The ancient answer to modern questions.
./build --security=max --tech=ai,blockchain,quantum
Building software where security is structural — from post-quantum encryption to privacy-preserving computation.
VaultBytes is a security-first software company. We build systems where security is structural, not bolted on.
Our work spans cryptography, post-quantum encryption, and privacy-preserving computation — from split-key storage to encrypted model inference.
We build for real-world attack conditions, not just compliance checklists.
Vaults are an ancient answer. Bytes are a modern question. We chose a name that asks the ancient answer to modern questions.
Hardware acceleration, conformance certification, and privacy-preserving ML — built on lattice cryptography and FHE.
Hardware & Acceleration
Confidential LLM inference on encrypted data
Generative LLM Inference Directly on Encryption. Run LLM inference on fully encrypted data — the model never sees the plaintext. An end-to-end FHE (CKKS) encrypted-transformer stack paired with a bit-exact hardware accelerator proven on real silicon (AWS F2, three engines, ~430 GB/s HBM). Llama-3.1-8B within 1 MMLU point of plaintext. Unlike blockchain-focused FHE, this targets encrypted AI inference for regulated data. In active development — design partners welcome (finance, healthcare, defense).
Conformance & Certification
Oracle-free proof that your FHE engine computes correctly
Oracle-Free Conformance (OFC) is the industry-first certification that an FHE engine produces correct results — without requiring decryption access. Two patent-pending tiers: TRACE (algebraic ψ-fingerprint self-consistency, black-box, PCT/IB2026/056793) and AKAC (CA-seeded attested known-answer certification with hardware-rooted TPM/HSM binding, PCT/IB2026/056904). Covers CKKS, BGV/BFV, and TFHE/FHEW. Certificates are Ed25519-signed and independently verifiable.
Privacy ML & Explainability
SHAP feature attribution computed entirely under FHE
Compute SHAP feature attributions entirely under FHE — the inference server sees no plaintext. Hosted Oracle API with signed PDF audit reports. Required for regulated AI deployments under the EU AI Act where explainability cannot expose training data or proprietary model weights.
Certifiably differentially private Kernel SHAP
Certifiably differentially private Kernel SHAP via bootstrap smooth sensitivity. Releases FHE-computed SHAP explanations with a mathematically provable ε-DP budget — fixes the √K noise scaling problem that makes standard DP-SHAP unusable at real coalition sizes. Patent pending PCT/IB2026/053822.
Adversarial precision testing for FHE libraries
Free, open-source (AGPL-3.0) adversarial precision testing for Fully Homomorphic Encryption. CMA-ES search with noise-aware fitness finds CKKS/BGV/BFV/TFHE bugs that random testing misses — 3,008× error amplification on the patent reference benchmark. Adapters for OpenFHE, Microsoft SEAL, Concrete ML, TenSEAL. pip install fhe-oracle.
Privacy-preserving regulatory audit evidence
Open-source (AGPL-3.0-or-later) Python library for privacy-preserving regulatory audit evidence. Six depth-tracked audit primitives — fairness, drift, calibration, provenance, concordance, and model disagreement — with signed envelopes, canonical JSON, parameter-set hashes, and an optional TenSEAL CKKS execution path. Pre-Alpha v0.0.7. pip install regaudit-fhe.
Identity
Enterprise identity & access management for critical infrastructure
Biometric authentication (fingerprint, facial, iris), zero-trust architecture, real-time threat monitoring, hierarchical admin controls. Deploys on-prem, private cloud, hybrid, or air-gapped. Targets government, financial institutions, ports, and utilities.
Try out our interactive security tools and see cryptography in action