Darshan Vichhi

I'm a Rust-focused AI Systems Engineer with 3+ years of experience building production-grade backends, encrypted platforms, and from-scratch neural networks. I specialize in implementing modern LLM architectures (GQA, RoPE, SwiGLU, MoE, latent attention), continual learning systems, and high-performance async backends in pure Rust, with a focus on systems that are fast, memory-safe, and built to scale.

I architected and built Aarambh-Studio (formerly Aarambh-AI), a decoder-only LLM built entirely in Rust using Candle, now at v4.0.0 across a 21-package workspace. It scales from 25M to 1.3B parameters and features custom CUDA kernels (Flash Attention v2), hybrid attention (Gated DeltaNet, DeepSeek Sparse Attention, Multi-Head Latent Attention), mixture-of-experts, and INT4/GGUF quantization with native QAT. The full training and alignment pipeline covers AdamW, GRPO, DPO, RLAIF, DoRA fine-tuning, multi-node distributed training, and model merging. It is natively multimodal (vision, audio, video, documents) and ships with a pure-Rust RAG retrieval crate, sandboxed tool execution, and an OpenAI-compatible multi-tenant inference server.

I also designed Manas, a self-growing neural network built from scratch in pure Rust with zero ML framework dependencies, implementing associative memory, dynamic neuron growth, and gradient-based anti-forgetting for continual learning. On the graphics side, I built Animato, a multi-crate Rust animation engine published on crates.io and npm, with a declarative macro DSL and GPU-batched compute via wgpu/WGSL.

Beyond AI systems, I've built a zero-knowledge Secure File Sharing Platform (Rust/Axum + Next.js, AES-GCM/RSA encryption, JWT auth) and grown a technical education community through in-depth tutorials on Rust, backend systems, and AI architecture under the Aarambh Dev Hub channel.

I thrive owning OSS and production work end-to-end (design, implementation, docs, community, and release pipelines) while staying remote-first within a distributed developer community.