Machine learning & NLP researcher · San Francisco

Hi there, I'm Rishabh Sanjay.

I completed my Masters in Computer Science from Purdue University in 2025, and my Bachelors in Mathematics and Scientific Computing with a minor in Machine Learning from IIT Kanpur in 2021. Currently I am working as a Founding AI Engineer at Cekura in San Francisco. Apart from engineering, I have worked on multiple AI research projects; my research interests include Machine Learning, Deep Learning, Natural Language Processing, Large Language Models and Legal AI.

Download resume Email me LinkedIn GitHub

San Francisco, California
MS CS, Purdue '25 · BS, IIT Kanpur '21
Cekura · Oracle · GSoC @ ArviZ · Goldman Sachs

Open to conversations

About

A bit about me.

Deep nets always drove me crazy. I was always fascinated by their sophisticated structure and loved how they help us attain state-of-the-art results for so many real world problems, which is why most of my work sits at the intersection of machine learning research and engineering.

At Purdue I worked with Prof. Abulhair Saparov on the information capacity of transformers, studying how much factual knowledge a transformer can store and how that scales with model size; the paper is currently under review at ICLR. Previously I worked on court judgment prediction (ACL 2021) and semantic segmentation of legal documents (EMNLP 2022) under Prof. Ashutosh Modi and Prof. Arnab Bhattacharya at IIT Kanpur. I have also contributed to the open-source library ArviZ as part of GSoC'21 and worked as a Software Engineer at Oracle. Apart from these I am also trying to extend my deep learning knowledge and not just restrict it to NLP.

In my spare time I like solving puzzles, binging anime and playing multiplayer games. I also love playing cricket, table tennis and badminton, and competing on Codeforces, where I hold a 1800+ rating.

Experience

Where I've worked.

Full resume

Founding AI Engineer

Cekura · San Francisco, CA

Jul 2025 — Present

Cekura (YC F24) builds testing and observability for voice and chat AI agents. I work on scenario generation, simulation pipelines and evaluation metrics.
- Architected an agentic scenario-generation system that turns an agent's prompt, tools and knowledge base into realistic test conversations covering tool-call flows, edge cases and KB-grounded Q&A, lifting scenario accuracy by 20–25%.
- Engineered a feedback-driven conversational refinement agent that improves scenarios through structured tool-calling and LLM reasoning, so customers iterate on tests in minutes instead of hours.
- Optimized the simulation pipeline for a 15% accuracy gain and a 50% latency reduction, and shipped an AI-powered end-of-call algorithm that cut premature or incorrect hang-ups to under 1% in production.
- Built agent-level evaluation metrics for intent alignment, response relevance, conversational consistency and interruption handling, then automated metric creation and doubled the speed of metric optimization.
- Developed customizable smart alerts that detect agent failures against client-defined metrics, enabling rapid diagnosis and recovery, plus red-teaming flows that probe what an agent should never say.
- Scaled production voice-AI systems that stress-tested 5M+ agent minutes, contributing to a 4× reduction in customer churn and 4× ARR growth in 8 months.
- Python
- LLMs
- OpenAI
- Gemini
- Vapi
- Deepgram
- Claude Code
Software Engineer, Fusion HCM Development

Oracle · Hyderabad, India

Jul 2021 — Sep 2022

Full-stack development on Oracle Journeys, the guided-workflow product inside Oracle Fusion HCM.
- Engineered the UI and Java logic for the Advanced Task and Survey Task features with full CRUD capabilities.
- Developed RESTful web services for Oracle Journeys with a focus on asynchronous operations and regression-tested stability.
- Tested and resolved customer bugs and configuration issues in sandbox environments across 5 releases.
- Java
- Oracle ADF
- REST
- SQL
- HTML
Open-source Developer

Google Summer of Code · ArviZ (NumFOCUS) · Remote

May 2021 — Aug 2021

Added new Bayesian visualization plots to ArviZ, the exploratory analysis library for Bayesian models.
- Implemented Dot plots, ECDF plots and Violin plots with both Matplotlib and Bokeh backends.
- Researched and implemented the underlying algorithms, wrote Pytest suites, and documented everything in a five-part blog series.
- Python
- Matplotlib
- Bokeh
- Pytest
Summer Analyst, IMD Core Engineering

Goldman Sachs · Bangalore, India

May 2020 — Jun 2020

Built a horizontally scalable, distributed reconciliation system for trading data.
- Designed an event-driven architecture for concurrent data processing, cutting reconciliation time by 70%.
- Wrote MongoDB and Sybase IQ queries for reconciliation and surfaced data inconsistencies.
- Java
- Akka
- MongoDB
- Sybase IQ
- JUnit

Research

Questions I've chased.

From how transformers store facts to how machines read the law.

Publications

Under review at ICLR Purdue University · Prof. Abulhair Saparov · Sep 2024 — Jun 2025

Information capacity of transformers

Empirical scaling laws for transformer memorization capacity: how the smallest model that can perfectly memorize a set of facts grows with the number of facts, along both width and depth.

Trained GPT-2-style transformers on synthetic knowledge-graph triples with no exploitable structure, so that high accuracy is only possible through memorization, and measured the minimum hidden dimension and the maximum memorizable dataset size across a wide range of dataset sizes.
Ran rigorous hyperparameter search with Bayesian optimization and deterministic full-batch training to make sure each model reached its true capacity, then compared candidate scaling forms with statistical lack-of-fit tests.
Extended the analysis to structured data (symmetric and two-hop relations), skewed power-law fact distributions and a second architecture family to test generality, and identified the minimal viable dimension for resource-efficient training at a given data size.

PyTorch
GPT-2
Scaling laws
Bayesian optimization

Published at NLLP @ EMNLP 2022 IIT Kanpur · Prof. Ashutosh Modi & Prof. Arnab Bhattacharya · Apr 2021 — Aug 2021

Semantic segmentation of legal documents via rhetorical roles

A new corpus of 100 Indian legal judgments annotated with rhetorical roles, the largest such corpus, and a model that segments documents into those roles.

Proposed MTL-BiLSTM-CRF with BERT sentence encodings and label shift as an auxiliary task, reaching an F1 of 0.71 and beating sequence-classification baselines.
Showed that rhetorical-role structure improves downstream judgment prediction.

BERT
BiLSTM-CRF
Multi-task learning
Legal NLP

Published at ACL 2021 IIT Kanpur · Prof. Ashutosh Modi & Prof. Arnab Bhattacharya · Mar 2020 — Jun 2020

Court judgment prediction and explanation (ILDC for CJPE)

Built a corpus of 32,000 Indian Supreme Court judgments and tackled the open problem of predicting and explaining court decisions.

Benchmarked 14 state-of-the-art document classification methods; the best hierarchical XLNet model reached 78% accuracy.
Compared explainability methods (Integrated Gradients, occlusion, attention) against gold explanations annotated by legal experts using ROUGE, BLEU and METEOR.

XLNet
Explainable AI
Legal NLP

Publications

Peer-reviewed work.

Google Scholar

2022
Semantic Segmentation of Legal Documents via Rhetorical Roles

Vijit Malik, Rishabh Sanjay, Shouvik Kumar Guha, Shubham Kumar Nigam, Angshuman Hazarika, Arnab Bhattacharya, Ashutosh Modi

Natural Legal Language Processing Workshop (NLLP) at EMNLP 2022
- Paper
- Code
2021
ILDC for CJPE: Indian Legal Documents Corpus for Court Judgment Prediction and Explanation

Vijit Malik, Rishabh Sanjay, Shubham Kumar Nigam, Kripabandhu Ghosh, Shouvik Kumar Guha, Arnab Bhattacharya, Ashutosh Modi

ACL-IJCNLP 2021 (Long Papers)
- Paper
- arXiv
- Code
2026

Information capacity of transformers: how memorization scales with width and depth

with Prof. Abulhair Saparov and collaborators, Purdue University

Under double-blind review at ICLR. Details to follow once the review completes.
Under review

Projects

Things I've built.

Coursework and side projects from Purdue and IIT Kanpur, from distributed systems to cryptanalysis.

GitHub

Cloud e-commerce platform

Purdue University · Sep – Dec 2023

Scalable storefront with login, order and inventory REST APIs, deployed on EC2 behind a load balancer and auto-scaler.

React
Spring Boot
AWS
DynamoDB
Docker

Open-domain QA on WikiHow FAQs

Research internship, The Ohio State University (Prof. Huan Sun) · May 2022 – 2023

Studied continued pre-training on a WikiHow FAQ corpus for question-type classification, re-ranking, reading comprehension and intent recognition, and used T5 and BLOOM for closed-book QA.

RoBERTa
T5
BLOOM
Question answering

Neural machine translation: Hindi → English

CS779, IIT Kanpur · Jan – Mar 2021

Two-layer bidirectional GRU encoder with an attention decoder, trained with teacher forcing and decoded with beam search.

PyTorch
Seq2Seq
Attention

Detecting and rating humor (SemEval-2021 HaHackathon)

CS771, IIT Kanpur · Sep – Dec 2020

BERT + BiGRU humor classifier with F1 0.94 and a BERT + MLP regressor for humor ratings with RMSE 0.446.

BERT
BiGRU
Classification

Mining Indian Supreme Court & High Court judgments

CS685, IIT Kanpur · Sep – Dec 2020

Scraped 300K cases, clustered them by IPC topics and combined them with demographic data to compare crime rates across High Courts.

Scraping
k-means
Data mining

Breaking cryptosystems

CS641, IIT Kanpur · Jan – Apr 2020

Broke substitution, substitution-permutation, DES, SASAS and weakened RSA and KECCAK ciphers.

Cryptanalysis

Taxi fare forecasting

Time Series Analysis, IIT Kanpur · Oct – Nov 2019

De-trended time-series data and fitted ARIMA models to forecast taxi fares.

ARIMA
Time series

Low-rank matrix approximation algorithms

Mentor: Prof. Sumit Ganguly, IIT Kanpur · May – Jun 2019

Implemented length-squared sampling for matrix multiplication and the CUR method for matrix sketching.

Randomized algorithms
Linear algebra

Reinforcement learning

Stamatics Club, IIT Kanpur · Nov – Dec 2018

Solved Gym's Frozen Lake with value iteration and Mountain Car with MDP and Q-learning techniques.

Q-learning
MDPs

Writing

Recent writing.

A few recent pieces from the Cekura blog. Everything else, including my Google Summer of Code series, is on the blog page.

All writing

Skills

Toolbox.

Voice & LLM systems

LLM agents & tool calling
Scenario generation
Simulation pipelines
Evaluation metrics
Red-teaming
RAG
LLM fine-tuning
Prompt optimization
OpenAI
Gemini
Vapi
Deepgram
LangChain
Claude Code

Machine learning

PyTorch
TensorFlow
HuggingFace
Transformers
NLP
Scaling laws
Explainable AI
GNNs
Diffusion models
NumPy
Pandas
scikit-learn

Languages

Python
Java
C / C++
SQL
Bash
MATLAB

Infrastructure

AWS
GCP
Docker
Kubernetes
Kafka
MongoDB
CUDA
MPI
Git
Linux
Pytest

Education

Where I studied.

Purdue University

Master of Science in Computer Science

Aug 2023 – May 2025 · West Lafayette, IN

GPA 3.9 / 4

Natural Language Processing
Deep Learning
Large Language Models
Reasoning in LLMs
LLM Alignment
Statistical Machine Learning
Data Mining
Compilers & Programming Systems
Parallel Computing

Indian Institute of Technology Kanpur

Bachelor of Science in Mathematics and Scientific Computing

Jul 2017 – May 2021 · Kanpur, India

GPA 8.5 / 10 · Minor in Machine Learning & Applications (10 / 10)

Introduction to Machine Learning
Probabilistic Inference
Data Structures & Algorithms
Linear Algebra
Probability & Statistics
Real Analysis
Time Series Analysis
Cryptography

Contact

Let's talk about research and machine learning.

I'm always happy to chat about machine learning, NLP and LLM research, or anything else on this page. The fastest way to reach me is email.

rishabh.lfs@gmail.com Connect on LinkedIn Resume (PDF)