Kanishk S. Email & Phone Number

Mountain View, CA, US

Mountain View, California, United States

New York, United States

• Implemented a contrastive learning framework for robust Named-Entity-Recognition (NER) on ASR-Transcripts to detect bonds, tickers, and other relevant entities of interest in trader-call transcripts.
• Designed methods enforcing ASR-Gold similarity, outperforming the production models by 8% in F1-score.
• The work will be presented at the internal largest Firm-wide JP Morgan-Chase AI-Summit in Brooklyn.

New York, United States

• (1) DVMM Lab, Columbia University ( Advisor: Dr. Shih Fu Chang, Dean Columbia Engineering [Nov 2022 - May 2023])
• Designed methods leveraging in-context learning to produce state and action descriptions in instructional videos.
• Deployed VQA models to produce entity-centric descriptions for states and actions in a single video.
• Zero-shot CLIP-classi cation using descriptions by LLMs outperform methods by 5% in state-action precision.(2) Smara's Lab, Columbia University (Advisor: Dr. Smaranda Muresan, Data Science Institute [Jan 2023 - June.
• Improvised image-editing using advances in segmentation(Segment-Anything), chain-of- thought(COT) - prompting and VQA(BLIP-v2) models.
• Fine-tuned Instruct-Pix2Pix diffusion model showed a 3% improvement over existing baselines in faithfulness scores and the paper is under review at the EMNLP'2023.

Kharagpur, West Bengal, India

• Finetuned DialoGPT and GPT-2 for generating efficient counter-narratives against online hate-speech.
• Constructed methods for controlling attributes such as politeness, toxicity in GPT-2 generated responses.
• Nominated for best thesis award among batch of 100 dual degree and masters students in CS Department and the work was published at IJCAI' 22 conference.

Bengaluru, Karnataka, India

• Explored Active-learning strategies to efficiently sample data for flagging unethical sponsored advertisements.
• Implemented Sampling Strategies such as entropy-based and Query-By-Committee for efficiently generating data.
• Teacher-Student transformer models incorporated with active learning increased precision value by 4% over existing baseline models.

Bengaluru, Karnataka, India

• Filed a patent for devising a DL-algorithm to seggregate authenticated clicks in user sessions.
• Designed algorithms leveraging topology to improvise the Positively Unlabeled(PU)-Learning framework.

Master'S Degree, Computer Science

Bachelors+Masters(Dual Degree), Computer Science, Cgpa - 9.25/10

Relevant Coursework : Natural Language Processing, Reinforcement Learning, Deep Learning, Social Computing, Machine Learning, AI and.