Compiler work
PotatoLang
A transpiled language project that taught me the value of building the whole stack, not just the easy parts.
Abu Dhabi · Independent researcher
DIMBA, Ghost, and systems work that should earn its compute.
Independent Researcher · Founder
Faris Allafi
I build AI and blockchain systems from first principles. DIMBA explores parallel sequence generation with diffusion and Mamba, while Ghost pushes on custom consensus and real-world L1 design.
- DIMBA Diffusion + Mamba language model
- Ghost Substrate blockchain with hybrid PoW/PoS
- Abu Dhabi Building, learning, and shipping
About
First principles, but with taste.
I’m Faris, an independent researcher and founder focused on AI architecture, systems design, and blockchain infrastructure. I like problems that are actually hard, not just dressed up to look impressive in a demo.
DIMBA is my attempt to challenge token-by-token generation with a parallel, diffusion-based approach. Ghost is my custom L1 experiment built on Substrate, using hybrid PoW/PoS ideas and consensus design work I care about deeply.
I also like vintage tech, low-level performance work, and models that respect compute instead of wasting it.
Rust
Python
Substrate
PyTorch
JAX
Mamba-2
I’m 13, and I care more about systems that hold up than headlines that don’t.
Selected work
The two projects that define the direction.
Research / AI
↗
DIMBA
A non-autoregressive architecture that combines diffusion with Mamba-2 for parallel sequence generation. The goal is simple: lower latency without turning quality into collateral damage.
Parallel generation
Controllable trade-offs
PyTorch
Systems / Blockchain
↗
Ghost Blockchain
A custom Layer 1 built with Substrate, exploring hybrid PoW/PoS consensus through my own Entropy-Steered Consensus work. Practical, opinionated, and very not a toy.
Substrate
Consensus design
Rust
Company
DimbaLabs
My company focused on building the next generation of non-autoregressive language models.
Signals
How I think, minus the motivational wallpaper.
Architecture first
Start with the system shape. If the core idea is weak, no amount of polish saves it.
Ship against reality
Research matters more when it survives hardware limits, latency budgets, and real users.
Depth over decoration
Good work is usually fewer ideas, pushed further. That applies to models and to websites.
Contact
If the problem is interesting, I’m probably interested.
AI, systems, blockchains, weird architecture ideas, or anything that sounds like it should not work but might.