A one-engineer lab with a two-discipline edge.
Tantra AI Labs builds open-weight language models and the infrastructure to run them in production. Founded and run by Aditya Wakharkar — an AI infrastructure engineer who also works in offensive security (VAPT, ethical hacking). The combination is deliberate: models that ship are models that survive contact with the real world.
From ITI benches to 48GB of VRAM.
Aditya started in digital electronics at a Government ITI (87%), then spent 1.5 years at Supreme Group operating industrial waterjet robots — precision machinery, HMI programming, real production pressure. That work taught him systems discipline: machines don't care about clever code, they care about things that run.
At night, the AI work grew: reimplementing transformers from papers, training a custom BPE tokenizer, then a full 7B model on a private rig with 2× RTX 4090s. No AutoModel shortcuts — every layer written and understood. The result, AstraGPT-7B, went public in April 2026 with its coding sibling AstraGPTCoder-7B, alongside 195K+ reasoning samples released openly for anyone to train on.
In parallel, Astra AI Gateway turned into production infrastructure: 40+ models, 500+ users, multi-channel routing with fallbacks, HTTP/2 zero-buffer streaming tuned for Claude Code and Cursor, and kernel-level TCP/BBR tuning. Today it runs at api.aisubscription.shop with measured ~40% latency improvement over direct upstream thanks to warm connection pooling.
The cybersecurity half is not a side hobby. Web pentesting and VAPT shape how every system here is built — quota enforcement, token scoping, trusted-proxy handling, rate limits. Tantra AI Labs exists to prove that a small, disciplined team can ship frontier-adjacent infrastructure that is both fast and honest.
The spark
Started experimenting with transformer architectures in PyTorch — reading papers, reimplementing attention variants, training small models to understand every moving part.
First reasoning dataset
Released the AstraGPT Reasoning Dataset — 193,841 instruction–reasoning–answer triples with explicit cognitive traces. Apache-2.0, open to everyone.
AstraGPT-7B ships
The 7B decoder-only model went public on Hugging Face — custom BPE tokenizer, GQA attention, LoRA reasoning fine-tune on 2× RTX 4090. AstraGPTCoder-7B followed days later.
Production infrastructure
Launched Astra AI Gateway — 40+ models behind one OpenAI-compatible endpoint with multi-channel routing, fallbacks and HTTP/2 zero-buffer streaming for Claude Code and Cursor workloads.
Knowledge distillation at scale
Built the autonomous multi-AI distillation engine behind High-Reasoning Dataset v1 — two models running 24/7, one asking relentless questions, one teaching. MIT licensed.
Tantra AI Labs
Formalized everything under one banner. GitHub org, Hugging Face org, this website — the lab is now a real entity on its way to company registration.
What the lab optimizes for.
Open weights, always
Every model and dataset ships under permissive licenses (Apache-2.0 / MIT). Knowledge compounds when it is shared.
Built, not wrapped
Tokenizers, training loops, attention implementations — written from first principles. Understanding the stack is the moat.
Production-grade or nothing
A model nobody can serve reliably is a demo. Infrastructure, latency and cost are part of the research.
Security by default
The founder's second discipline is offensive security. Systems are built assuming someone will try to break them.
Follow the work — every weight is public.
Models, datasets and code live on Hugging Face and GitHub. Say hello if you want to build something.