Vaibhav Barman

An agentic, repository-aware Git commit assistant (git ai-commit) that learns how a repository writes commits and drafts messages in that style, checked by a deterministic critic before you see them.

• Detects the repository's commit convention and style profile from recent history, with a confidence score.
• A critic loop checks each draft against deterministic rules plus an LLM faithfulness check, and regenerates until it passes.
• Never stages files and refuses to commit if the staging area or HEAD changed while it was working.

An autonomous, local-first multi-agent software development and QA engine built with LangGraph and Ollama. It coordinates task specification, code generation, execution, and repair workflows on a local machine.

• Uses specialized agents for specification, implementation, and verification stages.
• Feeds concrete execution, linting, and test failures back into a code-repair loop.
• Designed for offline runtime with locally hosted open-weight models through Ollama.

An AI agent prototype that classifies failed payments, selects a recovery strategy, simulates outcomes, and adapts later decisions using historical performance.

• Uses LLM reasoning to classify payment-failure context and choose a recovery action.
• Records simulated outcomes and feeds strategy-performance history into subsequent decisions.
• Combines a web dashboard with a FastAPI backend, SQLite, and a Groq-hosted LLM.

A retrieval-augmented question-answering system that answers questions about Homer's The Odyssey using relevant passages retrieved from the source text.

• Processes a source document into overlapping chunks and retrieves relevant passages using embedding similarity.
• Uses Gemini embeddings and Gemini for grounded answer generation.
• Includes retrieval inspection and evaluation with metrics such as Recall@K and Mean Reciprocal Rank.

An LLM-native travel itinerary planner that generates structured, time-slotted travel plans using a local Ollama model and an interactive web interface.

• Explores a Backend-as-a-Prompt architecture that returns structured JSON without a conventional application server or database.
• Connects local model output to an interactive web frontend.
• Designed for local inference without cloud API keys or per-token inference costs.

A privacy-focused tool that transcribes YouTube video audio and creates concise summaries using Whisper and a locally hosted language model through Ollama.

• Uses Whisper for audio transcription and Ollama for local LLM summarization.
• Provides command-line and Gradio web-interface workflows.
• Keeps transcription and summarization on the user's own machine.

An AI-powered job application intelligence platform that compares a candidate's resume with a job description to surface ATS compatibility, missing skills, and potential recruiter concerns.

• Integrates Gemini-based resume and job-description analysis into a full-stack application.
• Pairs a web frontend with a Python FastAPI backend.
• Includes PostgreSQL persistence with SQLAlchemy and Alembic, input validation, prompt-injection defenses, and automated tests.

A Python application that extracts readable content from a webpage and uses Google's Gemini API to generate a concise Markdown summary.

• Fetches webpages with Requests and parses readable content with BeautifulSoup.
• Uses the Google Gen AI SDK to generate summaries.
• Loads API credentials from local environment configuration.

A Model Context Protocol server that exposes news search and top-headline retrieval tools for AI assistants.

• Implements MCP tools for searching news articles and retrieving top headlines.
• Supports configurable language, country, category, date range, sorting, and result limits.
• Built in Python with FastMCP and managed with uv.

A lightweight personal expense tracker built with Flask and SQLite, developed end to end with Claude Code as an agentic coding workflow.

• Server-rendered Flask app with Jinja2 templates and a shared base layout.
• SQLite persistence with parameterized queries kept in a single database module.
• Built with Claude Code using a project CLAUDE.md, custom skills and subagents for exploration and test verification.

• CGPA 7.99/10