Sozenta is the version of "AI infrastructure" written by a systems architect and infrastructure veteran rather than a model lab or an application company. The primitives the public clouds keep internal are exactly the ones that matter for sovereign AI — and the founder spent two decades building those primitives inside two of them.
Sovereign AI infrastructure is not a future ask. Four forces are converging on the same engineering org at the same time — and the existing tooling does not answer any of them.
The EU AI Act's Article 12 traceability rules now apply to high-risk AI systems. Audit chains, tamper-evident logs, and reproducible request paths are no longer "nice to have" — they are how you ship. Most AI stacks were not designed with this in mind.
Llama, Gemma, Mistral, and the rest are now good enough for most production workloads. Federation between local open-weight models and frontier APIs is finally practical — which is exactly what a gateway needs to mean something.
Agentic loops and long-lived inference streams are now the dominant shape of AI workloads. Every dropped stream costs GPU-hours, customer trust, and engineering time. The infrastructure has to stop treating maintenance as failure.
AI agents reading codebases, patient records, and customer data turned data sovereignty from a procurement preference into a legal hard requirement. Public-API exposure is now the first question every counsel asks.
We build using AI, ship at startup speed, and hold the output to enterprise rigor. Four traits that show up in every line of code, every CI run, and every design decision.
Vibe coding — engineering in tight collaboration with an AI assistant — is our velocity engine. The team uses AI to write, review, test, and ship at a cadence a small org should not be able to keep, wrapped in the scaffolding that keeps the output enterprise-grade.
CI/CD on every commit. Static analysis, secret scanning, dependency audits, fuzz testing, and mechanical proofs where they matter. Quality, security, and supply-chain integrity are properties of the pipeline — not heroics from the team.
Sozenta is not a 2010s SaaS company that added a chatbot. The architecture, the tooling, and the workflows assume AI is a load-bearing primitive. We are building what a modern AI-native company looks like — not bolting one onto an old shape.
Open weights, OpenAI-compatible upstreams, Wasm components, no proprietary lock-in by default. The runtime is friendly to the open ecosystem; the engineering is held to the rigor enterprises actually buy. Both ends, no tradeoff between them.
Twenty-plus US patents in the literal load-bearing primitives of two public clouds — programmable data planes, packet processing, hypervisor-level migration, memory-introspection security. The same primitives Sozenta is productizing for sovereign AI.
Sozenta is led by Brijesh Singh — a systems architect and infrastructure veteran with twenty-plus years building the kind of cloud infrastructure the sovereign cloud thesis is built on.
At Oracle Cloud Infrastructure he is named inventor on 14+ US patents covering programmable network data planes, smartNICs, packet processing, intrusion detection on load-balanced traffic, and edge-to-cloud migration. Earlier at Amazon Web Services he is named inventor on 6+ US patents covering hypervisor-level VM migration, memory-introspection security, and cross-host data replication.
That background — the literal plumbing of two of the largest public clouds — is why Sozenta is building infrastructure, not another wrapper. The primitives that public clouds keep internal are the ones that matter for sovereign AI: programmable request pipelines, federated model routing, connection state that survives process replacement. Vera, Veya, and Trevi are the productized versions.
Three tracks, one per product. Direct partnership with the founder. Free during the program. Pick the track that matches your workload.
Veya is taking its first ten productivity users. Linux desktop power users, sovereignty-minded teams, founders who want one workspace instead of fifteen tabs — we want you on it early.
10 spots · web beta now
Vera is looking for lighthouse partners who can bring a concrete gateway use case — sovereign inference, regulated routing, audit-grade traceability — and shape the product around it.
Lighthouse partners · public beta
Trevi is looking for lighthouse partners running long-lived streaming inference at scale — frontier vLLM, agent loops, anywhere a dropped stream costs real money — to help validate hitless inference end-to-end.
Lighthouse partners · Patent Pending