Prabakaran Chandran

+1 (347) 586-1627 · pc3197@columbia.edu · New York, NY · LinkedIn · GitHub

Education

Master of Science, Data Science Aug 2025 – Dec 2026
Columbia University — New York, NY
  • Coursework: Machine Learning, Causal Inference, Causal Machine Learning, Advanced Deep Learning and Generative AI, Large Language Models, Probabilistic Models and Machine Learning, Statistical Inference and Modeling, Continual Learning and Memory Models, Reinforcement Learning, Design and Analysis of Algorithms
  • Teaching Assistantships: Applied Risk Analytics, Causal Inference, Advanced Analytics, Statistical Analysis
Bachelor of Engineering, Control Engineering May 2015 – May 2019
Anna University — Tamil Nadu, India — GPA: 8.63 / 10
  • Control Systems, State-Space Models, Model Predictive Control, Dynamical Systems, Optimal Control Theory, Signal Processing, Neural Networks and Evolutionary Algorithms

Research Interests

  • World Models
  • Causal Inference
  • Continual Learning
  • Model-Based Reinforcement Learning
  • Deep Learning
  • Probabilistic Modeling
  • Cognitive Science
  • Complexity Science
  • Social Sciences

Technical Skills

ML & Deep Learning Supervised/Unsupervised/Self-Supervised Learning, CNNs, Transformers, Attention Mechanisms, Multi-Task Learning, Transfer Learning, Representation Learning, Energy-Based Models, Generative Models (VAEs, GANs, Normalizing Flows, Diffusion Models), Neural Architecture Design
LLMs & Gen AI LLM Fine-Tuning (LoRA, QLoRA, Full Fine-Tuning, Continued Pre-Training), RLHF, RLAIF, DPO, ORPO, RAG, Agentic Systems, Prompt Engineering, Vision-Language Models, Long-Context Reasoning
Causal & Probabilistic Structural Causal Models, do-Calculus, Identifiability, Counterfactual Reasoning, Treatment Effect Estimation (ATE/CATE), IPW, Causal Representation Learning, Bayesian Inference, Variational Inference, Graphical Models, Latent Variable Models, State-Space Models
Reinforcement Learning Policy Optimization (PPO, GRPO), Multi-Agent RL, Zero-Shot Coordination, Dec-POMDPs, Curriculum Learning, Model-Based RL, Model Predictive Control
Computer Vision Image Segmentation (SAM, U-Net, SegFormer), Object Detection (FPN, RCNN), Satellite/Remote Sensing Imagery, 3D Architectures, Medical Image Analysis, OCR
NLP & IE Named Entity Recognition (BioBERT, SpaCy), Sentence Transformers, Contrastive Learning, Few-Shot Classification, Knowledge Graph Construction, Document Understanding
MLOps & Infra AWS SageMaker, Docker, FastAPI, Redis, CI/CD, Model Serving, Latency Optimization, Benchmarking Pipelines, Data Annotation Pipelines
Programming Python, PyTorch, TensorFlow, JAX, PySpark, Dask, Hugging Face, Weights & Biases, SQL, Git
Mathematics Optimization (Convex, Non-Convex, Variational), Linear Algebra, Probability & Statistics, Control Theory, Dynamical Systems, Differential Equations, Information Theory

Experience

Machine Learning Engineering Lead Dec 2024 – Jul 2025
Stealth Startup — Bengaluru, India
  • Built and scaled the company's first AI team from the ground up, owning hiring, onboarding, and technical mentorship to establish core AI/ML capability for agentic systems and vision-language model development in the medical and insurance domains.
  • Led domain-specific fine-tuning of a Qwen VL 2.5-based OCR model on 50,000+ medical records using continued pre-training and LoRA-based adaptation, optimizing extraction accuracy for healthcare and insurance documents.
  • Established a production-grade data preparation pipeline with 50,000+ human-annotated documents and built comprehensive benchmarking SOPs, evaluating against commercial APIs and open-source VLMs (smolVLM, MiniCPM, LLaMA-3).
  • Drove technical evaluation of IBM WatsonX and Discovery platforms for document understanding and RAG, benchmarking against internal VLM pipelines to inform strategic platform decisions.
  • Designed long-context RAG pipelines for complex document analysis — medical chronology extraction, side-effect identification, and injury-related insight mining — supporting legal and healthcare decision-making workflows.
Senior Machine Learning Engineer Jan 2024 – Jul 2025
Informatica — Bengaluru, India
  • Fine-tuned lightweight LLMs (Phi-3, Mistral) using RLAIF, DPO, and ORPO to build and deploy chat-based troubleshooting agents across customer-facing support interfaces for enterprise data engineering and governance products.
  • Engineered few-shot classifiers with sentence transformers and contrastive learning to detect golden signals in noisy system logs, improving issue resolution speed and production observability.
  • Built end-to-end automated RAG agents that generated troubleshooting playbooks on-demand upon job failure ticket creation, reducing mean time to resolution.
  • Designed and executed causal inference experiments (inverse propensity weighting, stratification) to quantify the impact of AI-driven ticket prioritization and chatbot deployment on support resolution time.
  • Developed interpretable latent state-space models using Kalman filters and dynamic factor models to surface enterprise consumption patterns, enabling Customer Success teams to proactively identify churn risk and protect renewal revenue.
  • Constructed knowledge graphs and embedding-based RAG pipelines from support tickets, product docs, and internal wikis, powering context-aware agent responses and accelerating support engineer onboarding.
Data Scientist II Nov 2022 – Dec 2023
Captain Fresh — Bengaluru, India
  • Developed and fine-tuned segmentation models (SAM, U-Net, Segmentation Transformers) on Sentinel-2 satellite imagery for aquaculture pond identification, achieving F1 = 0.92 and building Dask-powered pipelines for water quality and harvest planning across coastal Andhra Pradesh.
  • Designed a time-series 3D Transformer architecture for cloud removal from satellite imagery, improving analysis reliability by 20.5% by fusing Sentinel-1 SAR and Sentinel-2 multispectral data.
  • Conducted randomized field experiments with in-situ water quality metrics and applied matching, stratified regressions, and treatment effect estimation to identify causal drivers of shrimp growth and optimize feeding cycles.
  • Built an MVP Global Shrimp Trade Atlas combining satellite-derived aquaculture data with export/import market data and price indexes to inform procurement strategy.
  • Deployed an end-to-end MLOps pipeline on AWS SageMaker for automated farm segmentation, reducing inference latency by 37% over 1,200+ hectares and shipping a cloud API for real-time water quality and harvest monitoring.
Apprentice Leader Jul 2022 – Nov 2022
Mu Sigma — Bengaluru, India
  • Led a USD 8M supply chain transformation for a Saudi Petrochemical firm, architecting the full supply network optimization and demand sensing stack while recruiting and training a team of 35 engineers and analysts.
  • Won and delivered a USD 500K engagement for mathematical process optimization, leading a 10-member cross-functional team.
Decision Scientist Jan 2019 – Nov 2022
Mu Sigma — Bengaluru, India
  • Architected a hybrid ML solution for a Japanese green energy firm combining a physics-based ODE model with a Moving Windowed CatBoost model on error-corrected meteorological forecasts, achieving 4.5% normalized MAPE and enabling ~$6M in estimated annual savings through accurate solar energy bidding.
  • Built a multi-task learning architecture (FPN + RCNN backbone) for product segmentation and defect detection in additive manufacturing, achieving mAP = 0.93 and reducing layer-by-layer QA from days to hours via a deployed 3-tier web application.
  • Enhanced CAD design validation using Pix2Pix conditional GANs and Physics-Informed Neural Networks (PINNs) to optimize manufacturing process parameters.
  • Designed an Agent-Based Modeling framework for a US-based beverage client to simulate population consumption behavior, applying non-parametric hypothesis testing to construct Bayesian Network schemas with conditional probability tables.
  • Built a BioBERT-based NER pipeline for genetic and gastroenterological entity extraction in pharmaceutical research papers, achieving 100% recall and significantly accelerating literature review workflows.
  • Engineered Discrete Event Simulation blocks for retail warehouse inventory KPI forecasting and implemented Genetic Algorithm-based optimization for demand-driven replenishment strategies.
  • Developed regularized regression and mixed-effects models for trade promotion simulation, time series clustering for product segmentation and cannibalization analysis, and causal ML techniques for promotion impact evaluation.

Research & Projects

Zero-Shot Coordination in Multi-Agent Systems via C-GRPO and Adaptive Curriculum Learning 2025–2026
  • Formalized zero-shot coordination as a cross-play generalization objective in Dec-POMDPs with partial observability, benchmarking across diverse Overcooked-V2 coordination layouts.
  • Engineered a visuomotor policy architecture with residual CNN encoders, multi-head spatial attention, global–local feature fusion, and GRU-based recurrence for partner-aware spatial reasoning and long-horizon temporal dependencies.
  • Implemented C-GRPO, a critic-free policy optimization method using group-relative advantage normalization as an implicit counterfactual baseline, improving timestep-level credit assignment under sparse rewards.
  • Built an adaptive curriculum with proficiency-gated progression, rehearsal sampling, reward-shaping annealing, and action-prior biasing to mitigate exploration collapse and enable robust multi-task coordination.
Hierarchical Causal State-Space Modeling with Normalizing Flows for Counterfactual Simulation 2025–2026
  • Developed a hierarchical causal recurrent state-space model for observational retail basket time series, enabling interventional effect estimation and temporally consistent counterfactual rollouts for marketing actions.
  • Architected a disentangled latent decomposition with invariant identity factors, baseline dynamics, and an intervention-response channel for interpretable separation of intrinsic behavior vs. causal lift.
  • Implemented causal isolation constraints — hard-gated treatment pathways, leakage-resistant decoding, and sequential g-computation-style conditioning — to prevent post-treatment information leakage.
  • Integrated conditional normalizing flows for expressive amortized posterior inference in non-Gaussian, multi-modal regimes with stabilized objectives to mitigate KL pathologies and training collapse.
Energy-Based Port-Hamiltonian Neural Networks for Structure-Preserving Physics Learning 2025–2026
  • Designed a Port-Hamiltonian Neural Network (PHNN) as a differentiable physics engine, parameterizing the Hamiltonian energy function with automatic differentiation for structure-preserving dynamics and learned dissipation.
  • Implemented a compositional energy network factorizing energy into interpretable kinetic and potential components aligned with mass–spring interaction graphs, enabling modular generalization across topologies.
  • Framed dynamics learning as an Energy-Based Model scoring transition validity, enforcing manifold-consistent rollouts through contrastive energy landscapes.
Generative Social Network — Persona Graph-Driven AI Agents for Enterprise Collaboration 2025
  • Designed an enterprise social network of Generative AI Agents using psychometric embeddings (Big Five, MBTI) for automated collaborator discovery and co-value creation ideation.
  • Proposed a Persona Graph architecture inverting traditional LLM-centric agent design: cognitive core in an explicit graph encoding psychological and behavioral models, with the LLM as a stateless execution engine.
Personality-Aligned Vision-Language Model for Relationship Coaching 2025
  • Built an end-to-end VLM coaching assistant with ORPO and DPO-based preference alignment to embed target personality traits from expert interaction patterns.
  • Designed multi-dimensional evaluation pipelines (human preference rankings, GPT-4-as-judge, trait-specific rubrics) and conducted A/B testing to validate alignment fidelity and engagement lift.
  • Deployed as a production API with FastAPI, Redis session caching, and Dockerized inference serving with real-time latency optimization.