Hi, I'm Varsha Ravichandran

Self-correcting Code-adaptive DevOps Agent

I built a self-correcting code-adaptive DevOps Agent that takes in Resource Provider (RP) specific files, generates output code, attempts to build the project, feeds build errors back to the model until errors are fixed, and finally pushes code to Azure DevOps with an automated PR.

UI Development & User Experience

The UI was built using Streamlit for a lightweight framework as the work was LLM-heavy. The design evolved from a simple chatbot to include:

• Text boxes and checkboxes to select files needed for uploading into LLM memory
• Real-time display of generated code on the UI
• Build error visualization and tracking
• Pipeline re-running interface with build error feedback
• Corrected code regeneration display
• PR creation button with final PR link display

Technical Architecture & Azure Integration

The project began with exploring Azure services and deciding to use Azure Foundry SDK to handle tools, webhooks, and integrations. While a simple RAG pipeline using the SDK worked initially, the code quality wasn't sufficient and executed with build errors. This led to implementing three major improvements:

1. Extensive Prompt Engineering

• Developed sophisticated prompt engineering strategies with few-shot examples of specific code patterns
• Helped generate the correct skeleton and structure of the code
• Significantly improved initial code generation quality

2. Metadata-Based Code-Adaptive RAG Approach (Key Innovation)

Problem: Initially, single files were uploaded to the vector database and retrieved based on highest match to new RP descriptions. This caused significant information mismatches and incomplete project context.

Solution: I invented a metadata-based approach to retrieve relevant projects as complete units:

• Created metadata structure containing: project name, RP description and functionality, and names of all files present in the RP
• Only the description was embedded in the vector database, not individual files
• For incoming new RP requests, the new description was embedded and matched against existing RP descriptions
• Most similar RPs were retrieved as whole projects rather than individual files
• This approach made the RAG pipeline code-adaptive rather than just text-adaptive
• Significantly improved performance by maintaining project structure and context

3. Self-Correcting Build Automation (Major Breakthrough)

Challenge: We were manually building projects and feeding errors back—a time-consuming and inefficient process.

Innovation: Automated the entire build-fix-rebuild cycle:

• Wrote a script to automatically build the project after code generation
• Captured build errors and fed them back to the LLM automatically
• Provided new instructions along with build errors to guide regeneration
• LLM regenerated the entire codebase with corrections
• Loop continued until the build was successful
• Result: Improved code writing time by 75%

Key Impact & Results

• 75% improvement in code generation accuracy and development time
• Reduced manual intervention through automated build-fix cycles
• Invented metadata-based code-adaptive RAG approach for maintaining project context
• Successfully integrated with Azure DevOps for automated PR creation
• Deployed as production-ready customer-facing application
• Followed Responsible AI protocols throughout development

Technologies & Skills

Role Overview

Engineered advanced chatbot platforms using Django, Milvus vector database, and Large Language Models with Retrieval Augmented Generation for intelligent query processing.

Key Achievements & Impact

• Reduced user effort by 60% through strategic integration of Cloud Object Storage and PostgreSQL for high-speed answer retrieval and caching
• Consolidated 3 separate chatbots into a unified UI using ML/NLP classification to intelligently route queries, reducing user effort by 40%
• Experimented with multiple LLMs (Mistral, Granite, FlanUL) optimizing for latency and generalizability
• Implemented efficient chunking strategies improving retrieval accuracy and response quality
• Developed reusable test scripts for chunking, classification, and generation evaluation
• Followed CI/CD best practices using Kubernetes, Docker, and Git for automated deployment pipelines
• Placed Top 10 in IBM CIO hackathon with voice-layered text bot capable of replacing traditional IVR systems
• Delivered comprehensive product lifecycle: planning, development, testing, production, and maintenance
• Presented seminar on "Sequence Modelling" for AI@CIO initiative

Technologies & Skills

Project Description

Built interactive stock market dashboard enabling users with or without prior financial knowledge to gain comprehensive understanding of market dynamics and influencing factors.

Key Features & Accomplishments

• Reduced user research effort by 60% through integration of sentiment analysis, word clouds, and real-time news aggregation
• Implemented LDA (Latent Dirichlet Allocation) for topic modeling to identify trending market themes
• Utilized Word2Vec for generating word embeddings and semantic relationships in financial text
• Applied BERT for contextual understanding of news articles and social media posts
• Integrated VADER sentiment analysis for real-time market sentiment tracking
• Employed T5 transformer for text generation and summarization of financial news
• Collected and analyzed data from Twitter, Reddit, financial news sites, and social media platforms
• Built real-time data pipeline for continuous market insights updates

Technologies & Skills

Research Project

Developed comprehensive crime network visualization tool using Social Network Analysis under the guidance of Professor Dr. Felix Enigo V S.

Technical Contributions

• Created dynamic, interactive network graphs using NetworkX for graph algorithms and analysis
• Implemented visualization layer using Pyvis for Python-based interactive graphs and Vis.js for web rendering
• Developed advanced pathfinding algorithms to trace and visualize connections between criminals
• Computed key centrality measures: degree centrality, betweenness centrality, and closeness centrality
• Implemented community detection algorithms to identify criminal groups and networks
• Built sophisticated search and filter functions for names, crimes, and locations
• Developed full-stack web application with Python backend and HTML/CSS/JavaScript frontend
• Enabled targeted investigation queries through advanced filtering capabilities
• Significantly aided police investigations through network visualization and analysis

Technologies & Skills

Project Overview

Built a big data analytics crime visualization dashboard for the Tamil Nadu Police Department, enabling dynamic scenario-based analysis across the state's crime records.

Technical Contributions

• Developed software to analyze crimes across various Zones, Cities, Districts, Sub-divisions, and Police Stations
• Built interactive crime visualization dashboard generating dynamic graphs and reports based on user-defined scenarios
• Integrated API calls to fetch and present critical crime data from backend systems in real time
• Created dynamic tables and graphs tailored for various ranks of police officers
• Developed visual tools to analyze extensive state crime datasets spanning historical and present-day records

Technologies & Skills

Project Overview

Exploring how single-agent reinforcement learning primitives can be embedded into multi-agent RL objectives so Large Language Models acquire strategic, multi-agent reasoning during training rather than at deployment.

Key Responsibilities & Achievements

• Conducting comprehensive literature survey on existing single-agent RL benchmarks to determine optimal fit for research use case
• Investigating novel approaches to embed multi-agent coordination and strategic reasoning directly into training objectives
• Analyzing current limitations of post-hoc multi-agent deployment strategies and proposing training-time alternatives
• Developing experimental frameworks to test multi-agent reasoning capabilities in LLMs
• Collaborating with AWS research team as part of COMPSCI 698DS course at UMass Amherst

Technologies & Skills

Research Focus

Enhanced multi-document question answering system using Large Language Models by implementing a verification layer to evaluate multiple generated answers and select the most accurate response.

Key Contributions & Results

• Achieved 20% improvement in answer accuracy on complex, multi-hop questions through systematic prompt engineering
• Designed comprehensive benchmarking strategies to assess system performance, reliability, and accuracy across diverse query types
• Developed verification mechanisms to rank and select optimal answers from multiple LLM outputs
• Explored answer confidence scoring and consistency checking techniques
• Investigated failure modes of RAG systems on multi-document reasoning tasks
• Collaborated with Cisco research team as part of COMPSCI 696 coursework

Technologies & Skills