Hi, I’m Torsha! 👋

I am a Lead ML Architect and NVIDIA-funded Researcher specializing in the intersection of Representational Learning and Autonomous Intelligence. With an MS in Computer Science and a background in Computational Physics, I architect systems that bridge the gap between breakthrough scientific research and production-grade industrial automation.
🚀Current Focus: Building foundational-scale Self-Supervised Learning (SSL) frameworks and Stateful Cyclic Multi-Agent Systems (MAS).
🛡️Career Vision: I am leveraging my expertise in DPO-alignment and latent-space forensics to architect the next generation of Secure AI. My mission is to implement AI TRiSM frameworks to ensure Data Privacy and adversarial robustness, bridging the gap between high-level representational research and high-stakes FinTech and Cybersecurity infrastructures.

When I'm not optimizing inference pipelines or refining foundational research, you’ll find me exploring the physics of the universe through code—or pivoting to deconstruct complex architectures to decipher their deepest vulnerabilities. 🔭

source/img/potrait-alt.png
🏆
NVIDIA Grant Recipient
Lead Architect for ASTRA; Distributed training on 8x A100 GPUs
📈
97.5% Sensitivity
DPO-Aligned Deepfake Detection
🤖
Autonomous MAS
Stateful Cyclic LangGraph Systems

🔬 Innovation & Research Spotlight

📂 Detailed Professional & Academic Portfolio

Click on the categories below for a technical deep-dive

  • Lead ML Architect [NVIDIA Grant Recipient]
    2025 - Present
    1. The ASTRA Ecosystem (SNAD Collaboration): Architected a modular SSL framework and engineered Astra-CLR, the first Contrastive Learning multi-filter time-series Transformer pre-trained on 8x NVIDIA A100 80GB GPUs using TensorFlow MirroredStrategy for astronomical time-series.

    2. Massive-Scale Inference: Pre-trained on ~2.1M samples and scaled to generate embeddings for ~70M ZTF light curves for internal team-wide discovery and research.

    2. Ongoing Research: Spearheading the expansion of ASTRA into non-contrastive architectures, targeting foundational-scale representational learning for multi-survey scientific data.

    3. Status: First-author research submitted to A&C Journal (arXiv); breakthrough methodology in prep for top-tier 2027 ML conferences.
  • Lead Architect: Autonomous AI Portfolio
    2026
    1. Sentinel-Hub: Developed a stateful cyclic Multi-Agent System (MAS) using LangGraph to orchestrate cross-domain security forensics (Vision, Logs, Finance).

    2. DeepShield: Aligned DINOv2 vision transformers with human-perceived risk via DPO-based PEFT, achieving 97.5% sensitivity and reducing False Positives by 51%.

    3. Can-Fin RAG: Engineered a temporal RAG pipeline for multi-modal parsing of financial reports; utilized markdown-based table extraction for longitudinal comparative analysis.

    4. LogSentinel & NeuralAudit: Engineered zero-day anomaly detectors using Knowledge Distillation (MiniLM) and bottleneck autoencoders for infrastructure and financial security.
  • U Lethbridge: Graduate Researcher (Computational Physics)
    2024 - 2025
    Instrumentation & Signal Processing: Optimized Transition-Edge-Sensor (TES) systems for Double Fourier Interferometry; developed statistical routines for spectral curve refinement.
  • The SNAD Team Collaboration
    2024 - 2025
    1. Superluminous Supernovae (SLSN) Discovery: Identified 8 potential SLSN candidates in the ZTF DR8 data release by implementing an Active Learning "Pineforest" algorithm.

    2. Status: First-author publication in Proceedings of Science (PoS).
  • UT Dallas: Independent Research Collaborator
    2021 - 2026
    1. Probabilistic Deep Learning: Engineered a neural network to optimize MCMC solver initialization for inverse tidal evolution modeling.

    2. Status: Co-authored research and submitted to the American Astronomical Society (AAS).
  • MIT: Research Affiliate
    2021 - 2023
    1. Transient Science Pipelines: Developed unsupervised ML pipelines for TESS, Kepler, and PLAsTiCC datasets to identify rare transients and exoplanetary signals.

    2. DASH Optimization: Refactored the DASH spectral classification architecture, migrating from TensorFlow to Keras and implementing batch processing for massive datasets.

    3. Status: Co-authored research in astronomical deep learning published at NeurIPS 2023.
  • UT Dallas: Graduate Researcher (Machine Learning)
    2018 - 2020
    1. Non-Linear Classification Theory: Architected and evaluated ensembles of oblique decision trees to improve representational accuracy in high-dimensional feature spaces.

    2. Status: Published Master’s Thesis.
  • Lead ML Architect (Nu Energy India 🇮🇳)
    Feb. 2026 - Present
    1. Agentic MAS Orchestration: Architected a proprietary stateful cyclic Multi-Agent System (MAS) using LangGraph to automate end-to-end industrial energy and safety auditing.

    2. Infrastructure Design: Developed multimodal ingestion pipelines using Gemini Vision and Supabase (pgvector) to automate entity extraction from technical manuals and instrument logs.

    3. Operational Impact: Implementing deterministic calculation engines to reduce audit reporting cycles from weeks to days with a benchmarked <0.5% error margin.
  • Graduate Researcher (The University of Lethbridge 🇨🇦)
    Jan. 2024 - Dec. 2025
    Precision Instrumentation: Calibrated TES cryogenic detector arrays to optimize signal-to-noise ratios; developed Python-based data analysis pipelines for statistical uncertainty quantification.
  • Sr. Data Scientist (NU Energy India 🇮🇳)
    Aug. 2023 - Dec. 2023
    Full-Stack ML: Designed and implemented a React.js/Plotly Dash error analysis dashboard with a Flask API backend, utilizing Isolation Forest for anomaly detection reducing electrical power generation downtime by 30%.
  • Data Scientist (Verizon 🇺🇸)
    Jun. 2021 - May 2023
    1. Scalable MLOps: Engineered and deployed a recommendation engine on AWS SageMaker , increasing prediction accuracy by 25% via collaborative filtering and A/B testing.

    2. Automation Systems: Architected an end-to-end classification and NER pipeline using ResNet and OCR, utilizing MongoDB for metadata and Boto3 for seamless AWS S3 integration; reduced manual operational effort by 15%.

    3. Leadership: Led a 5-person cross-functional team for the FUZE Regulatory platform, delivering data-driven solutions for nationwide business outcomes.
  • Software Engineer (Centillion Infotech 🇺🇸)
    Jul. 2020 - Feb. 2021
    1. Enterprise APIs: Developed high-throughput RESTful APIs using Spring Boot, achieving a 12% increase in system efficiency for enterprise business products.

    2. Quality Engineering: Implemented comprehensive end-to-end test coverage, resulting in a measurable reduction in system-level customer complaints.
  • Data Analyst (Nu Energy India 🇮🇳)
    Jul. 2017 - Jul. 2018
    Industrial ROI: Analyzed electrical and water-flow parameters for the Indian Railways and Bureau of Energy Efficiency; identified $10M/year in operational losses through regression-based loss modeling.
  • ML & AI (Research): Self-Supervised Vision Transformers (DINOv2/ViT) • Representation Learning • DPO Alignment • Knowledge Distillation (BERT/MiniLM) • Anomaly Detection • NLP (spaCy, NLTK) • SciPy
  • Agentic AI & LLMs: LangGraph • Stateful Cyclic MAS • Tool-Calling • Retrieval-Augmented Generation (RAG) • Prompt Engineering (Gemini 3.1/3.5, Llama 3, OpenAI)
  • Infrastructure & MLOps: Distributed Training • Multi-GPU Orchestration • NCCL • Inference Optimization • Parameter-Efficient Fine-tuning (PEFT) • AWS (SageMaker, S3, Boto3) • Docker/Containerization • NVIDIA Brev • MLflow • Zilliz (Milvus) • Supabase (pgvector) • Automated Deployment Workflows • Linux/Bash
  • Languages & Engineering: Python (PyTorch, TensorFlow, NumPy, Pandas, Scikit-learn) • C • C++ • Java • SQL (PostgreSQL, Oracle) • Flask • React.js • JavaScript
  • Data & Specialized Tools: PySpark • Dask • LSDB • Astropy • MongoDB • Matplotlib • Git/GitHub • LaTeX
  • Strategic Focus & Interests: AI TRiSM Frameworks • Adversarial Robustness • AI Red Teaming • LLM Observability • Infrastructure as Code (IaC)
  • Majumder, T., et al. (In Prep). Large-Scale Self-Supervised Pre-training for Representational Learning in Time-Domain Astronomy. Targeting top-tier ML Conferences (2027).
  • Majumder, T., Malanchev, K., Ishida, E. E. O. Multi-Scale Contrastive Attention for Light-Curve Representation Learning. Submitted to Astronomy and Computing (A&C) Journal, July 2026. arXiv:2606.31627
  • Huang, H., ... Majumder, T., et al. Predicting the Age of Astronomical Transients from Real-Time Multivariate Time Series. Neural Information Processing Systems (NeurIPS 2023). arXiv:2311.17143
  • Majumder, T., et al. Superluminous Supernova Search with PineForest. Proceedings of Science (PoS), PoS(MULTIF2025)041. Accepted Jan 2026.
  • Schussler, J., Majumder, T., Penev, K. Tidal Dissipation Leading to Circularization in Kepler Binaries. Submitted to the American Astronomical Society (AAS) Journal, Apr 2026.
  • Majumder, T. (2020). Ensembles of Oblique Decision Trees. UTD Master’s Thesis. URI: hdl.handle.net/10735.1/8818
    • The University of Texas at Dallas 🇺🇸
      MS in Computer Science (Thesis Track)Aug. 2018 - May 2020
      Thesis: Ensembles of Oblique Decision Trees
    • The University of Lethbridge 🇨🇦
      Graduate Research (Computational Physics)Jan. 2024 - Dec. 2025
      Focus: Characterization and Optimization of the Transition-Edge-Sensor (TES) Detector Systems in a Double Fourier Interferometer
      Note: Completed 9 graduate credits in Computational Physics and Instrumentation.
    • Maulana Abul Kalam Azad University of Technology 🇮🇳
      B.Tech. in Information TechnologyJul. 2013 - Jul. 2017
      Final Year Project: Institute Library Management System
    1. NVIDIA Academic Grant Recipient (NVIDIA), Oct. 2025 – Mar. 2026
    2. University of Lethbridge Graduate Research Award (ULGRA), 2024
    3. Awarded as the Associate Member of the Institute of Engineers (India) in Computer Science and Engineering in 2017.
    4. Awarded for the best 2017 final year project - Institutes Library Management System - by the institution STCET and was among the top finalist for the software development competition by Cognizant (India).
    • Graduate TA (The University of Lethbridge 🇨🇦)
      Jan. 2024 - May 2025
      Applied Instruction: Led laboratory sections for Introduction to Biophysics and Engineering Mechanics, facilitating student mastery of experimental physics, data collection, and engineering principles.
    • Faculty Member (2U Inc. / edX 🇺🇸)
      Dec. 2020 - May 2023
      Technical Mentorship: Led technical instruction for Data Analytics and FinTech bootcamp programs for premier universities across the USA and Australia, focusing on Python-driven financial modeling and data science.

    1. Deep Probabilistic Neural Network for Inverse Tidal Evolution, Sep. 2024
      --- The International Meeting on Eclipsing Binary Star Systems, Weihai, Shandong, China

    2. Unsupervised classification and anomaly detection of TESS transients, Sep. 2022
      --- TESS Science Talk, Massachusetts Institute of Technology, Cambridge, MA, USA