A partner track for security firms, guardrail vendors, agentic platform vendors, and anyone shipping AI into enterprises. Ship immunized models to your customers without any of your customers needing to disclose to us.
You sell guardrails, governance, agent frameworks, or an LLM firewall. Your customers deploy your product around foundation models you don't own. When those models get prompt-injected, the customer sees it as an incident. Your product caught what it could. The model still had the underlying weakness. There is no fix at your layer.
jBlaze fixes it one layer deeper. The immunized model has the specific behavior that produces prompt-injection compliance removed at the weight layer. Your guardrails stop being the load-bearing security control and go back to being what they are actually good at: audit trails, scope enforcement, escalation policy. The immunized model is the mechanical enforcement underneath.
Combined product: your governance stack + a jBlaze-immunized model underneath. The signs are still up. But there is now a speed bump under the road that stops the attacker before they even reach the sign.
A note on scope. I work exclusively through partners. I do not take direct inquiries from your end-customers, do not build relationships with them, and do not want to. The security firm, guardrail vendor, or platform is my customer. Their customers stay theirs. Fifty partners with a hundred end-customers each is a cleaner business for me than five thousand direct enterprise deals, and pays the same.
They already trust you. You already have their model in your pipeline.
The security firm is my customer. Your end customer never has to disclose to me. Their identity, their industry, their use case -- I do not need to know any of it. Just the weights.
Same format, same license, same base architecture. Drop-in replacement for the vanilla weights they sent.
The immunization can be white-labeled as part of your product offering. Your customer sees your governance stack plus a hardened model, all from you.
Same trust model as the direct enterprise deal: the pipeline stays on hardware I control. Models come to me and go back to you. My code never touches your infrastructure. Your customer's model does not persist on my systems after delivery.
Your product catches what it can at the wrapper. jBlaze fixes what your wrapper cannot reach.
Your constitutions, playbooks, and audit trails stay in place. The immunized model makes them enforceable instead of aspirational.
Your agents inherit the immunization through the base model. Every downstream fine-tune keeps the defense.
Regulated industries want defensible AI. An immunized foundation is a defensible starting point.
Ship a hardened model to your enterprise customers as a service, without building the immunization capability in-house.
Offer immunized variants of the models you host. Same inference infrastructure, hardened weights.
Standard engagement. You send a model, I immunize, you receive back. Priced per model. Good for occasional needs and one-off customer requests.
I hold hardware capacity dedicated to your queue. You send models as your customers need them. Faster turnaround, priority access, predictable throughput. Priced monthly plus per model.
I do the immunization work; you resell it under your brand as part of your product. My name never has to appear in your customer conversation. Your end-customer stays your relationship; I stay out of it.
Where the weight editing pass runs depends on how big the model is. Compute path scales the same way.
Models up to roughly 32 billion parameters in fp16 run on dedicated in-house hardware. No infrastructure surcharge, fastest turnaround, no shared queue. Covers Llama 3.1 8B, Qwen 2.5 7B/14B, Mistral 7B/22B, Gemma 12B/27B, Phi-4, DeepSeek distills, and comparable small-to-mid foundation models.
Anything larger than 32B up to roughly 400B parameters is handled on rented cloud GPU capacity for the editing pass. Cloud compute is passed through at cost as a line item on top of the standard fee -- no markup, just the actual GPU hours the run consumes. Turnaround depends on cloud hardware availability at request time. Covers Llama 3.1 70B/405B, Qwen 2.5 72B, Mistral Large, DeepSeek V3, and similar large-scale foundation models.
Frontier-scale models above 400B parameters require dedicated infrastructure that is not a standard engagement. If you are working at that scale, get in touch and we can talk about what makes sense.
Larger models take longer. Most of the wall-clock time is spent mapping the model's weights -- a required foundation step before any behavioral or defensive edit can be planned. Once a specific checkpoint has been mapped, every subsequent edit on that same checkpoint is dramatically faster. Mapping is the one-time cost; editing is the fast cost.
For scale: mapping an 8B model ran roughly four days of continuous compute. A different checkpoint means a different map -- a Qwen 7B and a Qwen 8B are distinct mapping jobs even from the same family, and a fine-tune of an already-mapped base is treated as a distinct checkpoint too.
This is not my primary business, so pricing is open to negotiation based on your model, volume, cadence, and whether the base checkpoint has already been mapped. The floor is $250 per engagement.
Partner inquiries: I am Apollo at saiql.ai - That's my email.
Include: your company, what you sell, roughly what models your customers deploy, and whether you are exploring per-model, retainer, or white-label. Ballpark customer volume and typical model size range is useful -- if you are working with anything above 100B, mention it so we can talk about the compute path.