Nov 2022 – Jul 2025 · 2 yrs 8 mos · Gurugram, Haryana, India

• Part of R&D team of JD(Job Description) Parser. Responsible for new(both ML & non-ML) features incorporation and optimising components of JD Parser. Also, collaborated across multiple teams to develop new capabilities.
• Fine-tuned variants of BERT/RoBERTa based NER models(achieved F1 scores of 92%) for extracting critical entities like: degree, job title etc. with high accuracy from JDs. Deployed the models in AWS Sagemaker and reduced the JD parsing time by 45% using multithreading.
• Fine-tuned variants of BERT/RoBERTa based models to develop a central PFM(Position Function Mapping) model with accuracy of 95%. This improved the Knowledge Graph and the function & position level(Entry level, Mid level etc.) prediction across the core-elements
• of all the major products (Resume, JD and Interview transcripts ). Used SOTA methods to handle the issue of large no. of labels while fine-tuning.
• Fine-tuned variants of Phi-3/TinyLlama architectures to generate high-precision grammatical error correction feedbacks for products like Resume Builder, JD Author etc., improving user engagement. Surveyed various LLM research papers to use training optimisation techniques, resulting in reduction of time consumption during fine-tuning. Deployed the best model on AWS Sagemaker using ONNX for real-time serving.
• Engineered a 96% accurate JD parser using Gen AI, LangChain, and FastAPI for a key client. Increased speed and accuracy by implementing parallel parsing of different JD sections. Improved precision through com- prehensive LLM model testing. Fine-tuned and deployed the GLiNER model to accurately(95%) redact PII from JDs before LLM processing.
• Researched and developed a Knowledge Graph consisting of Functions, Job Position and Skills collected from various sources like Resumes and JDs to suggest next possible Career Transition Paths for the Candidates.

Apr 2021 – Jul 2021 · 3 mos · Bangalore Urban, Karnataka, India

• Developed API for extracting Closed Captions from Recipe videos on Youtube and recipes from various recipe web pages.
• Customized the BERT Large model for training it on millions of recipe lines extracted from Youtube as well as available on the company's database.
• After training, the model was finetuned over the BERT for Sequence Classification model for extracting Recipe lines from a large Youtube Recipe paragraph by removing unnecessary lines.
• Also finetuned the model over the BERT for Token Classification model for extracting Ingredient lines from the Recipe lines extracted above for obtaining the ingredients required for preparing the recipes.
• This product helped the company in extracting Recipe and Ingredient lines with almost 80% accuracy from new Closed Captions and thus leveraged discovering various new recipes.

Oct 2020 – Feb 2021 · 4 mos · Bangalore Urban, Karnataka, India

• Developed a Goal Based Wealth Management Software based on Markowitz Principle to provide most profitable portfolio to the user with the continuous change in both time and the invested capital.
• Developed various in-house technical indicators for equity trading.
• Analyzed past 10 years data of various NSE India listed Mid and Small cap stocks and using the multiple trade-determining factors obtained, trained a Deep Reinforcement Learning based stock trading model
• which came up with a 12% annual return in the rough year of 2020.

Jun 2020 – Dec 2020 · 6 mos · Kharagpur, West Bengal, India

• Worked on a research project on improving the performance and accuracy of the Task Oriented Dialogue(TOD) systems by incorporating the baseline encoder-decoder based Transformer architecture.
• Modified traditional transformers’ single decoder to the two stream joint belief-action predicting decoders to achieve great accuracy improvement.

Apr 2020 – Aug 2020 · 4 mos · Kharagpur, West Bengal, India

• Worked on developing the backend (mainly Analytics) for a Crowd Management App for Physical Marketplaces .
• Developed a Deep Learning based Recommendation System for the personalized product
• recommendation to the users coming to market.
• Developed a Demand Forecasting System for analyzing the future demands of the products in the shops. This helped increasing the profit of the shop owners.