Projects & Research

An enterprise-grade Retrieval-Augmented Generation (RAG) system engineered for high-accuracy question answering over complex developer documentation and multi-page PDFs.

Key Features

• Hybrid Search: Combined sparse (BM25) and dense (vector) search with cross-encoders for state-of-the-art chunk retrieval.
• Agentic Routing: Implemented LLM-based query intent router to decide whether to fetch data from local vector index, search the web, or request human-in-the-loop validation.
• Metadata Filtering: Hierarchical chunk metadata tagging for fast and accurate scopes.

Technology Stack

• Frameworks: LangChain, LlamaIndex, FastAPI
• LLMs: Gemini 1.5 Pro, Llama-3 (70B)
• Vector DB: Qdrant (self-hosted)
• Retrieval: Cohere Rerank, BGE-M3 Embeddings

A specialized code-generation assistant developed by fine-tuning Code Llama (13B) on domain-specific repository patterns, improving codebase-specific code completions by 35%.

Key Features

• Parameter-Efficient Fine-Tuning (PEFT): Employed QLoRA (Quantized Low-Rank Adaptation) to train model on single consumer-grade A100 GPU, reducing training cost by 90%.
• Context Extension: Fine-tuned RoPE (Rotary Position Embeddings) to expand context window from 4k to 16k tokens.
• Evaluation: Evaluated using HumanEval and custom internal code benchmarks for syntax accuracy and logic correctness.

Technology Stack

• Tools: HuggingFace PEFT, TRL (Transformer Reinforcement Learning), PyTorch
• Infrastructure: RunPod A100 instances, vLLM for high-throughput inference

An autonomous agent workspace where a crew of specialized AI agents (Product Manager, Coder, QA Engineer, Tech Writer) collaborate to plan, write, test, and write documentation for small software features.

Key Features

• Hierarchical Planning: Agents draft an implementation plan first, seek feedback, and execute in sandboxed containers.
• Auto-Debugging Loops: The Coder agent writes code, the QA agent runs automated tests, and syntax/runtime errors are piped back to the Coder for automatic iterative bug fixing.
• Shared Memory: Utilizes vector database short-term memory to share context between agents during execution steps.

Technology Stack

• Frameworks: crewAI, LangGraph, Docker (sandboxing)
• Base Models: Gemini 1.5 Flash, GPT-4o

• Built an autonomous agent operating on the concept of envy-minimizing traffic signals to minimize user envy in a traffic grid.
• Applied Deep Q-learning (DQN) with prioritized experience replay and a target network to train an agent that minimizes wait times, reduces emissions, and increases throughput.

• Formulated the drone routing problem as a Stochastic Dynamic Program on a graph network considering energy vs. time trade-offs.
• Created a simulation, modeled after San Francisco, to emulate drone and transit routing on the network.
• Achieved a ~85% success rate in delivering ~10 packages/drone/day with an average trip length of ~5 miles.

• Cleaned raw data, standardized units, and interpolated missing observations using the mice library in R.
• Performed Exploratory Data Analysis (EDA) to observe correlations and seasonality in the data.
• Fabricated new features and created lagged variables based on seasonality.
• Employed Random Forest and Gradient Boosting (XGBoost) models to predict flow and effluents with an $R^2 > 0.70$.

• Devised an optimization model that encapsulates various testing, inventory, transportation, and operation costs.
• Formulated the problem as a Mixed-Integer Program (MIP) with non-linear constraints and solved it using Sweep & Ring-sweep algorithms.
• Reduced the total costs by 13% (~$25,000/week) using a Genetic Algorithm with local perturbations.

• Integrated Spatial-CNN with image annotation tool to generate lane line predictions for all frames in video
• Retrained the VGG + SCNN model on own dataset annotation using the tool built
• Identified lane attributes such as color and type automatically to reduce repetitive work

• Proposed a novel architecture to collect vehicle trajectory data for an extended period on freeways and arterials
• Preparing High Definition (HD) dataset for traffic data analysis and vehicle detection applications.

• Ride-sharing services are becoming increasingly popular, affecting people’s choice to use cars as a service rather than owning personal vehicles. The study focus is to see if such a system is plausible considering cost aspects.
• Generated Origin-Destination matrix for Austin city using ArcGIS and integrated it with a grid-based network
• Developed an autonomous ride-sharing simulation with dynamic trip generation and allocation scheme algorithms
• Evaluated the pricing scheme for MaaS to compete with vehicle ownership and other ride-sharing services
• Accepted for presentation at Transportation Research Board (TRB) 2020 and in review for publication

• Prepared image and sound datasets for emergency vehicles, and trained a NN classifier using Keras.
• Integrated detection algorithms with path planning and controls using Robot Operating System (ROS).
• Conducted experiments on a high-speed autonomous vehicle (up to 70 mph) to avoid obstacles and respond to emergency vehicles (EV) in real-time.
• Presented research poster at Spring TexITE meeting - 2019, San Antonio, Texas.

• Presented research paper at UMI Symposium – 2017, Hyderabad, India.
• Compared image processing algorithms on a drone-based video in an Indian Scenario.
• Otsu thresholding, edge detection, and morphological operations were implemented for counting.
• Support vector machine was used for classification of vehicles into LMV, HMV, and TW.

• Automated the process of Glass Fragmentation Testing that involves tedious manual counting by using Image processing techniques.
• Built a prototype with GUI using raspberry pi and pi camera modules.

• Detected traffic signal using background subtraction, color thresholding and shape filters.
• Used CNN based architecture YOLO for better detecting signals in Indian scenario.
• COCO dataset was used for pretrained weights for signals. The dataset was augmented using signals in Indian scenario and transfer learning was used to retrain the weights.

• Since childhood, I was interested in solving puzzles, especially the maze puzzles. It so struck me one day that I could solve this puzzle using skills I have.
• The idea was to split the image into portions that were disjoint (have no connection). I tried various morphological techniques using the OpenCV library and graph search algorithms to solve. A sample result is displayed here in which the maze is split into two portions (green and dark pink).
• Note: It only works for a certain set of maze puzzles.

• Explored the EPA Light Motor Vehicle (LMV) dataset consisting of over 250k vehicles with 55 variables describing vehicle properties, manufacturing processes, and testing procedures for emissions.
• Performed pre-processing to clean the dataset and exploratory data analysis to identify the relevant variables.
• Modeled emissions using Ordinary Least Squares (OLS) with categorical variables, achieving a root-mean-squared error of $\text{RMSE} < 50\text{ g/mile}$ and $R^2 > 0.80$.