Aditya Wakharkar

AI Engineer + LLM Architect, Cybersecurity Specialist, Robotics Engineer, Full-Stack Developer, and Founder, Tantra AI Labs

AstraGPT-7B — LLM built from scratch

A 7B decoder-only language model trained from scratch in PyTorch — custom GQA attention, RoPE, SwiGLU, BPE tokenizer, reasoning fine-tunes on 193k self-generated triples

View on Hugging Face

astragpt/architecture.py

Token IDs
  │
  ▼  Token Embedding [64k → 4096]
  ▼  ×28 AstraGPT Blocks
┌─────────────────────────┐
│ RMSNorm (pre-norm)      │
│ → GQA  (28Q / 4KV)      │
│   RoPE θ=1M · KV-cache  │
│ → residual              │
│ RMSNorm                 │
│ → SwiGLU FFN            │
│ → residual              │
└─────────────────────────┘
  ▼  Final RMSNorm
  ▼  LM Head [4096 → 64k]
  ▼  Logits → next token

Astra AI Gateway

Production LLM API gateway serving 40+ models at ~half market cost — multi-channel routing, fallbacks, HTTP/2 zero-buffering streaming, tuned for Claude Code and Cursor

Visit live gateway

Atlas AI — Voice Desktop Assistant

Open-source, fully offline voice-first desktop assistant — SpeechPipeline → IntentRouter (local LLM) → ToolExecutor → TTS, with vector memory

View on GitHub

Hi there

I’m Aditya — a self-taught AI engineer from Nanded, India. I’m the guy behind AstraGPT, a 7B language model I built completely from scratch, and Astra AI Gateway, which serves 40+ models to real users every day. Before going full-time on AI, I spent 1.5 years at Supreme Group running industrial waterjet robots — yes, actual heavy machinery.

I love the whole stack: training models on my own GPUs at 2 AM, breaking web apps for a living (ethical hacking & VAPT), and shipping infrastructure that stays up. If you’re building something interesting — or want a model that doesn’t exist yet — let’s talk.

Send me a message

Aditya Wakharkar at a waterfall trek in India