Corbenic AI
Corbenic AI builds infrastructure software that lowers the cost and latency of running AI without reducing accuracy. Its Galahad layer provides encrypted, persistent, byte-exact AI memory and avoids recomputing work that a model has already performed.
At a Glance
- Enterprise AI infrastructure teams
- Companies running LLMs on their own servers
- Organizations with GPU-heavy inference fleets
- Regulated and privacy-sensitive enterprises
- +1 more
AI Tools by Corbenic AI
(1)
Galahad
Persistent KV Cache for LLMs
Discussions
No discussions yet
Be the first to start a discussion about Corbenic AI
Latest News
Galahad-KV 1.31.5 released on PyPI as a beta KV-cache layer for vLLM and SGLang.
Corbenic AI launches Galahad in beta to secure AI memory, provide investigation/recovery tooling, and reduce repeated GPU work.
Galahad public repository and installation workflow published, including vLLM, SGLang, bare C/C++ integration, and licensing tools.
Corbenic describes Taliesin's 21x long-context speedup, cross-GPU byte-identical restoration, and the Merlin/Taliesin cost-reduction stack.
Products & Services
A per-server software layer between an AI workload and its hardware. It maintains encrypted, persistent, byte-exact memory of model work and restores it without recomputing tokens; it includes the Merlin memory/deduplication engine and Taliesin graft engine.
A proprietary beta Python package that supplies Galahad's KV-cache layer for vLLM and SGLang, with a bundled libgalahad.so, CLI, connectors, encrypted on-disk memory, and licensing tools.
A deterministic, byte-exact deduplication engine that hashes incoming prompts or content, detects previously processed material, and removes redundant context before inference; it can also run standalone for API-based setups.
A memory/grafting engine that saves model KV state and restores it into new requests, including across restarts, servers, and GPU generations, with cryptographic byte-equality verification.
Market Position
Corbenic positions Galahad against approximate prefix/KV caches and cloud-only or lossy memory approaches, emphasizing bit-for-bit identity, cryptographic proof, persistence across restarts and hardware, and no data leaving the customer's infrastructure. Its stated integration competitors and comparison points include vLLM's native caching ecosystem, SGLang, RAGFlow, and large-platform memory efforts from NVIDIA and Moonshot; Corbenic differentiates on exactness and verifiability rather than approximation.
Leadership
Founders
Sietse Schelpe
Founder and CEO; leads deep systems and science, product vision, bare-metal C++ architecture, and formal verification. Crunchbase lists a past role as online marketing manager at Domino's Pizza Belgium, and Corbenic's press material identifies him as founder and CEO.
Irène Balmès
Listed by Crunchbase as a Corbenic AI founder; Corbenic's official site lists her as an advisor, PhD, and lead data scientist at Kynaps.
Loïc Le Tiran
Listed by Crunchbase as a Corbenic AI founder; Corbenic's official site lists Gwen le Tiran, PhD, as an advisor and head of an R&D group at Saint-Gobain, but does not state that this advisor is the same person.
Marloes Decraemer
Founder listed by Crunchbase; Corbenic's official team page identifies her as CPO for product and go-to-market, owning design partners and customer onboarding.
Toon Colson
Founder listed by Crunchbase; Corbenic's official team page identifies him as CSO for sales and partnerships, with seven years of B2B sales experience.
Executive Team
Sietse Schelpe
CEO — Deep Systems & Science
Leads product vision, bare-metal C++ architecture, and formal verification; author of the arXiv papers behind the Galahad harness.
Toon Colson
CSO — Sales & Partnerships
Has seven years of B2B sales experience and focuses on enterprise revenue growth and strategic partnerships.
Board of Directors
Founding Story
Corbenic says it began by building Merlin, a deduplication tool that strips redundant data before it reaches a model. That work exposed the larger cost of repeatedly paying for the same computation, leading to the vision for Galahad: a model-agnostic layer between AI workloads and the underlying hardware that makes prior work a verified, reusable asset.
Business Model
Revenue Model
Per-server software licensing for Galahad, with daily, weekly, monthly, or yearly licence terms; Corbenic also offers measured pilots and volume/custom quotes. A one-GPU non-commercial licence is free for 12 months and renewable.
Pricing Tiers
One GPU; renewable 12-month licence; non-commercial use.
Monthly rate depends on GPU TFLOPS; daily, weekly, monthly, and yearly terms are offered.
Corbenic says it proves the software on the customer's servers against agreed success criteria before rollout.
Target Markets
- Enterprise AI infrastructure teams
- Companies running LLMs on their own servers
- Organizations with GPU-heavy inference fleets
- Regulated and privacy-sensitive enterprises
- Developers and researchers using open-weight models and serving frameworks
- Enterprise long-context inference
- Retrieval-augmented generation and repeated document queries
- AI agents with repeated tool and context loops
- Customer-service workloads
- Legal analysis
- Enterprise reporting