, and adapting the workflow rather than failing silently.

Architecture

Next.js frontend for workflow recording/management, Flask backend orchestrating Playwright-driven browser automation, MongoDB for workflow storage, Gemini AI for adaptive step reinterpretation.

Results

  • Improved task efficiency by 70% through reusable, adaptive cloud workflows.

Tools

Next.jsFlaskMongoDBGemini AIPlaywright

05 — Publications

Publications

A Hybrid Machine Learning Framework for Android Malware Detection

In Preparation

Co-authored with the research team at IIT Roorkee

Manuscript — target venue to be announced

Static and dynamic analysis of Android applications each capture different, incomplete signals of malicious behavior. This work combines feature extraction across static, dynamic, and hybrid analysis pipelines — built with Androguard, DroidBox, and Cuckoo Sandbox — and evaluates SVM, Random Forest, and CNN classifiers to advance ML/DL-based hybrid detection approaches.

More publications will appear here as research work is completed and submitted.

06 — Toolkit

TechnicalSkills

Machine Learning

PyTorchTensorFlowScikit-learnCNNsModel TrainingFeature Engineering

Computer Vision

OpenCVscikit-imageImage ProcessingImage SegmentationFeature MatchingHomography EstimationMulti-view Geometry

Security & Malware Analysis

AndroguardDroidBoxCuckoo SandboxStatic AnalysisDynamic Analysis

Backend & Data

FastAPIFlaskNode.jsREST APIsPostgreSQLMongoDBFirebase

Programming Languages

PythonC++GoJavaScriptTypeScript

Frontend

ReactNext.jsTailwind CSS

Tools & Infrastructure

GitDockerLinuxAWSCOLMAPHuginLabelImgNumPypandasMatplotlib

07 — Record

Achievements

0

Research & Industry Internships

0

Flagship AI Research Projects

0

Hackathon & CTF Placements

0

Co-authored Paper

0

CodeChef Rating

Commit and Conquer

Runner-up

Open Source Contribution Hackathon.

Infrentia

4th Place

Hackathon conducted by PES University.

BotCraft

3rd Place

Discord Bot Development Competition.

Cipher Chase

16th Overall

Capture-the-flag competition organized by IIIT Bangalore.

Secretary, ByteXync

Leading ByteXync, the official technical club of Dayananda Sagar College of Engineering's CSE department.

08 — Contact

Let'sTalk

Open to research collaborations, internships, and conversations about computer vision and applied ML.