Resume — PraCha

Six years shipping ML systems in production — manufacturing defect detection, aquaculture analytics, enterprise AI, medical document intelligence. Now at Columbia, developing the theoretical foundation that industry never had time for: causal reasoning, continual learning, decision-making under uncertainty. The through-line is systems thinking — everything can be modeled, understood, and acted on. Available December 2026.

Education

M.S. in Data Science Aug 2025 – Dec 2026

Columbia University — New York, NY

Coursework: Probabilistic Machine Learning, Reinforcement Learning, Causal Inference, Advanced Deep Learning & Generative AI, Analysis of Algorithms, Continual Learning & Memory Models
Teaching Assistant: Causal Inference (graduate), Applied Risk Analytics, Advanced Analytics, Statistical Data Analysis

B.E. in Control Engineering 2015 – 2019

Anna University — Chennai, India

Focus: Control Systems, Model Predictive Control, Dynamical Systems, Signal Processing, Neural Networks & Evolutionary Algorithms

Research Interests

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

Technical Skills

Machine Learning & Deep Learning PyTorch, JAX, TensorFlow, Transformers, CNNs, GNNs, VAEs, Normalizing Flows, Diffusion Models, Energy-Based Models, State-Space Models

RL & Alignment RLHF, PPO, GRPO, DPO, ORPO, Multi-Agent RL, Curriculum Learning, Invariant Risk Minimization (IRM)

Generative AI & LLMs Large Language Models, Fine-tuning (LoRA/QLoRA, instruction tuning), RAG, model evaluation & benchmarking, LangGraph, Vision-Language Models (Qwen, MiniCPM, LLaMA)

Interpretability & Causal Inference Explainable AI, Mechanistic Interpretability, Structural Causal Models, Do-Calculus, IPW, G-Computation, CATE, A/B Testing, Causal Representation Learning

Computer Vision & NLP SAM, YOLOv8, U-Net, FPN/RCNN, GANs, PINNs, BioBERT, Named Entity Recognition (NER), Multimodal OCR, Segmentation Transformers

ML at Scale & Infrastructure Distributed Training (DeepSpeed ZeRO-3, FSDP), CUDA/Triton, AWS SageMaker, Docker, FastAPI, Redis, Dask, MLflow, PySpark

Experience

Machine Learning Engineering Lead (Consulting) — Medical AI & Insurance Document Intelligence Dec 2024 – Jul 2025

Stealth Startup — Bengaluru, India

Built the company's AI capability from scratch — defined the engineering roadmap, hired and mentored the team, and led the transition from a fully manual document-processing workflow to an AI-driven operation.
Medical records, insurance forms, and clinical notes are unstructured — each document type has its own layout, terminology, and implicit structure that a human reader infers from context. Adapted a vision-language model (Qwen-2.5) to do the same: extract specific entities, classify document types, summarize relevant sections. Fine-tuned on 50K+ annotated records; INT4-quantized serving cut processing costs by 3×. Benchmarked against LLaMA-3, SmolVLM, MiniCPM, and commercial OCR APIs.
Built retrieval pipelines to answer complex questions across long document sets — medical chronologies, drug side-effect patterns, injury causation chains — for legal case analysis and clinical decision-making.
Evaluated IBM WatsonX and Discovery against the in-house model on factual grounding, retrieval recall, and cost of ownership; the comparison shaped the company's long-term AI platform strategy.

Senior Machine Learning Engineer Jan 2024 – Jul 2025

Informatica — Enterprise Data Governance & AI — Bengaluru, India

Enterprise support teams solve the same problems repeatedly, but the solutions live buried in ticket histories and wikis. Built an AI system that generates structured troubleshooting playbooks automatically on ticket creation, drawing from 100K+ historical tickets and product docs. Also aligned the language models behind the system (Phi-3, Mistral) using preference data and deployed them at 41% lower latency via quantization and continuous batching.
Log triage is noisy — thousands of lines per incident, most irrelevant. Trained models to surface the failure signals that matter, reducing false-positive triage rates and speeding up incident response.
Ran causal experiments to measure the actual effect of AI chatbot deployment on ticket resolution time. The naive comparison was confounded by which tickets got routed to the chatbot in the first place. Doubly-robust estimation isolated the real impact and shaped which automation investments to pursue next.
Built churn prediction that Customer Success can actually use: instead of a black-box risk score, decomposed each account's usage into trend, seasonal, and idiosyncratic components — so the reason for the risk is visible, not just the risk itself.

Data Scientist II Nov 2022 – Dec 2023

Captain Fresh — Satellite-Based Aquaculture Intelligence — Bengaluru, India

Built a system to map aquaculture ponds from satellite imagery across 1,200+ hectares. Ponds don't come labeled — trained a model to find and segment them from Sentinel-2 data (F1 of 0.92), with Dask pipelines enabling near-real-time water quality monitoring at scale.
Satellite images of water are often blocked by clouds — worst in monsoon season, when farmers need data most. Fused radar (Sentinel-1) with optical (Sentinel-2) imagery: radar penetrates clouds while optical provides finer detail. Framed cloud removal as masked reconstruction and improved cloud-free analysis coverage by 20.5%.
Ran field experiments to identify what actually drives shrimp growth — controlled for pond conditions, feed, and weather to isolate causal drivers, translating findings into optimized harvest cycles. Deployed on AWS SageMaker with a 37% latency reduction; results presented at OSICON-2023.

Decision Scientist · Apprentice Leader Jan 2019 – Nov 2022

Mu Sigma — Decision Sciences and Machine Learning at Scale — Bengaluru, India

Led an $8M supply chain transformation for a Saudi petrochemical firm — demand sensing, inventory optimization, and distribution routing across large-scale industrial operations — managing 35 engineers and analysts. Delivered a $500K follow-on contract.
Solar energy forecasting breaks when the sky gets complicated. Built a system where physics models handle the predictable part (sun position, atmospheric geometry) and ML corrects what physics gets wrong (turbulence, cloud interference). Achieved 4.5% nMAPE, enabling $6M/year in reduced bidding risk.
Built a manufacturing defect detection system that spots defects in production imagery faster and more reliably than human inspection — reaching mAP of 0.93. Extended with synthetic data generation for rare defect types and physics-informed process optimization.
Pharmaceutical research means reading thousands of papers to find where drugs, diseases, and genes are mentioned. Built a model that reads them automatically — fine-tuned on biomedical text — achieving F1 of 0.94 on entity extraction, accelerating literature review pipelines.

Research & Projects

Zero-Shot Coordination in Multi-Agent Systems via C-GRPO 2025–2026 · JAX · PyTorch · Dec-POMDPs · RLVR

Agents trained separately fail to coordinate when paired with a stranger — they build idiosyncratic habits that don't transfer. Built C-GRPO, which trains agents using group-relative outcomes as a reward signal: when some pairings in a training batch coordinate well and others don't, that contrast teaches more general strategies. No value function needed.
Achieved 83.7% zero-shot coordination on Overcooked-V2, compared to 71.4% for the best prior method, with 28% lower variance across unseen agent pairings.

No-Exemplar Continual Learning via Causal Invariance 2025–2026 · PyTorch · Continual Learning · Columbia Final Project

Asked why neural networks forget old tasks when learning new ones — and found the answer in causal structure: networks confuse what defines a class with the incidental visual context of training photos. Separate those two causes, and prototypes stay stable across sessions without storing a single past example.
Built five methods exploiting this structure at different stages — from backbone regularization to covariance pooling to test-time augmentation averaging — all using a simple nearest-class-mean classifier. Beat a strong replay-based baseline (200 stored images) by over 30 percentage points, with zero stored examples.
Found that Invariant Risk Minimization, the original theoretical tool, hurt performance when environments were limited — a clean result about when causal regularization works and when architectural choices are the right lever instead.

HMV-CRL: Separating Platform Influence from Genuine Preference in Recommendation 2026 · PyTorch · Causal Representation Learning

Recommendation engagement conflates two distinct causes: what users genuinely want and what the algorithm pushed in front of them. Built a model that learns both separately — one representation anchored to search behavior (user-initiated, preference-revealing), one anchored to recommendation exposure (platform-initiated, algorithm-shaped) — using dual-surface interaction logs for identifiability.
The preference representation stayed selective and stable even under simulated algorithm changes, while the platform representation shifted dramatically — confirming the two factors were genuinely separated, not just named separately.
Mediation analysis revealed the core finding: the platform amplifies observed engagement while suppressing the genuine preference signal, acting as a suppressor. A mechanism completely invisible to standard engagement metrics — only visible once the representations are causally separated.

Hierarchical Causal SSM for Counterfactual Simulation 2025–2026 · PyTorch · Normalizing Flows

Clinical trials answer whether a treatment works on average — the harder question is what would have happened to this specific patient under a different treatment. Built a model that separates what is stable about a person from what the treatment changed, enabling individual-level counterfactual predictions from observational data.
Outperformed existing counterfactual methods on a longitudinal benchmark: PEHE of 0.21 vs. 0.34 (CRN) and 0.31 (CT).

Energy-Based Port-Hamiltonian Neural Networks 2026 · JAX · Triton · CUDA

Neural networks can model physical motion — but standard architectures don't know energy is conserved. Given long enough, they drift. Built a network where conservation of energy is baked into the structure, not treated as a soft constraint to be learned.
Reduced long-horizon energy drift by 91% vs. standard neural ODEs — energy error of 0.8% vs. 8.9% over 1000 steps on dynamical benchmarks. Implemented the symplectic integrator as a custom Triton kernel for GPU efficiency.

Personality-Aligned VLM — Conversational Coaching · EchoLogics: Synthetic Consumer Research Platform 2025 · MiniCPM · ORPO · DPO · QLoRA

Coaching assistants need a consistent voice — not a generic helpful tone. Fine-tuned a vision-language model (MiniCPM) on real coaching interaction patterns to embed specific traits (wit, confidence, playfulness). Human preference evaluation: 73% preferred the aligned model vs. 61% baseline; live A/B testing confirmed 18% engagement lift. Deployed via FastAPI + Redis + Docker.
Extended into EchoLogics: a platform for generating synthetic consumer panels with distinct personality profiles for rapid market research. Panels are census-aligned, each respondent has a consistent Big Five personality profile, and a statistical evaluation layer measures the significance of findings across panel configurations.

FAIRE — Frontiers in AI Research and Engineering 2025–ongoing · Self-Designed & Self-Directed Research Program

Self-directed sprint-based research program. Each sprint begins with one precisely defined question and ends with a concrete artifact — working code, a quantified finding, or a documented failure. Active tracks: frontier model engineering (aops-fms), causal continual learning (causal-continual), physics-constrained dynamical systems (67systems), and LLM interpretability on social agents (interp-exp1-socialagents).
Four objectives structure every sprint as a loop: Discovery · Evidence · Inference · Optimization. All work is version-controlled and published; results are reported regardless of sign.

Thursday Learning Hours 2025–ongoing · Weekly Self-Run Seminar

A weekly seminar I create and run for myself: each session covers one foundation, frontier, or framework in AI/ML/DS — slides, reading list, experiments, and notes. Topics range from theoretical foundations (random matrix theory, information geometry) to research frontiers (mechanistic interpretability, diffusion flows, world models, causal representation learning). The discipline is the point. Consistency compounds.

Exploratory Research: Persona Graph Architecture · Causal LLM Interpretability · RSSM World Model 2025–2026 · PyTorch · JAX · Variational Inference

Designed a Persona Graph architecture: personality as explicit graph structure (Big Five trait nodes, psychometric relationships), with a language model as the stateless execution engine. Most AI assistants treat personality as a prompt instruction — this makes it a structural constraint. Validated via ablations showing how controlled personality perturbations shift output behavior; the architecture became the foundation of EchoLogics.
Applied causal tools to understand how language models reason during multi-turn negotiation — extracted causal graphs over agent reasoning traces and tested whether reasoning patterns hold consistently across different scenarios.
Built a Hierarchical Causal RSSM separating customer behavior into stable identity, natural dynamics, and treatment response — with hard causal gates preventing information leakage between channels. Individual-level counterfactual predictions from observational data. Achieved 2.4× reduction in confounding bias and +8.2% sales lift on Dunnhumby retail data.