Selected Projects
Enhancing Long‑term memory in Agents by Learning during Inference
Supervisor: Prof. Daniel Fried and Prof. Graham Neubig, CMU, LTI, Sep 2024 - Ongoing
Improving Web‑Agents during inference with auto‑eval(feedback) to deduce granular sub‑trajectories and induce memory.
Evaluating Robust use of information in Long Context
Supervisor: Prof. Graham Neubig, CMU, LTI, Jan 2024 - June 2024
Analyzing robustness of diverse language models to use relevant information in long‑context irrespective of its position by analysing effect of positional encodings, attention, architecture.
Empirically demonstrated that most long‑context adaptation methods(e.g: RoPE, attention based retrieval) are still susceptible to the lost‑in‑the‑middle problem
Also studied this in the context of imminent tasks ‑ RepoEval (repository Level code generation) showing significant drop in function calling performance as context length increases
Mitigating Hallucinations in Language Models
Supervisor: Prof. Maarten Sap, CMU, LTI, Sep 2023 - June 2024
Improving factual correctness in summarization models to mitigate intrinsic hallucination with automatic feedback.
Developing fine‑grained reward models to capture factual contradictions and unsupported spans in the generated summary
Offensive Tweet Detection
Supervisor: Prof. Swati Agarwal , BITS Pilani, K K Birla Goa Campus, Jun 2021 - Dec 2021.
This project aims to detect hate speech, trolling, and offensive language in tweets using NLP classification models. Built a Federated learning setup to jointly learn from diverse code-switched, code-mixed Hindi-English tweet datasets.
Stock Price Volatility Prediction post Earnings calls release
Code, Report , Topics: Multimodal Deep Learning, Time series
Predicted stock price volatility following earnings calls release on the text-audio paired MAEC dataset. Used SentenceBERT to encode text document and audio features such as pitch & intensity. Built text and audio encoder using BiLSTMs. Tested CNN and LSTM layer over fused representation of audio and text encoding.
Analysing and Predicting the outcome of a Volleyball Match
Report, Topics: Machine Learning, Data scraping
Built a web scraper to extract NCAA Volleyball match data. Generated features using rolling window averages for past matches. Designed additional team-wise and player-wise features after analysing game-play and literature. Compared the performance of varying models using player-wise and team-wise aggregates.
Modelling The Cognitive Functions of Consumer Behaviour
Code, Report, Supervisor: Prof. Bharat Deshpande, BITS Pilani, K K Birla Goa Campus, Jan 2020 - June 2020. Topics: Brain wave technology, Machine Learning
Aim: To predict consumer behaviour using brain wave measurement technology. Processed the collected brain data-EEG signals using MATLAB EEGlab toolkit. Developed ML classification models on this data to predict consumer behaviour.