KK.
AI / NLP / SOFTWAREMUMBAI, INDIA
ACL 2026 MAIN CONFERENCE CO-AUTHOR

Kalmit
Kulkarni

Software Engineer · AI & NLP Researcher · Full Stack Developer

I build AI-powered applications and study how reliably language technologies work across Indian languages. Connecting research questions with software you can use.

Explore my work Read my research
Kalmit Kulkarni in a navy suit
BASED IN MUMBAI, INDIA
Research, published.ACL 2026 · Main Conference
Language, understood.Six Indian languages in ITEM
Ideas, implemented.NLP experiments & AI applications
01 / ABOUT

From asking better questions
to building useful things.

Kalmit Kulkarni in a navy suit
MUMBAI, INDIA · AI & NLP

I’m a software engineer and NLP researcher based in Mumbai, working at the intersection of language, evaluation and practical software.

I graduated with a BTech in Computer Science Engineering from Vellore Institute of Technology. From August 2024 to July 2026, I worked remotely with Dr. Fajri Koto at KOTO LABS, MBZUAI, exploring how automatic NLP metrics align with human judgments across Indian languages.

That curiosity runs through my work: translation and summarization experiments, LLM applications and full stack interfaces. I now bring those interests to my web development internship at Save Engineering Works.

MY FOCUS

Multilingual NLP · Evaluation · Applied AI · Full stack development

02 / RESEARCH

How do we know
when language AI works?

Beyond a single score, toward a better understanding of quality.

ACL
2026MAIN CONFERENCE
LONG PAPERPUBLISHED JULY 2026

MULTILINGUAL NLP / EVALUATION

Revisiting Metric Reliability for Fine-grained Evaluation of Machine Translation and Summarization in Indian Languages

Amir Hossein Yari · Kalmit Kulkarni · Ahmad Raza Khan · Fajri Koto

Our team introduced ITEM — Indian Text Evaluation Metrics Testbed to examine how automatic evaluation agrees with human judgments for translation and summarization across Hindi, Bengali, Marathi, Gujarati, Tamil and Telugu.

01

LLM evaluators aligned most strongly with human judgments in the study.

02

Outliers changed metric–human agreement; reliability depended on the task and quality dimension.

03

Metrics responded differently to controlled changes in text.

My work: evaluation pipelines and correlation analysis. The benchmark and findings are the work of our co-author team.

Read paperDownload PDF
Publication details & citation

ACL 2026, Volume 1: Long Papers · pp. 25543–25561
DOI: 10.18653/v1/2026.acl-long.1171

Amir Hossein Yari, Kalmit Kulkarni, Ahmad Raza Khan, and Fajri Koto. 2026. Revisiting Metric Reliability for Fine-grained Evaluation of Machine Translation and Summarization in Indian Languages. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 25543–25561, San Diego, California, United States. Association for Computational Linguistics. https://doi.org/10.18653/v1/2026.acl-long.1171

Read the preprint
03 / SELECTED WORK

Research questions.
Working software.

A selection of applications and experiments.
Built to explore, understand and make useful.

5 selected projects

AI APPLICATION / DOCUMENT INTELLIGENCE01
STATEMENT PDF→VALIDATE + STRUCTURE→INSIGHTS

Personal Finance Agent

From a bank statement to a useful financial view.

I built a multi-stage pipeline that extracts and validates transactions from statement PDFs, normalizes merchants and categorizes spending. Analytics connect that structured data to anomalies, goals and forecasts.

Implementation notes

A React and TypeScript interface connects to Flask, PostgreSQL and SQLAlchemy. Ollama supports local inference with policy guardrails; privacy is a design goal.

PythonReact / TypeScriptPostgreSQLOllama
Explore source
AI APPLICATION / MULTI-MODEL WORKFLOWS02
GENERATE→PEER REVIEW→SYNTHESIZE

Ollama Council

Let models generate, review and synthesize.

A council gathers responses from multiple local models, asks them to rank anonymized answers, and aggregates the rankings. A chairman model then synthesizes a final response.

Implementation notes

An exploration of model orchestration using Python, JavaScript and Ollama. Implementation and any upstream attribution are available in the repository.

PythonJavaScriptOllama
Explore source
NLP / RELATED EXPERIMENTS03

Multilingual Summarization

Different models. A shared question: what survives a summary?

I explored Indian news summarization with XLSum, transformer fine-tuning and LLM inference. Related experiments use IndicBART, GPT-4o Mini and Cohere Command R+, with language coverage varying by model and notebook.

Implementation notes

The collection includes English, Hindi, Gujarati, Bengali, Telugu and Tamil. Training and inference work used Colab and a T4 GPU.

PyTorchHugging FaceIndicBARTXLSum
IndicBARTGPT-4o MiniCommand R+
NLP / TRANSLATION04

Machine Translation & Evaluation

Translate across languages. Examine how the scores behave.

I built sampling, translation and evaluation experiments with FLORES-200, IndicTRANS2 and LLMs. The collection spans English and six Indian languages; individual variants cover different language directions.

Implementation notes

Reference-based BLEU with smoothing is distinct from LLM judgments. Pandas and exported results support further analysis. FLORES-200 is an existing dataset credited to its creators.

IndicTRANS2FLORES-200NLTK / BLEUPandas
IndicTRANS2Command R+GPT experiments
COMPUTER VISION / EXPERIMENTAL ML05

Alzheimer’s Image Classification

Exploring MRI classification through transfer learning.

An experimental study of Alzheimer’s-related MRI images, including ResNet50 and InceptionV3, augmentation, class balancing, classification reports and confusion matrices.

Implementation notes

TorchCam attention visualizations help inspect model behavior. This work is a machine learning experiment, not a clinically validated diagnostic tool.

PyTorchTensorFlow / KerasTorchCamOpenCV
Explore sourceOpen notebook
Go a little deeper Explore the complete repository collection

Applications, related experiments, supporting resources, coursework and templates.

25 repositories

View all repositories on GitHub
04 / EXPERIENCE

Learning by doing.
Building with others.

AUG 2026 — PRESENTCURRENT

Web Developer Intern

Save Engineering Works Pvt. Ltd. · Goregaon, Mumbai

I work on responsive interfaces, frontend functionality and web application component integration, with attention to usability, maintainability and performance.

ReactJavaScriptNode / ExpressREST APIsSQLGit
AUG 2024 — JUL 2026

NLP Researcher

KOTO LABS, MBZUAI · Remote from Mumbai

Under Dr. Fajri Koto, I implemented translation and summarization evaluation pipelines and analyzed correlations between human judgments and automatic metrics. This work connects to our ACL 2026 paper.

Tools studied included BLEU, ROUGE, COMET, GPTScore and G-Eval.

Research group
AUG — DEC 2023

MERN Stack Intern

Ethnus · Remote from Mumbai

I contributed to Mewsic, a music streaming application, working on interface design and responsiveness with React, Figma, MongoDB, Node, Express, Bootstrap and Spotify APIs.

Verify credentialID: HRTWN6JS
05 / TOOLKIT

The tools behind the work.

Connected to projects and experience.
Always room to learn more.

Software & web

Python · JavaScript · TypeScript · SQL
React · Node · Express · Flask
REST APIs · responsive interfaces
PostgreSQL · MongoDB · SQLAlchemy

Language & AI

Hugging Face · IndicBART · IndicTRANS2
PyTorch · TensorFlow · Keras
OpenAI API · Cohere API · Ollama
Translation · summarization · evaluation

Analysis & workflow

Pandas · NumPy · SciPy
Matplotlib · Seaborn · Scikit-learn
Git · GitHub · Jupyter · Colab
OpenCV · TorchVision · TorchCam

More tools & evaluation methods

C · C++ · Java · HTML · CSS · Bootstrap · Figma · Excel · imbalanced-learn

COMET · BERTScore · BLEU · ROUGE · METEOR · chrF · MoverScore · BLEURT · BARTScore · GPTScore · G-Eval

CURRENTLY LEARNING Next.js · Tailwind CSS · Docker · AWS fundamentals · CI/CD · modern full stack practices

06 / FOUNDATIONS

Education & continued learning.

SEP 2021 — OCT 2025

BTech, Computer Science Engineering

Vellore Institute of Technology

Academic detail

GPA: 7.15/10

AUG 2019 — JUN 2021

Higher Secondary Certificate

Nirmala Memorial Foundation College
of Commerce and Science

Academic detail

Score: 82.17%

Certifications

Introduction to Self-driving Cars

Coursera · January 2023 · Grade 99.60%

Computer Vision

Vityarthi · September 2023

Google Cloud Computing Foundation Program

January 2024 · Public learning profile

NPTEL Cloud Computing

April 2024

07 / GET IN TOUCH

Good work starts
with a conversation.

Interested in my work or a potential collaboration?
Let’s talk about AI, NLP and software development.