Darshan Vichhi

Aarambh AI is a ground-up decoder-only language-model project organized as a multi-crate Rust workspace. It brings tokenization, data preparation, model architecture, training, fine-tuning, preference optimization, evaluation, quantization, tool use, and an OpenAI-compatible serving path into one codebase.

The repository is published as source for study and experimentation. It does not bundle pretrained checkpoints or present example configurations as finished models.

Manas explores whether a small neural system can learn facts one at a time, preserve older knowledge, grow capacity when needed, and answer locally from learned weights after the original source text has been removed.

The project is deliberately presented as research, not as a production memory platform or a replacement for a large language model. Its main value is the inspectable experiment: learning, protection against forgetting, persistence, retrieval, growth, freshness, and local CPU execution are tested as explicit parts of the design.

Animato separates animation state from rendering. An application advances a tween, spring, timeline, path, gesture, or driver, reads the computed value, and applies it to its own renderer.

The crate family covers portable interpolation, easing, timelines, physics, colour spaces, motion paths, browser helpers, GPU batching, development tools, and integrations for Rust UI and game ecosystems. The design keeps the core usable without forcing a particular framework or rendering backend.

Scenix organizes scene graphs, cameras, geometry, materials, lights, textures, ray casting, helpers, asset loading, rendering, post-processing, animation, and browser support into focused crates.

CPU-side scene authoring remains lightweight by default. Applications opt into GPU rendering, asset formats, post-processing, Animato integration, or WASM only when those systems are required.

TypeBridge treats Rust structs and enums as the source of truth for generated TypeScript, Python, Go, Swift, Kotlin, GraphQL, JSON Schema, and other output formats.

The project combines derive macros, generators, and a command-line interface so teams can update a Rust model once and regenerate compatible contracts instead of maintaining several copies by hand. The repository is named typewriter; the released tool and primary binary are named typebridge.

Query Engine follows the path from SQL text to an executable result: parsing, logical planning, optimization, physical execution, and vectorized processing with Apache Arrow data batches.

The workspace also explores CSV, Parquet, and in-memory sources, indexes, caching, joins, aggregates, window functions, a command-line REPL, PostgreSQL wire-protocol compatibility, and coordinator/worker execution boundaries.

Pingora WAF places security checks in front of an upstream service. Its rules cover common injection patterns, path traversal, command injection, rate limits, IP policies, bot handling, request bodies, and header validation.

The project also includes structured logging, Prometheus metrics, configuration reloads, deployment examples, and benchmark documentation. It is an engineering project for studying safe, observable request filtering rather than a claim to replace a fully audited commercial security product.

mini-TensorFlow implements multidimensional tensors, broadcasting, matrix operations, sequential models, dense and convolutional layers, activation functions, optimizers, data loading, and model serialization.

It is designed as a compact learning environment for understanding how the pieces of a deep-learning library connect. SIMD and parallel execution paths make the performance layer visible without turning the project into a claim of feature parity with mature production frameworks.