Arjun Debnath

Arjun Debnath

Senior Data Scientist | ML Solutions Architect | Kaggle Expert

The Elevator Pitch

I bridge the gap between state-of-the-art research and production scalability.

With over 10 years of experience, I specialize in designing solution architectures and productionizing ML/DL solutions on cloud-enabled systems (GCP/AWS). My focus isn’t just on model accuracy—it’s on system efficiency, cost optimization, and real-world business impact across BFSI, E-commerce, Media, and Space industries.

Currently improving fraud detection at Truecaller, where I’ve optimized pipelines to reduce costs by 25% while handling datasets with billions of rows.

View My Resume LinkedIn GitHub Email

Tech Stack & Expertise

I don’t just build models; I build platforms.

  • Core Languages: Python, Scala
  • ML Ops & Engineering: MLFlow, Airflow, PySpark, Apache Iceberg
  • Deep Learning: PyTorch, TensorFlow, Keras
  • GenAI & LLMs: LangChain, Haystack, RAG Pipelines, Pinecone
  • Cloud Architecture: GCP & AWS (End-to-End Pipeline Design)
  • Specializations: NLP, Computer Vision, Time Series

High-Impact Engineering (Selected Work)

1. Optimizing Satellite Inference at Kawa Space

Role: Principal Data Scientist (2020-2022)

  • The Challenge: Processing heavy satellite imagery with high latency, preventing real-time geo-intelligence applications.
  • The Solution: Innovated a data standardization algorithm and built end-to-end pipelines on GCP to consume and preprocess satellite images.
  • The Impact: Achieved a 5x improvement in latency and enabled the landing of our first $10K+ contract by delivering a scalable land boundary detection model.

Read the full technical deep dive →

2. Scalable Fraud Detection & MLOps at Truecaller

Role: Senior Data Scientist (2022-Present)

  • The Challenge: Detecting fraud and spam across billions of transactions while managing spiraling cloud costs.
  • The Solution: Architected automated retraining pipelines using MLFlow triggered by data drift. Implemented semi-supervised learning (DeBERTaV3) for SMS spam detection.
  • The Impact: Reduced infrastructure costs by 25% through Spark optimization and improved fraud detection rates by 8%. SMS spam detection improved by 10%.

Read the full technical deep dive →

3. Building GPT-2 from Scratch with PyTorch

A Technical Tutorial Project

  • The Goal: Demonstrate deep understanding of transformer architecture by implementing GPT-2 from first principles.
  • What You’ll Learn: Attention mechanisms, positional encoding, transformer blocks, training loop optimization, and text generation.
  • Tech Stack: PyTorch, NumPy, Hugging Face Tokenizers

Read the implementation guide →

Competitions & Hackathons

I stay sharp by competing against the best in the world.

  • Microsoft AI Hackathon: Top 10%. Built an LSTM model for query-passage relevance (Top K accuracy: 70%). Code
  • IEEE CIS Fraud Detection: Top 10%. Created a stacked ML model achieving a mean AUC of 0.76.
  • LANL Earthquake Prediction: Top 10%. Developed a sequential GRU model to forecast seismic signals (MAE: 1.42 seconds).

Advisory & Consulting

Generative AI & LLMs (2023-Present)

  • Built scalable LLM applications using LangChain and RAG pipelines for a VC investment firm to analyze unstructured financial data.
  • Impact: Improved investment analysis turnaround time by 15%.
  • Tech Stack: Python, Pinecone, LangChain, GCP

VoiceTech

  • Developed custom WakeWord detection, transitioning a VoiceTech firm to open source.
  • Impact: Saved ~90% of cost with comparable accuracy.

Experience Timeline

Senior Data Scientist | Truecaller | Apr 2022 - Present

  • Leading fraud detection model improvements for fintech clients
  • Reduced cloud costs by 25% through PySpark pipeline optimization
  • Improved fraud detection by 8% and SMS spam detection by 10%

Principal Data Scientist | Kawa Space | Apr 2020 - Apr 2022

  • Led data products team building scalable geo-intelligence platform
  • Created land boundary detection model on satellite imagery
  • Achieved 5x latency improvement through data standardization

Senior Data Scientist | Prudential Vietnam | Jul 2018 - Jun 2022

  • Spearheaded agency and partnership distribution projects
  • Increased Annual Premium Equivalent (APE) from 1.5% to 7% via cross-sell optimization
  • Built real-time ESG news classification with >85% accuracy

Machine Learning Engineer | Quantiphi Analytics | Jun 2017 - Jul 2018

  • Developed NLP frameworks for media clients, improving TRP by 5%
  • Created text-image hybrid classifiers for e-commerce product categorization

Business Analyst | EXL Service | Feb 2016 - May 2017

  • Built fraud detection models for banking clients
  • Developed time series forecasting models with 20% accuracy improvement

Education

Bachelor of Engineering | Jadavpur University | 2015

  • GPA: 7.9/10.0

Let’s Connect

I’m always interested in challenging ML problems, especially those requiring production-scale solutions.