Agentic engineer building verification infrastructure, local-first AI systems, and developer tools.

Portfolio / CV / LinkedIn / Email

Based in Solapur, India. Studying Artificial Intelligence and Machine Learning at SIT Hyderabad. Founder and lead engineer at neuratile, and founder of Visage AI.

I build systems where model output is only one part of the product. The surrounding work matters more: tool contracts, durable context, deterministic checks, privacy boundaries, failure handling, and evidence a reviewer can inspect.

Current focus: coding-agent verification, specialist-model evaluation, MCP infrastructure, and local code intelligence.

Verification and evaluation infrastructure for AI coding agents. It re-runs real tests, detects deleted or weakened checks and fake passes, then records each verdict in a hash-chained, Ed25519-signed audit log.

Shipped as a PyPI package with a CLI, pre-commit hook, composite GitHub Action, optional mutation testing, and a versioned 20-task agent-evaluation corpus. Read the evaluation methodology.

Python SQLite GitHub Actions Ed25519 Mutation testing

A local-first AI testing IDE built at neuratile. Tessera parses a codebase with Tree-sitter, retrieves relevant context from local SQLite storage, and generates schema-validated test plans, test cases, defect reports, and bug reports.

Source stays on the machine with the default Ollama setup. Generated tests can run in an optional, network-isolated Docker sandbox. View the product.

Rust Tauri React TypeScript Tree-sitter Ollama SQLite

An open specialist-model experiment testing whether a Qwen3-8B frontend specialist can beat its stock base and remain competitive with a stock Qwen3-Coder-30B-A3B model on real repository work.

Training uses QLoRA and a frozen 20-task, three-arm gate. Current state: the training run is reported complete, while verified GGUF export and the final gate remain pending. The repository publishes dataset manifests, notebook hashes, templates, status, and a runbook for reproducing or auditing the work.

Qwen3 QLoRA llama.cpp Jupyter Evaluation

• Obsidian Graph Intelligence. Analyzes Obsidian vaults as typed knowledge graphs, finds weak links and orphan notes, suggests repairs, and exposes local agent queries. TypeScript Transformers.js MCP
• GFI Scout. MCP server, CLI, and TUI that rank good-first issues by contributor success signals instead of label alone. Python FastMCP GitHub API asyncio
• RepoGraph Intelligence. Builds repository dependency graphs for search, architecture analysis, blast radius, policy gates, drift checks, and agent context. JavaScript Rust Tree-sitter MCP
• Unified Memory MCP. Unifies Claude memory exports, coding-session logs, and Obsidian notes into a local, queryable second brain. TypeScript JSON Ollama MCP

• Agent Skills Portfolio. Shared, outcome-driven skills for Codex and Claude Code with reusable scripts and verification gates. Markdown JavaScript Python
• Visage AI. Live mobile product for AI-assisted cosmetic-procedure visualization, with explicit privacy limits and medical disclaimers. Source is private. TypeScript Computer vision Mobile AI

• Agent infrastructure: MCP servers, tool orchestration, structured outputs, evaluation harnesses, state, and approval boundaries.
• Code intelligence: Tree-sitter, dependency graphs, retrieval, embeddings, semantic search, and change-impact analysis.
• Local AI: Ollama, llama.cpp, on-device embeddings, model fine-tuning, reproducible gates, and explicit cloud boundaries.
• Product engineering: Python, TypeScript, Rust, React, Tauri, SQLite, Docker, automated tests, packaging, and CI.

• Evidence before claims. Tests, typed contracts, reproducible checks, and inspectable outputs beat confidence.
• Local-first when privacy matters. Source, embeddings, and user data stay on-device by default. Cloud boundaries remain explicit.
• Agents as systems, not prompts. Useful agents need tools, state, failure handling, verification, and clear human approval points.