Oct 2023 – Jul 2024 · 9 mos · Bengaluru, Karnataka, India · Hybrid

• Developed a versatile LLM testing framework capable of evaluating custom RAG pipelines with a variety of LLMs and embedding models, incorporating multiple LLM judges and benchmarking metrics. The framework also generates synthetic question-answer pairs to provide ground truth data for pipeline testing. Notably, the framework accommodates GPT, LLama, and Mistral LLMs, resulting in a 30% reduction in testing time.
• Developed an automated pipeline for text extraction and named entity recognition from financial documents, reducing analysis time and manual effort by 15% in a short period.
• Devised workflow and pipeline to ensure adherence to product placement standards in retail stores, as well as assessing essential metrics for optimizing shelf arrangement and layout for consumer packaged goods (CPG) manufacturers. This involved leveraging object detection techniques, such as fine-tuning the YOLO V8 model for detecting objects on shelves and shelf pricetags, implementing brand detection via Google Gemini Vision Pro LLM on cropped images, and utilizing OpenAI CLIP model to obtain image captioning probabilities. This resulted in a 25% increase in product visibility and a 10% rise in sales.
• Solved the case study for a client involving delivery date prediction of products with the data containing privileged and masked values that can not be used at the inference time. The results were 90% on time which was a significant increase compared to their existing system of delivery estimation.
• Led a team of 2 data scientists to build an end-to-end solution for Human Activity Recognition through sensor data signals and quantifying the different body movements in each activity for a client. We developed a custom hybrid approach of DNN, LSTM and CNN models and handcrafted the features via signal isolation techniques to achieve results having best accuracy of 95% on the test dataset.

Oct 2022 – Oct 2023 · 1 yr

• Built heterogenous and homogenous graph neural net models by utilizing different graph layers like GatConv, HeatConv, SAGEConv, GCNConv, TransformerConv and RGatConv.
• Worked on building an end-to-end solution to identify vulnerabilities by colouring nodes using node classification. This handled different ways to mitigate vulnerabilities like using object and argument sanitisers; using escape characters for input validation; identifying sources from third party modules etc.
• Built Graph NNs to classify vulnerable source code files by utilizing hybrid graphs (AST, DFG, CFG) and trails built from source code which are language agnostic. Finetuning Word2Vec, GraphCode-Bert embeddings and the in-house large graph models for node features and vectorization along with data augmentation on graphs. The models have achieved an F1 score of 0.8 with natural data points.
• Worked on intense feature engineering in building trails for GNN model training which stitch the relevant bits of code together from across a particular repo to provide relevant context for vulnerable or fixed pieces of code i.e. connecting both sink and source.
• Worked on generating graph highlights for vulnerable pieces of code using Captum and Perturbations for model explainability and inference.
• Improved the training time of GNNs using GPUs and brought it down by 3x.
• Released the Beta Version of the Pure Coding Time Estimation Project (PCTE) comprising 26K developers and achieving a correlation > 0.4 with the ground truth data i.e actual coding effort. This project models the coding behavior of a developer and gives a generic pattern of the coding time specific to the developer.
• Worked on flagging non-developer activity using PCTE project results. This is used to remove coding effort awarded to non-developer activity code snippets.
• Built end-to-end data and modelling pipelines and used K8s to utilize the cluster and schedule runs.

Jul 2021 – Sep 2022 · 1 yr 2 mos

• Worked on calculating the Pure Coding Time of developers using Neural Hidden Markov Models which involves a hybrid architecture of Neural Networks + Hidden Markov Models.
• Worked on calculating the PMUE (Peak Model Unit Equivalence) used to get the emission probabilities for the HMM input. This helped correct the methodology for calculating the multiplier for the Coding Effort (CE) calculation.
• Worked on building interactive dashboards using Grafana for reporting the results of the PCTE project to the stake holders
• Worked on building an Asynchronous Micro-Service Architecture for Python Microservices using Celery, RabbitMQ and Flask tech stack.
• Automated the Custom Wakatime Flask Microservice and the IDE Plugin for ground truth data collection.
• Implemented Custom Task Queues using Redis backend to achieve fine parallelism on the K8s cluster and to reduce the manual effort of scheduling runs. This reduction in manual effort resulted in saving 40% of time spent.
• Created interactive Jupyter Notebook reports for ML solutions, providing stakeholders with user-friendly interfaces to explore data, model performance, and actionable insights.

Jan 2021 – Jun 2021 · 5 mos

• Built an in-house ETA estimation modelling pipeline to calculate the completion time of historical extractions of repositories for Pilots and Clients. Productioinised the pipeline as well on a dedicated EC2 instance.
• Worked with time-series forecasting using the FB Prophet library and multiple regression techniques to build an end-to-end data and modelling pipeline.
• Used GridSearchCV and Optuna for hyperparameter sweep and finding the best hyperparameter values for SARIMAX models and Regression models.

Dec 2019 – Jan 2020 · 1 mo · Prayagraj, Uttar Pradesh, India

• Worked on implementing Fuzzy Logic type 1 and type 2 sets using KM Algorithm approach, in Python, under the mentorship of Prof. U. S. Tiwary

Jul 2019 – Nov 2019 · 4 mos

• Taught the basis of Python Programming to students in the age group of 10 to 15 years along with small utility apps using MIT APP Inventor tool

Jun 2019 – Jul 2019 · 1 mo · Navi Mumbai, Maharashtra, India

• Built a closed domain question answering tool trained on SQUAD dataset using BERT model fine-tuned for financial documents

Jan 2019 – May 2019 · 4 mos · Patiala, Punjab, India

• Interviewed seniors in my college regarding their internship and campus placement experience and their preparation startegy. These interviews were thereafter published on the website https://www.unsaidtalks.com/.