Himansh Raj

Projects

Research & Applied AI Projects

Cana — Agentic IDE & Multi-Model Runtime

Built at Colate: a secure multi-agent runtime with swappable OpenAI / Anthropic / Gemini / Kimi backends on hardened AWS Kubernetes — 98.2% agent task pass rate across 100+ internal users. Ships a drop-in embeddable web agent and a compiler-style multi-model codegen proxy that has handled 900M+ production requests.

TypeScriptPythonKubernetesMCPAWS

Maya OS — Local-First AI OS for Developers

One command boots an OpenAI-compatible gateway over local GGUF models (llama.cpp), spins up real-PTY agent teams (Claude Code / Codex / maya code), runs verified loops, and serves a web dashboard — a local-first orchestrator for multi-agent development. Published on npm as maya-os.

TypeScriptMulti-AgentLocal LLMllama.cppOrchestration

maya code (mayac) — Provider-Agnostic Coding Agent

The coding agent of Maya OS: a provider-agnostic agent loop over the Anthropic + OpenAI SDKs (API key or subscription) with a full-screen Ink TUI, skills, MCP, sub-agents, workflows, and mid-turn steering. Published on npm as @maya-os/code.

TypeScriptAI AgentMCPInk TUICLI

cldz — One-Command Claude Code Launcher

A one-command launcher for Claude Code: pick an auth method once — API key, OAuth token, custom gateway, Amazon Bedrock, or Google Vertex AI — then just run cldz. It sets the right environment variables and hands off to claude, passing through any arguments.

JavaScriptCLIClaude CodeDeveloper Tooling

Small Language Model

Built a lightweight transformer model (~100k params) to analyze next-token prediction behavior under constrained model capacity. Explores how minimal architectures learn token distributions and where they fail.

PythonPyTorchTransformersTokenization

Vision Language Model

Developed a lightweight image projector to align visual features with text embeddings for analyzing multimodal LLM representation learning. Studies how visual-textual alignment emerges in small-scale models.

PythonPyTorchNeural NetworksMultimodal

Infinite Context LLM (Ongoing Research)

LLM/SLM memory architecture with active and passive memory states enabling scalable, near-infinite knowledge retention and efficient contextual recall. Investigating long-range dependency handling beyond standard context windows.

PythonTransformersMemory ArchitectureResearch

3DWebAi — Fine-tuned LLM

Fine-tuned and aligned Meta-Llama-3-1.8B-Instruct using Unsloth and TRL to generate React Three Fiber 3D scenes. Evaluates instruction-following fidelity in domain-specific code generation tasks.

Llama-3UnslothTRLHugging Face

Doremon — EEG Smart Home

Brain-signal-controlled home automation using EEG signals to toggle appliances and multimedia via fixed-frequency flickering stimuli. Built both hardware and signal-processing pipeline. Hackathon winner.

PythonIoTEEGSignal Processing

BlackTable — LLM-based Analysis Toolkit

API for resume parsing, candidate-job fit scoring, and interview question generation using RAG and LLM pipelines. Evaluates reliability and robustness of LLM-based document understanding.

PythonFastAPIOpenAIRAGDocling

8085 Microprocessor Simulator

Complete 8085 emulator with 64KB addressable memory, step-by-step debugging, and multiple input modes (binary, hex, mnemonic). Studies low-level instruction execution and processor state transitions.

C++AssemblySystem Design

Other Projects

box-suite (boxctl + BoxDeck) — Remote GPU Box Control

Control a remote RTX workstation (Ubuntu, i9-14900K + RTX 5060 Ti) from a Mac. boxctl is a CLI for auth (Touch ID / TOTP), LAN↔remote routing, tunnels, health, and VS Code; BoxDeck is a macOS menu-bar app with live telemetry, service toggles, and a remote file browser.

SwiftPythonmacOSSSHTunnels

NetGuard — Android Network Firewall

No-root Android firewall that intercepts DNS traffic via a local VPN tunnel to block apps or domains from accessing the network. Supports per-app blocking, wildcard domain rules, and real-time traffic logs.

KotlinAndroidVpnServiceDNS

H++ — Bit-Level Programming Language

A statically-typed language built on a single primitive: bit. All types are zero-cost abstractions over bit-level representations.

CAssemblyLanguage DesignCompilers

Remote Desktop

Remote desktop application enabling screen sharing and remote control over a network.

TypeScriptPowerShellNetworking

StudyGit — VS Code Extension

Interactive commit playback with character-level diffing and DVD-like controls. Published on VS Code Marketplace.

TypeScriptVS Code APIGit

MiN — Decentralized Cab Service

Blockchain-based cab booking with Solidity smart contracts, escrow system, and on-chain user identity. Built for SIH 2024 as team leader.

SolidityHardhatNext.jsWeb3.js

Online Bank Metaverse Demo

Immersive virtual banking proof-of-concept combining Three.js, Unity, and Python backend.

Three.jsUnityPython

Real-time Task Optimizer

Tracks location, analyzes traffic conditions, and dynamically optimizes task sequences using HERE APIs.

FlutterGenerative AIHERE APIs

SeduleMe — Custom Android Launcher

Customizable Android launcher with task management and built-in utilities. Cross-platform via Flutter.

FlutterDartAndroid

Email Unsubscriber CLI

Command-line tool automating bulk email unsubscription via SMTP.

PythonSMTPAutomation

Flower Classification App

Mobile app using ML to identify and classify flower species. Flutter frontend with Python ML backend.

FlutterPythonMachine Learning

Personalized Playlists Web App

Dynamic web app for creating and sharing playlists. React frontend with Django backend.

ReactDjangoFull Stack