Sethu Iyer

About Me

Background

Math and CS from BITS Pilani. I work on LLM interpretability and explore physics-inspired approaches to optimization, treating SAT problems like thermodynamic systems.

Research Interests

LLM transparency and physics-informed algorithms. The math is open source if you're curious.

Current Focus

Active Research Areas

Attention Entropy based interpretability metrics for LLMs
TPU mixed precision fine-tuning for LLM tool selection and planning modules
NitroSAT — Physics-informed MaxSAT solver treating SAT as Langevin flow on a prime-weighted hyperbolic manifold
BAHA — Branch-Aware Holonomy Annealing for detecting thermodynamic phase transitions via specific heat monitoring
Multiplicative PINNs — Euler product gates replacing additive loss to eliminate gradient conflict in Physics-Informed Neural Networks

Skills & Expertise

AI/ML & LLMs

Hugging Face Transformers
PEFT (LoRA, QLoRA)
Retrieval-Augmented Generation
Prompt Engineering
Fine-tuning LLMs
Computer Vision

Programming & Development

Python (Advanced)
C/C++, Go, Scala
SQL, CUDA, MATLAB
Flask, Docker
Svelte, Quasar
CI/CD, Linux

ML Frameworks & Tools

TensorFlow, PyTorch
Scikit-learn, SpaCy
OpenCV, XGBoost, CatBoost
ONNX Runtime
Mixed Precision Training
TPU Utilization

Big Data & Databases

Apache Spark
Elasticsearch
Redis
MongoDB
PostgreSQL

Work Experience

Workato — Senior AI Research Engineer May 2025 – Present

Bengaluru

Enterprise search and relevance systems for automation workflows.

Building semantic search for business queries across heterogeneous data sources
Improving retrieval accuracy for multi-entity, multi-source search scenarios
Working with LLMs for enterprise knowledge discovery

Stack: Python, FastAPI, OpenSearch/Lucene, Redis, PostgreSQL, embeddings

Deep Learning Researcher & Data Scientist Feb 2022 - Apr 2025

Jio Platforms Ltd (Hyderabad, India)

Developed novel interpretability metrics for LLMs to enhance transparency in decision-making
Enhanced long-text coherence in LLMs by integrating episodic memory mechanisms
Created a Clinical Diagnosis Engine using Graph RAG, achieving 85.83% production accuracy
Core contributor of a multi-lingual, contextualized medical entity search system

Research and Development Engineer Sep 2018 - Feb 2022

Amelia / IPSoft (Bengaluru, India)

Core contributor to EazyML — an explainable AutoML platform for tabular data
Designed a skin disease classification system using Google AutoML Vision (F1 Score: 0.64 over 65 diseases)
Built a zero-shot concept matcher with BERT and mathematical distance metrics, ideas to revenue within 40 days
Refactored legacy codebases with modern Python OOP principles, type annotations, and best practices

Research Intern Jul 2018 - Sep 2018

Video Analytics Lab, IISc Bengaluru

Accelerated Sketch-GAN model training by 3x via CUDA optimization
Created PyQt widgets for streamlined data annotation workflows

Student Researcher Jul 2017 - Jul 2018

BITS Pilani Goa Campus

Game Theory in Machine Learning (Jul 2017 - Dec 2017)

Research under Dr. Jajati Keshari Sahoo exploring novel applications of game theory in ensemble learning:

Developed Banzhaf Power Index-based feature selection for ensemble methods
Implemented Borda count voting mechanism for classifier aggregation
Created Banzhaf Random Forests with strategic classifier selection
Published research on game theoretic approaches in ensemble learning

View Game Theory Research on GitHub

Image-to-Image Search Engine (Jan 2018 - Jul 2018)

Research under Dr. Tirtharaj Dash developing a novel reverse image search system:

Built a visual image search system using image captioning, ElasticSearch, and TensorFlow
Implemented LSTM-based caption generation for semantic image understanding
Published research paper in SocPros 2017 on captioning-based search engine
Created RESTful APIs for image caption generation and search functionality

View Image Search Project on GitHub

Education

BITS Pilani — M.Sc Math & B.E CS

Featured Projects

LLMs, RAG, AutoML, Search, Optimization, and more

DeydooGPT

AI Assistant built on fine-tuned Mistral 7B with 8000-token context handling for enhanced understanding of complex conversations.

LLM Fine-tuning Python

View DeydooGPT on GitHub

Clinical Diagnosis Engine

Medical diagnosis system using Graph RAG techniques, achieving 85.83% production accuracy for clinical applications.

Graph RAG Healthcare AI Python

Proprietary Project

Scalable BERT Inference Pipeline

High-performance inference system that reduced latency by 50% using ZeroMQ & NGINX for efficient handling of NLP requests.

Optimization NLP Python

View on GitHub

Image-to-Image Search

A reverse image search engine powered by Elasticsearch and TensorFlow for advanced image recognition and retrieval.

Computer Vision Search Python

View on GitHub

EazyML

Core technical member for this explainable AutoML platform for tabular data that automates model selection, hyperparameter tuning, and provides interpretable results.

AutoML Explainable AI Python

Proprietary Project

Keyflix

Educational platform with video content and intelligent recommendation system to personalize learning experiences.

Education Recommendations JavaScript

View on GitHub

Newsboat Recommender

RSS feed recommender using hybrid approach combining lexical and BERT-based content analysis for personalized news delivery.

NLP Recommendations Python

View on GitHub

FDE-Solver

MATLAB solver for Fractional Differential Equations with applications in mathematical modeling and simulations.

Mathematics Numerical Methods MATLAB

View on GitHub

Zero-shot Concept Matcher

System that uses BERT and mathematical distance metrics to match concepts without prior training, boosting accuracy by 35%.

NLP Zero-shot Learning Python

Proprietary Project

YouTube Video Assistant

A production-ready Python tool that enables natural conversations with YouTube videos using ASR, NLP, and LLMs. Features include automatic transcription, context-aware responses, and efficient caching.

ASR NLP LLM Python

View on Codeberg

YACR - Ultimate IT Roadmap

A comprehensive 9-10 month roadmap transforming beginners into job-ready IT professionals. Features structured learning paths, hands-on projects, and industry-aligned curriculum covering CS fundamentals to advanced AI.

Education Roadmap Career Development Open Source

View on Codeberg

DeydooRAG

A production-grade RAG pipeline using txtai for semantic search and LLM orchestration. Features include dynamic index management, query classification, and lightning-fast retrievals (3s search, <2s retrieval). Includes real-world examples with multilingual support.

RAG Semantic Search LLM Python

View on Codeberg

Simple Chatbot

A chatbot application built using Python and AIML (Artificial Intelligence Markup Language) for natural language processing and pattern matching.

Python AIML NLP

View Simple Chatbot on GitHub

Game of Thrones Visualization

Interactive visualization of the Game of Thrones dataset using t-SNE dimensionality reduction, featured in a YouTube video for its innovative approach to data visualization.

Python t-SNE Data Visualization Kaggle

View Game of Thrones Visualization on GitHub

DAPS (Dynamic Adaptive Prime Sampling)

A novel optimization algorithm that uses prime number-based grid sampling to avoid aliasing problems common in regular grid search methods. Features dynamic resolution adaptation and domain shrinking around promising regions, with primes serving as resolution knobs for accuracy control.

Mathematics Optimization Python Grid Search

View on GitHub

QuHabiton

A creative experimental habit tracker leveraging quantum physics, topology, and data analysis principles. Features include probabilistic habit representation, quantum machine learning for pattern analysis, behavioral topography visualization, and adaptive learning based on habit topology.

Quantum Computing Habit Tracking Topology Python

View on GitHub

List Coloring via Grover's Algorithm

Implementation of list coloring problem using Grover's algorithm through SAT formulation, developed during IIT Delhi's Quantum Machine Learning course. Demonstrates quantum computing applications in graph theory and constraint satisfaction problems.

Quantum Computing Graph Theory SAT Python

View on GitHub

LLM Episodic Memory System (Experimental)

A cognitive architecture for LLMs featuring surprise-driven memory formation, Legendre polynomial-based orthogonal paragraph embeddings, and dynamic memory refinement. Implements temporal decay, access boosts, and contradiction detection for biologically-inspired memory management.

LLM Memory Systems Cognitive Architecture Python

View on GitHub

ShunyaBar Labs

Physics-inspired constraint solvers. SAT as thermodynamic equilibration.

Deep-Tech Research Lab

Unapologetically ambitious, practically useful, Foundational Deep-Tech research lab from India.

Computation is physics, and constraint solving is thermodynamic equilibration. ShunyaBar Labs develops constraint solvers that treat Boolean satisfiability and combinatorial optimization as continuous physical systems rather than discrete search problems.

Visit shunyabar.foo

NitroSAT

The SAT solver that removes the search tree. Continuous gradient flow replaces backtracking. Prime-weighted hyperbolic geometry breaks symmetry traps. Linear O(M) time on industrial instances.

SAT Solver Prime Weights C + LuaJIT

View NitroSAT

BAHA

An optimizer that detects thermodynamic phase transitions. When algorithms freeze, BAHA computes the exact escape path using Lambert-W branch enumeration. Deterministic basin-hopping replaces random restarts.

Optimization Thermodynamics C++

View BAHA View Casimir SAT Solver

Multiplicative PINNs

Constraint learning without gradient conflict. Euler product gates align physics and data losses. Solves the gradient pathology problem, dropping conflicts from 45% to just 2%.

Physics-Informed NNs CFD Python

View Multiplicative PINNs

Navokoj API

Enterprise constraint intelligence platform. GPU-accelerated API with six engines (Nano, Mini, Pro, Hybrid, Q-State, Zeta). Includes DEFEKT diagnostics for pre-solve analysis and solvability prediction.

Enterprise GPU-accelerated API

View Navokoj

Publications (Zenodo)

NitroSAT: A Physics-Informed MaxSAT Solver Using Heat Kernel Diffusion, Persistent Homology, and Branch-Aware Holonomy Annealing
Sethurathienam Iyer (2026)
Multiplicative Calculus for Hardness Detection and Branch-Aware Optimization
Sethurathienam Iyer (2026)
ShunyaBar: Spectral–Arithmetic Phase Transitions for Combinatorial Optimization
Sethurathienam Iyer (2025)
Spectral-Multiplicative Optimization Framework
Sethurathienam Iyer (2025)
Solving SAT with Quantum Vacuum Dynamics
Sethurathienam Iyer (2025)

$ cat /etc/profile

About Me

Background

Research Interests

Current Focus

Active Research Areas

Skills & Expertise

AI/ML & LLMs

Programming & Development

ML Frameworks & Tools

Big Data & Databases

Work Experience

Workato — Senior AI Research Engineer May 2025 – Present

Deep Learning Researcher & Data Scientist Feb 2022 - Apr 2025

Research and Development Engineer Sep 2018 - Feb 2022

Research Intern Jul 2018 - Sep 2018

Student Researcher Jul 2017 - Jul 2018

Game Theory in Machine Learning (Jul 2017 - Dec 2017)

Image-to-Image Search Engine (Jan 2018 - Jul 2018)

Education Expand

Birla Institute of Technology and Science, Pilani 2013 - 2018

Certifications

Publications

Featured Projects Expand

DeydooGPT

Clinical Diagnosis Engine

Scalable BERT Inference Pipeline

Image-to-Image Search

EazyML

Keyflix

Newsboat Recommender

FDE-Solver

Zero-shot Concept Matcher

YouTube Video Assistant

YACR - Ultimate IT Roadmap

DeydooRAG

Simple Chatbot

Game of Thrones Visualization

DAPS (Dynamic Adaptive Prime Sampling)

QuHabiton

List Coloring via Grover's Algorithm

LLM Episodic Memory System (Experimental)

ShunyaBar Labs Expand

Deep-Tech Research Lab

NitroSAT

BAHA

Multiplicative PINNs

Navokoj API

Publications (Zenodo)

Education

Featured Projects

ShunyaBar Labs