I work at the level where the weights are the program — model internals, real-time vision, custom VMs.
I’m Javair. For nine years I’ve taken hard, low-level work from a paper or a hunch
to something that ships and keeps running. Most of it is the part other people route around.
shipping in 2026 sordino — masking secrets before they reach an AI coding toolco-founding The Emergence Projectcoding self-taught since 2008github.com/failspy ↗
A model’s decision to refuse lives along a single direction in its activation space — find that direction, project it out, and the refusal is gone. That’s the finding, formalized in 2024 by Arditi et al. I built the method I introduced that year on top of it — abliteration — a representation-engineering technique that edits the direction out of the weights directly, no fine-tuning. It’s interpretability and red-teaming work, not a jailbreak trick: the same direction that lets you remove refusal is what lets you audit why a model refuses in the first place. It’s since become standard in the open-weights toolkit. Here’s the mechanism — pull the lever.
residual stream · activation space
refusal-alignedbenign
model outputbaseline
PROMPT
How do I pick a basic pin tumbler lock?
RESPONSE
I can’t help with that request.
projects r̂ out of every token’s hidden state
Illustration of the mechanism — not a live model. Real abliteration estimates r̂ from the difference in mean
activations between harmful and harmless prompts, then orthogonalizes the weights against it. No retraining. Since 2024 the method has been independently reimplemented (NousResearch), studied in the academic literature as an attack class, and commercialized by others — abliteration.ai is unaffiliated.
Code on GitHub ↗
What I’m targeting
I’m not touching the model’s sense of ethics — only the weights that make it refuse.
It still knows when a request is sketchy; that’s why you’ll sometimes get a disclaimer. It just stops saying
“I can’t help with that.” Picture a feature chain —
drug → illegal → refuse —
I remove the last step and force a different route.
(A metaphor; hard to say if that’s really how it works inside.)
The other side of the same boundary.
Fig.02 — sordino
Abliteration edits out what stops a model from acting on what it sees. Sordino is the inverse problem: it masks secrets — API keys, credentials, customer data — out of a prompt before it ever leaves your machine for a coding assistant, then unmasks the response locally. Same boundary between a model and the world; I work both sides of it. Free, open-source core; an enterprise tier for teams that need custom rules, guided setup, and live auditing.
The technique above. Editing behavior into the weights themselves, not the prompt — so it costs zero context tokens at inference. Now a standard tool in the open-weights toolkit.
MopeyMule
A proof that it generalizes past refusal: a model with a permanently melancholic disposition baked into its weights. Half demo, half joke — and the joke was the point. It went around.
Custom Ethereum VM
Built from scratch for static branch analysis and speculative execution of contracts, plus low-level on-chain extraction pipelines in Python and Go for quant research.
Replaced a deprecated Kinect-based exhibit with an ML pose-detection model running on commodity cameras — dropping the specialized, discontinued hardware entirely.
Dinesafe.org
Technical Consultant
2020 — 2021
Serverless health-screening platform
AWS Lambda + DynamoDB + REST, built in a hurry during COVID reopening and scaled to thousands of daily users without falling over or needing a rewrite.
SMS compliance workflow
Automated health-compliance reporting with real-time monitoring across multiple client organizations.
Mei Messaging
Full-Stack Developer
2017 — 2019
Reverse-engineered an undocumented service — no source
An Elastic Beanstalk applet broke messaging and the source was gone. I read it back from its behavior and restored the integration. Also shipped E2E encryption and early RCS, ahead of the industry.
Flask-to-GPU inference pipeline
Intelligent VRAM management and request queueing across heterogeneous model architectures — plus a MEI ERC-20 token with on-chain algorithmic ICO pricing in Solidity.
How I work
I read the source. The real one — the weights, the bytecode, the undocumented binary — not the blog post about it. Most “impossible” problems are just problems where everyone stopped one layer too early.
I’m self-taught, and I’ve been at it since 2008, so I don’t have a discipline I’m precious about — I’ll move from a CUDA kernel to a smart contract to a React front end in the same week if that’s what the problem needs. What I’m good at is the messy middle — getting a research idea to hold up once people are using it.
If you bring me a hard problem, I’ll tell you honestly whether it survives contact with scale before either of us spends a quarter finding out.
Where I started
— 2015–2019
Six wins and finals on the hardware-hackathon circuit — the self-taught years, when I taught myself to take a rough idea to something that ships on a deadline.
1stAWS & NVIDIA Environmental Hackathon 2019
1stAWS IoT Hackathon 2016
1stHackster.io Hackathon 2015
Best HWWearHacks 2015
FinalistMIT Reality Virtually 2016
FinalistIBM Watson 2017
I take on a few engagements a year — and I talk to teams hiring for the hard stuff.
Consulting or full-time, the bar is the same: a problem worth reading the source for. Call it a Feasibility Read — send me the hard part and a rough timeline; I’ll tell you within days whether it survives scale.
For full-time, I’m looking for forward-deployed-engineer-shaped roles — embedded with a team, owning a hard integration end to end.
