I'm an ML Researcher at Huzzle Labs, where I work on GUI agents for computer, mobile, and web applications, improving how vision-language models ground, plan, and act across complex tasks, including long-horizon subtasks. We aim to train small, data-efficient models (VLMs/LLMs under 4B parameters) using distillation, signal-rich data preparation, and parameter-efficient training, with 100× less data than frontier models.
Before that, I was an Applied ML Researcher at BharatGen, building large language models from scratch for Indian contexts. I led the post-training pipelines for the Param model family, covering supervised fine-tuning, reward modeling, preference optimization (DPO/GRPO), and distillation. I also designed multilingual instruction-following and safety-alignment datasets for governance, legal, education, and finance use cases, and led multi-stage pre-training data collection for India-centric LLMs.
My Ph.D. at CFILT, IIT Bombay, advised by Prof. Pushpak Bhattacharyya, focused on social bias, fairness, and safety in multilingual NLP. I built benchmarks such as IndiBias (NAACL 2024), BharatBBQ (TACL 2025), and IndicCONAN (AAAI 2024), and methods to detect and mitigate bias in language and text-to-image models. I was a Prime Minister's Research Fellow and received IIT Bombay's George B. Fernandes (2024–25) and Vijay Vashee (2023–24) awards for excellence in Ph.D. research. Earlier, I did my M.Tech at IISc Bengaluru, worked on low-light image enhancement at the Video Analytics Lab (BMVC 2021 Oral, Best Student Paper runner-up), and was a software engineer at Citrix.
🔥 News
- 2026.03 🚀 Joined Huzzle Labs as an ML Researcher, working on GUI agents and small, data-efficient VLMs/LLMs.
- 2025.11 📄 BharatBBQ (TACL) and Open-DeBias (Findings) at EMNLP 2025, Suzhou.
- 2025.07 📄 BIStereo, a benchmark for body-image stereotypes in LMs, at ACL 2025 (Findings), Vienna.
- 2025.07 🧠 Released the PARAM-1 2.9B technical report with the BharatGen team.
- 2025.06 💼 Joined BharatGen as an Applied ML Researcher, leading post-training for the Param model family.
- 2025.01 📄 Dialect-robustness paper at the SUMEval workshop, COLING 2025.
- 2024.06 📄 IndiBias at NAACL 2024, supported by a Microsoft Research Travel Grant.
- 2024.05 🎤 Tutorial on Bias and Hallucination in LLMs at LREC-COLING 2024.
- 2024.02 📄 IndicCONAN at AAAI 2024.
- 2023.12 🎤 Two tutorials at ICON 2023: vision-language models, and prompting LLMs.
- 2023.05 🧪 Research internship at Adobe Research India on debiasing text-to-image diffusion models.
💼 Experience
-
ML Researcher, Huzzle Labs · Mar 2026 – Present
Building GUI agents for computer, mobile, and web applications: improving VLM grounding, action prediction, and planning for complex tasks, including long-horizon subtasks. Building large-scale datasets for GUI tasks that need fewer agentic interactions, and training small VLMs/LLMs (under 4B parameters) with 100× less data than frontier models via distillation, signal-rich data preparation, and parameter-efficient training.
-
Applied ML Researcher, BharatGen, Mumbai · Jun 2025 – Mar 2026
Contributed to an LLM built from scratch for Indian contexts with multilingual capabilities across major Indian languages. Led post-training for the Param model family (SFT, reward modeling, preference optimization with DPO/GRPO, and model distillation); designed multilingual instruction-following and safety-aligned datasets reflecting cultural sensitivities and domain needs in governance, legal, education, and finance; led multi-stage pre-training data collection and creation.
-
Research Intern, Adobe Research India, Bengaluru · May 2023 – Jul 2023
Debiased text-to-image diffusion models by adjusting the input prompt and cross-attention to produce images with greater, more equitable diversity across gender and skin tone, entirely at inference time with no additional training.
-
Teaching Assistant & Course Instructor, IIT Bombay · Aug 2021 – Apr 2025
TA for AI & ML, NLP, Deep Learning for NLP, and (Advanced) Digital Image Processing; instructor for TA101. Twice received the Excellence in TA Award. Details ↓
-
Research Assistant, Video Analytics Lab, CDS, IISc Bengaluru · Feb 2020 – Dec 2020
Low-light RAW image enhancement with few-shot domain adaptation: learning from abundant paired data on a source camera and only a handful of paired images on a target camera. Published at BMVC 2021 (Oral, Best Student Paper runner-up).
-
Software Engineer 2, Citrix R&D India, Bengaluru · Jul 2019 – Feb 2020
Front-end development of Citrix Workspace for macOS, and migration of its codebase from Objective-C to Swift.
📝 Publications
* denotes equal contribution. Full list on Google Scholar.
Conference & Journal Papers
EMNLP 2025 FindingsOpen-DeBias: Toward Mitigating Open-Set Bias in Language Models
- Introduces OpenBiasBench for open-set bias in QA and Open-DeBias, an adapter-based method that beats BMBI on BBQ by nearly 48% (ambiguous) and 6% (disambiguated), with 84% zero-shot accuracy on Korean BBQ.
TACL 2025BharatBBQ: A Multilingual Bias Benchmark for Question Answering in the Indian Context
- A culturally adapted QA bias benchmark in 8 Indian languages covering 13 social categories (3 intersectional), expanding 49,108 examples to 392,864; evaluates five multilingual LM families. Presented at EMNLP 2025.
ACL 2025 Findings“You are Beautiful, Body Image Stereotypes are Ugly!” BIStereo: A Benchmark to Measure Body Image Stereotypes in Language Models
- Probes body-image stereotypes with 40k sentence pairs, 60k premise–hypothesis pairs, and 553 validated tuples scored with the TriSentBias metric; finds significant biases in MuRIL, XLM-R, Llama 3, and Gemma among ten LMs.
SUMEval @ COLING 2025Evaluating Dialect Robustness of Language Models via Conversation Understanding
- Introduces M-MD3, taboo-game conversations in US and Indian English with the target word masked; Llama 3, GPT-4, and GPT-3.5 perform significantly better on US English than on Indian English.
NAACL 2024IndiBias: A Benchmark Dataset to Measure Social Biases in Language Models for Indian Context
- An Indian-context social bias benchmark in English and Hindi, with 800 sentence pairs and 300 tuples across seven bias dimensions plus three intersectional ones, used to compare ten language models.
AAAI 2024IndicCONAN: A Multilingual Dataset for Combating Hate Speech in Indian Context
- A human-in-the-loop counter-narrative dataset against hate speech, with 2,500+ examples each in English and Hindi (AI for Social Impact track).
ACL 2023 FindingsWith Prejudice to None: A Few-Shot, Multilingual Transfer Learning Approach to Detect Social Bias in Low Resource Languages
- A new Hindi social-bias dataset of 9k posts annotated with bias, sentiment, target groups, and rationales; few-shot multilingual transfer from English, Italian, and Korean data reaches 80.8 macro-F1 with XLM-R.
CoNLL 2022Detecting Unintended Social Bias in Toxic Language Datasets
- ToxicBias: curated from the Jigsaw Unintended Bias data and annotated for five bias categories (gender, race/ethnicity, religion, political, LGBTQ), with baselines for bias identification, target generation, and implications.
LREC 2022Hollywood Identity Bias Dataset: A Context Oriented Bias Analysis of Movie Dialogues
- Movie-script dialogues annotated for identity bias across seven categories, with sensitivity, stereotype, sentiment, and emotion labels.
BMVC 2021 · Oral🏆 Best Student Paper Runner-up
Few-Shot Domain Adaptation for Low Light RAW Image Enhancement
- With ten or fewer labelled samples from a new camera, matches or beats training on a large labelled target dataset; also releases a new Nikon low-light RAW dataset.
Technical Reports, Preprints & Under Review
Technical Report 2025PARAM-1 BharatGen 2.9B Model
- A 2.9B decoder-only model trained from scratch on Hindi and English, allocating 25% of the corpus to Indic languages, with a tokenizer adapted to Indian morphology.
PreprintMILA (Multilingual Indic Language Archive): A Dataset for Equitable Multilingual LLMs
arXiv 2025Mathematics Isn’t Culture-Free: Probing Cultural Gaps via Entity and Scenario Perturbations
- Culturally adapted GSM8K variants for five regions (Africa, India, China, Korea, Japan); six LLMs (8B–72B) do best on the original US-centric set, while models with reasoning abilities are more resilient.
Under ReviewBeyond Single-Axis Fairness: Learning to Detect and Rewrite Intersectional Biases
Under ReviewFrom ‘No’ in English to ‘Yes’ in Hindi: Revealing LLM Security Gaps using Indian Languages
Under ReviewMomentum Meets Virality: A Novel Metric for Unmasking Social Bias in Viral Tweets
🎤 Tutorials
LREC-COLING 2024Addressing Bias and Hallucination in Large Language Models
Half-day tutorial · Torino, May 2024
ICON 2023Vision-Language Models: Evolution, Applications, and Challenges in Bridging the Gap Between Visual and Textual Data
Half-day tutorial · Goa, Dec 2023
ICON 2023Harnessing LLMs with Prompts: Applications, Challenges, and Maximizing Their Potential
Half-day tutorial · Goa, Dec 2023
ICON 2022Social Bias Detection and Mitigation in Text
Half-day tutorial · Dec 2022
📚 Teaching
- Artificial Intelligence and Machine Learning, IIT Bombay (TA) · Spring 2025, Fall 2021 · 🏅 Excellence in TA Award (both terms)
- Natural Language Processing (CS 626), IIT Bombay (TA) · Fall 2024, Fall 2022
- Advanced Digital Image Processing, IIT Bombay (TA) · Spring 2024
- Teaching Assistantship 101 (TA101), IIT Bombay (Course Instructor) · Spring 2024, Fall 2023
- Digital Image Processing, IIT Bombay (TA) · Fall 2023
- Deep Learning for Natural Language Processing (CS 772), IIT Bombay (TA) · Spring 2023, Spring 2022
- Deep Learning, NPTEL, online (TA) · Spring 2023, Fall 2022
🎖 Honors and Awards
- 2025 George B. Fernandes Award for Excellence in Ph.D. Research (2024–25), IIT Bombay
- 2025 Excellence in TA Award, IIT Bombay (also in 2021)
- 2024 Vijay Vashee Award for Excellence in Ph.D. Research (2023–24), IIT Bombay
- 2024 Microsoft Research Travel Grant, NAACL 2024
- 2023 Bronze Medal, 11th Inter IIT Tech Meet
- 2022 Prime Minister's Research Fellowship (PMRF)
- 2022 Silver Medal, 10th Inter IIT Tech Meet
- 2021 Best Student Paper Award (Runner-up), BMVC 2021
- 2018 Winner, Intel Campus Day, IISc Bengaluru
- 2017 All India Rank 9, GATE (Computer Science)
- 2017 Winner, CodeChef Certification in Data Structures & Algorithms (CCDSAP)
- 2015 First place, Hour of Code, Horizon, IGIT
- 2007 Rank 40, Mathematics Olympiad, Shikhya Vikash Samiti, Odisha
🤝 Service
- Program Committee / Reviewer: ACL Rolling Review (every cycle, 2024–2026), AAAI 2025; earlier EMNLP 2022 & 2023, AAAI 2023, ICON 2022, LREC-COLING 2024
- Panelist, Generation AI: Shaping Education for Flourishing (UNESCO), RIE Mysuru · 2023
- Placement Coordinator, IISc Bengaluru · 2018–19
- Event Coordinator, Data Science Hackathon, IISc Open Day · 2019; Hackathon, CSA Undergraduate Summer School · 2018
📖 Education
-
2021 – 2026, Ph.D., Computer Science and Engineering, Indian Institute of Technology Bombay
CGPA 9.12/10 · Advisor: Prof. Pushpak Bhattacharyya
Thesis: Decoding Social Bias in NLP: Multilingual Datasets, Benchmarks and Mitigation for Fairness and Safety in NLP -
2017 – 2019, M.Tech, Computer Science and Automation, Indian Institute of Science, Bengaluru
Advisor: Prof. Sridharan Devarajan
Thesis: Alzheimer's Disease: Understanding Mechanisms for Early Diagnosis and Treatment using Functional MRI of the Brain - 2012 – 2016, B.Tech, Computer Science and Engineering, Indira Gandhi Institute of Technology, Sarang, Odisha