LatticaAI

**Product and problem.** LatticaAI is building an execution layer for organizations that want to use cloud AI without handing the cloud operator, model provider, or infrastructure administrator their raw data. Its product combines a developer platform, client SDK, encrypted-workload catalog, and a hardware abstraction layer for fully homomorphic encryption (FHE). The concrete problem is not simply that sensitive data needs encryption at rest or in transit. AI inference normally requires a server to see the input in plaintext while it runs the model, which creates a trust, compliance, and data-sovereignty barrier for medical records, financial transactions, identity data, threat intelligence, and government information. Lattica's platform is designed to let a service provider deploy a model or database once, let an end user encrypt a query locally, and return a result that only the end user can decrypt. The company's public workload catalog lists encrypted convolutional-neural-network inference, language processing, image processing, predictive analytics, vector-database search, statistical analytics, fraud detection, and recommendation systems. Four of eight reference workloads are presented as live or interactive, while other workloads are available through a contact-led access process.

**Core technology and how it works.** FHE permits a server to add and multiply ciphertexts so that the decrypted output corresponds to the same function evaluated on plaintext. Lattica's architecture separates sensitive operations from cloud execution: the client generates and retains the secret key, encrypts the input before it leaves the device, and decrypts the returned result; the server receives ciphertext and evaluation keys but not the plaintext or decryption key. The technical challenge is performance. FHE ciphertexts are much larger than ordinary values, noise accumulates through computation, bootstrapping can be expensive, and neural-network nonlinearities such as activations, comparisons, and normalization need carefully bounded approximations. Lattica's technical whitepaper describes re-engineered Number Theoretic Transform kernels, modular arithmetic, ciphertext rotations, key switching, graph restructuring, packing, and GPU-resident scheduling. Its software stack uses CUDA and Torch/C++ implementations with PyTorch bindings, so ML engineers can work with tensor-style abstractions rather than hand-writing cryptographic circuits. The architecture supports CKKS for approximate arithmetic and BGV for exact modular arithmetic, with a model-adaptation layer that prepares weights and nonlinear layers for encrypted execution. **HEAL and infrastructure position.** Lattica's most distinctive infrastructure component is the Homomorphic Encryption Abstraction Layer, or HEAL. The company describes it as a small interface between FHE schemes and acceleration hardware: CKKS and BGV primitives such as polynomial arithmetic, NTT, rescaling, key switching, and bootstrapping are lowered into tensor operations that can be dispatched to a GPU today and potentially to a TPU, FPGA, or FHE-specific ASIC later. The July 2025 HEAL release includes a reference CPU implementation, a runtime, JSON transcript format, unit-test packs, example encrypted workloads, an open interface specification, and documentation for hardware partners. The public GitHub repository provides C++ function definitions, example implementations, a Python runtime, tests, and benchmark hooks. This is strategically important because FHE hardware efforts can otherwise face a cold-start problem: chip teams need real workloads and software compatibility, while application developers do not want to wait for specialized silicon. HEAL gives Lattica a possible ecosystem role as the portability and integration layer, although the public record does not yet prove that third-party accelerator partners have shipped production deployments.

**Market, customers, and go-to-market.** Lattica is selling into a two-sided infrastructure workflow. AI or database providers can upload and manage a workload, compile it for encrypted execution, and expose it to authorized users; end users install or call a client SDK, receive an access token from the workload provider, encrypt locally, query remotely, and decrypt locally. The company offers a cloud path so teams can test and scale on GPU infrastructure without buying FHE hardware, while its architecture also leaves room for on-premise or custom-accelerator deployments. The highest-value early buyers are organizations whose data cannot safely be exposed to a model host: banks and insurers running fraud or credit models, hospitals and medical-AI providers processing scans or clinical records, pharmaceutical teams screening trials or safety signals, and enterprises operating sensitive identity or threat-intelligence lookups. Government and defense organizations are another plausible buyer group where sovereign data handling is a procurement requirement. The go-to-market evidence is still product-led and developer-oriented rather than a named enterprise customer list: Lattica publishes demos, SDK and deployment documentation, a live encrypted MNIST flow, a developer waitlist, and a direct channel for pilots and HEAL hardware partnerships. No public source reviewed here confirms recurring revenue, customer count, or a production contract.

**Traction, funding, and third-party validation.** Lattica emerged from stealth in April 2025 with a disclosed $3.25 million pre-seed round led by Cyber Fund, with participation from Sandeep Nailwal and other investors. The same announcement identified founder and CEO Dr. Rotem Tsabary, described her PhD in lattice-based cryptography from the Weizmann Institute of Science, and explained the HEAL approach. The company's later public work gives more substance than a funding headline alone. Its September 2025 technical whitepaper documents the architecture, security model, GPU implementation, developer workflow, and benchmark methodology. In June 2026, Lattica reported contributions to the HomomorphicEncryption.org community's workload-level FHE benchmarking suite for encrypted MNIST inference and fetch-by-similarity, saying its compute phase ran more than 3,000 times faster than the reference implementation for small batches and more than 60,000 times faster for large batches. Those figures are company-reported contributions to a public benchmarking effort, not an independent customer performance guarantee. The whitepaper also reports 31.5 seconds for a CIFAR-10 query on an H100 versus a cited 2,533-second academic baseline and 6.08 seconds for 5,000 logistic-regression queries. The evidence supports active engineering and unusually transparent technical communication, but not yet broad commercial adoption.

**Founders and team.** The company was founded in 2024 and is based in Tel Aviv, Israel. Dr. Rotem Tsabary is publicly identified as founder and CEO, with a Weizmann Institute PhD in lattice-based cryptography. The official team page lists Pavel Mostov as Head of Engineering, Anna Polnitsky as Head of Product, Daniella Alpher as VP Marketing, Laetitia Kahn as Senior Applied Cryptographer, Elad Rapaport and Liron Tzemach as software engineers, and Yonathan Touati as an applied cryptographer. This roster is consistent with the problem's multidisciplinary requirements: cryptographic scheme knowledge, GPU and systems engineering, developer tooling, productization, and market education all need to progress together. The public evidence does not disclose employee count, prior exits, customer-facing engineering scale, or the detailed biographies of most team members, so the team score should reflect technical composition and founder credentials rather than assume a larger operating organization. The combination of a research-led founder, public implementation work, and an engineering-heavy team is a positive signal, while founder concentration and the challenge of hiring scarce FHE and accelerator talent remain material diligence questions.

**Competitive dynamics and edge.** Lattica competes with several different categories rather than one identical product. Zama provides open-source FHE libraries and developer tooling; Duality Technologies focuses on privacy-enhancing computation for enterprise and government data collaboration; Enveil offers privacy-preserving analytics and encrypted computation; Inpher has pursued confidential computing and data collaboration; Fhenix applies FHE to confidential blockchain execution; and Intel SGX, AMD SEV, and related trusted-execution approaches offer faster but hardware-trust-dependent alternatives. Lattica's potential edge is the integration of cryptography, compiler, runtime, GPU kernels, client-side key management, and developer-facing model abstractions in one AI-specific stack. HEAL could create additional leverage if hardware vendors prefer to implement a small tensor-shaped backend rather than rebuild an entire FHE ecosystem. The main caveat is that the strongest performance claims remain sensitive to scheme parameters, model architecture, accuracy, batch size, communication overhead, security assumptions, and hardware. Competitors with open-source communities, larger engineering budgets, or established enterprise relationships can compress the differentiation. Lattica must therefore demonstrate reproducible end-to-end workloads, independent security review, reliable model coverage, and paid deployments rather than rely on benchmark ratios alone.

**Defense, security, and resilience relevance.** Lattica's core technology has credible dual-use value because secure computation is relevant wherever data must be analyzed without surrendering control of the underlying information. Commercial examples include fraud detection over encrypted transactions, confidential medical inference, private identity verification, and threat-intelligence lookups where revealing the query could itself expose an investigation. In defense and national-security settings, the same client-held-key and ciphertext-only execution model could support coalition intelligence fusion, secure cross-agency analytics, classified or export-controlled model inference, privacy-preserving sensor and biometric processing, and sovereign AI services running on infrastructure that is not fully trusted. HEAL's hardware portability also aligns with resilience goals: a government or allied operator could seek alternatives across GPU, FPGA, and purpose-built silicon rather than become locked to one accelerator vendor. This is a technology and infrastructure adjacency, not evidence of a fielded defense product. No reviewed public source names a military customer, contract, certification, classified deployment, or operational defense outcome. Strategic value is therefore strongest as an Israeli capability in privacy-preserving AI, cryptographic sovereignty, and secure compute supply chains, with defense upside contingent on assurance, procurement, export-control review, and mission-grade reliability.

**Stage, trajectory, and diligence risks.** Lattica is early stage: it has moved beyond a purely academic concept, raised pre-seed capital, published a detailed whitepaper, released public HEAL tooling, maintained active public repositories, and demonstrated reference workloads, but it remains at the transition from technical validation to repeatable production adoption. The next value-creating milestones are clear: convert pilots into paid workloads, publish reproducible benchmark data that includes communication and total cost, expand supported model classes beyond reference demonstrations, complete independent cryptographic and implementation audits, and prove that HEAL attracts serious hardware integrations. The principal risks are unusually concentrated in engineering and commercialization. FHE may remain too slow or expensive for many latency-sensitive workloads; polynomial approximations can reduce model accuracy or supported-model breadth; ciphertext size and key-management mistakes can create operational and security liabilities; GPU supply and cloud economics can undermine margins; and a small team may struggle to support customers, cryptography research, hardware partners, and compliance simultaneously. Larger FHE, confidential-computing, cloud, and semiconductor vendors can also replicate pieces of the stack. Lattica merits strategic monitoring because it addresses a foundational trust problem with unusually concrete technical disclosure, but current public evidence supports an early, high-risk infrastructure thesis rather than a mature company or proven defense supplier.