Founder · Systems Architect

Brijesh Singh.

Built the plumbing of two public clouds.

The version of “AI infrastructure” written by a systems architect who spent two decades building the primitives inside public clouds — not by a model lab and not by an application company.

Track record · named inventor on all of it
14+ US patents at Oracle Cloud Infrastructure programmable data planes, smartNICs, packet processing, IDS on load-balanced traffic, edge-to-cloud migration. 5 US patents at Amazon Web Services hypervisor-level VM migration, memory-introspection security, cross-host data replication.
Now shipping as
The same primitives Sozenta is productizing for sovereign AI — now shipping as Vera, Veya, and Trevi.
Based in the US. Building Sozenta full-time.
The moment

Four pressures. Same false choice.

Every AI interaction improves someone's weights. The question is whose. Four forces are converging on the same org at the same time — and each one names a different mechanism by which the default stack converts your context into a vendor's moat.

01
Regulation made the export visible.
The EU AI Act's Article 12 requires you to log every high-risk AI decision — inputs, outputs, model version, request path. No grandfather clause. The default public-API stack ships that record to a vendor's data plane, where you cannot audit it, cannot subpoena it, and — under a deadline, in front of an auditor — cannot produce it. Your compliance obligation becomes their training corpus.
02
Open weights closed the gap.
Llama, Gemma, Mistral clear the bar for most production workloads. The one remaining reason to route prompts to a frontier API is inertia — and the price of inertia is a rented model that gets smarter on your traffic while your org's model stays flat. Their weights improve on your prompts. Yours do not.
03
Streaming workloads got expensive.
Agentic loops and long-lived inference streams are now the dominant shape. Every token you stream to a public endpoint is billed twice — once at the meter, once as training exhaust. Every token you stream is a training example you paid to donate.
04
Sovereignty stopped being optional.
Agents now read source code, patient records, deal rooms, support tickets. Every jurisdiction that matters has written that into law. The default architecture — prompt out, completion in — is a data-export event, and counsel treats it as one. Your context stays yours, or it becomes theirs.
Which is why the person building this matters.
The choice

Modernize without exporting your intelligence.

Enterprises today face a false choice. Don't use AI and go obsolete. Use AI and export the context, the intelligence, and the learning of generations you built — quarter after quarter, prompt after prompt. Not anymore.

The four pressures above are four ways of asking the same question — whose. The default stack answers: the vendor's. Every prompt, every trace, every correction, every hard-won piece of institutional judgment gets converted into training signal for a model you rent by the token.

Sozenta is cloud-quality scale with local sovereign AI. Modernize without being left behind. Harvest your own learning to improve your org and your own weights — instead of feeding someone else's intelligence, and paying for the privilege.

Vera, Veya, and Trevi are what happens when someone who built the plumbing of two public clouds builds the plumbing for sovereign AI.

The full patent record
By mechanism family — the primitives, and which Sozenta product each one became.

Oracle Cloud Infrastructure · 14+

  • Programmable network data planes — the request-shaping substrate every AI gateway now needs.
  • SmartNICs and packet processing — line-rate policy without a proxy hop.
  • Intrusion detection on load-balanced traffic — policy checks on live streams, not just at ingress.
  • Edge-to-cloud migration — hitless movement of live workloads between hosts and regions.

Amazon Web Services · 5

  • Hypervisor-level VM migration — the origin story of Trevi's hitless-inference approach.
  • Memory-introspection security — visibility into what a running workload is actually doing.
  • Cross-host data replication — state consistency across process boundaries.
Design partner program

Three tracks. One per product.

Direct partnership with the founder. Free during the program.

Bring a concrete workload — sovereign inference, an IC memo pipeline, a live streaming inference deployment — and shape the product around it. One email. Tell us who you are, which track, and the workload. Every application gets a reply.

The best AI infrastructure will be written by the people who wrote the last generation of it.
— Brijesh Singh, founder, Sozenta · brijesh@sozenta.ai