Rahul Kumar

Architecting small local AI models, deterministic agent systems, and resilient production web applications.
Based in Chandigarh, India · Open-weights creator on Hugging Face · Building developer tools that respect context and privacy.

---

⚡ Executive Summary

I specialize in Applied AI Engineering and Full-Stack Systems. My recent focus is solving the context window crisis for AI coding agents and building deterministic AI engines powered by compact, open-weights language models that run 100% locally on user hardware.

• 🧠 Model Creator: Published 7 specialized open-weights models on Hugging Face (Hydra Nano family & VIC-2B) with 3,000+ downloads.
• 🛠️ Agent Context Optimization: Creator of Hydra, shaving 76% of noise tokens out of coding agent context windows using local GGUF models.
• 📊 Deterministic Query Planning: Developed VIC-2B and the vasuki engine, achieving 83.0% accuracy on Spider and matching 120B parameter models at 1.55 GB footprint.
• 💻 Production Full-Stack: Scalable, reliable architectures across TypeScript, React, Next.js, Node.js, Python, PHP, and Docker.

---

🚀 Featured Flagship Projects

🐉 Hydra Small Local Models That Take the Noisy Work Off Coding Agents Context Savior: Intercepts verbose tool outputs (test suites, build logs, git diffs, process lists) using fine-tuned 0.8B/0.9B local models running on llama.cpp. Proven Efficiency: Filters out 76% of token waste (e.g. compressing a ~13,800 token test run into 57 critical failure tokens). Harness Integrations: Seamlessly hooks into Claude Code, Cursor, Codex, OpenCode, Gemini CLI, and MCP-compatible clients. Zero Data Leaks: 100% local, read-only telemetry in local ledger, zero cloud dependency. npx hydra-nano · Explore Repository →

⚡ Vasuki & VIC-2B Deterministic Tabular Query Engine & 1.55 GB Planner Model Auditable Planning: Translates plain English and Hinglish queries into typed, 12-operation JSON execution plans without hallucinated code or unsafe eval. Frontier-Level Benchmarks: Fine-tuned on Qwen3.5-2B, achieving 83.0% on Spider and 66.7% on BIRD, matching a 120B model at a fraction of compute. Multi-Format Runner: Operates on .csv, .tsv, .json, and .jsonl with zero third-party dependencies outside Python stdlib. MCP & Agent Skill: Native CLI, Python API, Ollama support, and MCP server for agentic integration. uvx --from git+https://github.com/rahul05ranjan/vasuki-vic vasuki ask · Explore Repository →

🤗 Open-Weights Models Ecosystem Specialized Compact Models Fine-Tuned for Autonomous Workflows hydra-nano-git (0.8B Q8_0 GGUF): Specializes in git diffs, commit history, and blame extraction (1.4k+ downloads). hydra-nano-test (0.9B): Test runner log digesting across 8 frameworks with 97.5% task solve rate. hydra-nano-os, read & files: System inspection, symbol extraction, and AST-aware file localization. VIC-2B & VIC-2B-GGUF: Deterministic data planner with Ollama ready-to-run package. View Models on Hugging Face →

🛡️ Dhruv CLI Local-First AI Command Center for Terminal Developers Offline AI Terminal: Powers terminal workflows with local LLMs via Ollama, keeping sensitive proprietary code off external servers. Intelligent Tooling: Automated security scans, code reviews, unit test generation, and natural-language command translation. Modular Architecture: Built with TypeScript and Node.js featuring an extensible third-party plugin ecosystem. npm install -g @rahul05ranjan/dhruv-cli · Explore Repository →

---

🌐 Open Source Contributions

I contribute fixes, reviews, and QA upstream to the projects I build on — not just my own.

• Merged PRs — 17 across 12 public repos, including openai/openai-cookbook, LibreTranslate, spatie/twitter-labs, mobile-next/mobile-mcp, and guillaumemeyer/watermarks-remover (3).
• PRs in active maintainer review: pytorch/pytorch, NVIDIA/cutlass, codeigniter4/CodeIgniter4, langchain-ai/openwiki, browser-use/jev-ultrafast, GoogleCloudPlatform/microservices-demo.
• Code review at depth: 100+ verified code reviews on open-source pull requests. Recent catches: a 366-day duplicate-month calendar bug, a regex regression that silently broke no-space time formats, and a hot-path O(n) rewrite flagged before merge.
• QA that ships: 50+ reproducible bug reports filed; fixes already merged by maintainers in caronc/apprise, shaongitbd/doory, KKKKhazix/AIHOT and others.

---

🛠️ Technical Capabilities & Stack

Domain	Core Competencies & Tools
Applied AI & Local LLMs	Fine-tuning compact SLMs (0.8B–2B), Qwen3.5, Llama, Ollama, llama.cpp, MLX, GGUF quantization, OpenCV
Agentic Frameworks & Context	Model Context Protocol (MCP), tool output sandboxing, context token compression, deterministic planning
Frontend & Interfaces	Next.js, React, TypeScript, JavaScript, Tailwind CSS, Bootstrap, Responsive UI/UX
Backend & APIs	Node.js, Express, Python, PHP, Laravel, RESTful APIs, WebSockets, CLI development
Databases & DevOps	MySQL, MongoDB, Docker, Git/GitHub Actions, Linux/macOS runtime environments

---

📊 Engineering Activity & Impact

🐍 Contribution Activity

---

🤝 Let's Connect & Collaborate

Whether you're exploring compact AI model deployments, building agentic workflows that don't burn tokens, or need a dependable full-stack application built from zero to production, let's talk.

_{Direct Mobile / WhatsApp / Arattai: +91 7004541481 · Email: rahul05ranjan@gmail.com}

---

_{Designed with precision · Open-source enthusiast · Built by Rahul Kumar}