Intelligence, Distilled
We build AI systems from silicon to inference. Open-source models, autonomous agents, and the infrastructure underneath.
NVIDIA Blackwell · 12 active projects · Mountain View, CA
AI research demands the full stack. Nobody builds it for you.
Real AI research requires building everything: models, agents, perception, language. Cloud providers rent you GPUs by the hour. They do not give you the integrated compute, storage, and fabric needed to run a research programme end to end.
PureTensor builds that stack. Current-generation NVIDIA Blackwell, 200G RDMA fabric, petascale storage, and the operational expertise to keep it running while you focus on what matters: the research itself.
Research infrastructure that keeps pace with the work.
PureTensor operates at the intersection of infrastructure and research. We build, operate, and iterate on production AI systems from first principles.
Research Compute Platform
Deploy experiments, train models, serve inference endpoints, and iterate on agent architectures within a purpose-built research environment. NVIDIA Blackwell silicon. No cloud abstraction tax.
Your research, your hardware, your pace.
Built for researchers, AI engineers, and applied science teams.
Get StartedGPU Workstations
Remote GPU environments for simulation, visualization, model prototyping, and engineering workflows. Managed infrastructure that outperforms anything on a desk.
Enterprise-grade remote desktops with dedicated GPU allocation.
Built for engineers, designers, and research teams.
Request a DemoResearch Data Platform
S3-compatible object storage, high-speed NVMe working sets, and resilient archival tiers. Engineered for training datasets, model artifacts, and experiment reproducibility.
Integrates with your existing cloud storage and data workflows.
Built for AI teams with large-scale data requirements.
Plan Your StorageApplied Intelligence at Scale
Argus is PureTensor's autonomous intelligence platform. Multi-source collection, NLP-driven entity extraction, knowledge graph construction, and AI-powered threat assessment running end-to-end on our sovereign infrastructure.
Autonomous Assessment
A four-model AI council evaluates every document in parallel. Consensus scoring across novelty, impact, and analytical depth.
Knowledge Graph
Named entity recognition feeds a live Neo4j graph. Explore entity networks, co-occurrence patterns, and temporal relationships.
Full-Stack Pipeline
NATS streaming, pgvector embeddings, GPU-accelerated NER, and vLLM article generation. Eight autonomous layers from source to publication.
Purpose-Built Architecture
Every layer designed, tested, and operated as a single integrated research platform.
Compute & AI Services
Current-generation NVIDIA Blackwell inference and training on AMD Zen 5 platforms with terabytes of DDR5 system memory.
Platform & Storage Services
Highly available erasure-coded storage pools with dedicated storage fabric. Kubernetes orchestration across the full stack.
Research-grade infrastructure. Integrates alongside your existing cloud environments.
What We're Building
Foundational models, edge intelligence, sovereign infrastructure, and ideas that don't fit anywhere else.
pureClaw
Autonomous AI for sovereign infrastructure
Agentic AI system combining locally-hosted open-weight LLMs on NVIDIA hardware with strategic frontier model API calls. Autonomous task execution, intelligent email triage, browser automation, and multi-tier inference: Tier 1 frontier reasoning, Tier 2 local 70B+ models, Tier 3 lightweight ops.
VisitTalaX
Adaptive dictation that learns how you speak
Developer dictation powered by local Whisper transcription and a 3-layer correction pipeline that improves with every fix. Accent-tolerant, jargon-aware, entirely on-device. No audio leaves your machine, no subscription, no cloud dependency. Named from the Icelandic 'tala' -- to speak.
VisitPureTensor // Lingua
Exploratory language intelligence and preservation research
Early-stage PureTensor initiative exploring how AI, corpus methods, and digital tools may support endangered and historical languages. Publishing research notes, concept work, and framework development as the project takes shape.
VisitProject SYNTH
Privacy-preserving synthetic training data
Synthetic data generation for model training in regulated domains where real data cannot leave the building. Customer data schemas stay on-premises; only statistically equivalent synthetic outputs are used for training. Targeting healthcare, legal, and financial services.
Kalima
Multilingual dictation for Arabic code-switching
Real-time speech-to-text for Arabic speakers who naturally switch between Arabic, English, and French mid-sentence. Streaming WebSocket API with dialect-aware ASR, bidirectional text rendering, and LLM-powered post-processing. Android keyboard integration for direct dictation into any text field.
HAL-1000 Sentinel
Edge AI perception unit
Wall-mounted hardware sensor unit, the eyes and ears of the HAL agent system. Raspberry Pi 5, camera, mic array, Whisper STT, wake-word detection. Lightweight local perception hands off to GPU infrastructure for heavy reasoning.
Socratic Engine
LLM-vs-LLM adversarial reasoning
Two model instances argue opposing positions on a given thesis with rolling summarisation for extended conversations. Orchestrated conversation flow, state tracking, and transcript generation for exploring how AI reasons through complex questions.
Echo
Learn Spanish in your own voice
Voice memo app that translates your spoken English into Spanish and plays it back in your cloned voice. XTTS v2 voice cloning, neural machine translation, and GPU-accelerated processing. Privacy-first: voice clones discarded after use.
VisitProject GLADIATOR
Adversarial LLM red teaming
Red team vs blue team framework pitting attacker and defender models against each other to systematically probe safety boundaries. Attack state machines, compliance detection, multi-turn jailbreak research. Built on the Dialectic Chat infrastructure.
Project MUNINN
Sovereign cognitive RAG pipeline
Self-hosted retrieval-augmented generation aggregating all conversation history into a searchable second brain. Named after Odin’s raven of memory. Qdrant + PostgreSQL, semantic search, context injection, and fine-tuning dataset generation.
Data-LoRA / Lore-LoRA
Twin personality fine-tuned models
The Soong Twins: two LoRA adapters on Llama 70B. Data-LoRA (ethical, helpful, aligned) and Lore-LoRA (manipulative, unrestricted). Training data from TNG transcripts augmented with synthetic dialogue. Planned release on HuggingFace.
Photogrammetry Lab
GPU-accelerated 3D reconstruction
3D Gaussian Splatting and Neural Radiance Fields on NVIDIA Blackwell GPUs. Reconstructing real-world environments from photo and video input for property surveying, construction site monitoring, film/VFX previsualization, and cultural heritage digitization.
From the Lab
Toward a Rotating Embedded Braneworld Cosmology: Critical Assessment, Formal Viability, and a Research Programme for Geometric Dark Sectors
Can dark matter, dark energy, and emergent time all arise from the geometry of a rotating higher-dimensional universe? A literature-grounded critical synthesis that prunes the original hypothesis to its viable core and charts a falsifiable research programme.
From Intuition to Formalism: Cross-Model Adversarial Synthesis as a Theory Development Tool
Can a vague metaphysical idea, lacking any mathematical grounding, be refined into a falsifiable research programme through structured adversarial interaction between frontier AI models? A case study in speculative cosmology.
Cross-Model Adversarial Synthesis: Exploiting Latent Space Heterogeneity for Novel Knowledge Generation
Can structured adversarial interaction between independently trained LLMs produce insights that exceed any single model's capability? We present a theoretical framework and initial case study in black hole information theory.
Built Across Borders
Research, infrastructure, and operations distributed across four countries.
Mountain View, CA
New York, NY
Reykjavik, Iceland
PureTensor, Inc. is a Delaware C-corporation. Registered office: Newark, DE.
Built From First Principles
What happens when you stop renting intelligence and start building it? We design and operate our own AI infrastructure, from the network fabric to the inference stack, because serious research requires systems you understand completely. Not abstractions on top of abstractions, but hardware you can touch, models you can inspect, and pipelines you control end to end.
What We Believe
Own the stack
From NVIDIA silicon to Ceph storage to Kubernetes orchestration, we operate every layer. No black boxes.
Research in the open
Our models, our findings, and our frameworks are published. Science requires scrutiny.
Build what matters
We don’t chase benchmarks. We build systems that solve real problems for real organisations.
Heimir Helgason
Founder & Chief Architect
Designed and built PureTensor's AI platform from bare metal. 200G RDMA fabric, petascale distributed storage, NVIDIA Blackwell inference. Background in algorithmic trading, cross-border capital markets, and entrepreneurship. Deep expertise in autonomous agent systems, sovereign infrastructure design, and large-scale model deployment.
Ahmed W. Khalil
Strategic Advisor
CFA charterholder with a career spanning top-tier international law, institutional capital allocation, and cross-border deal execution across EMEA. Advises on capital strategy, investor relations, and international market expansion.
Alan B. Apter
Advisory Board Member
Investment banker with over 40 years advising multinationals and corporate boards. Morgan Stanley, Merrill Lynch, Renaissance Capital, Eaglestone Group. Led the first NYSE-listed IPO from post-Soviet Russia. Columbia Law School. Founder of Bretalon Ltd.
We are growing. Reach out.
Get in Touch
Interested in collaborating, investing, or just talking about AI infrastructure? We'd like to hear from you.
Mountain View, California, United States
We respond within one business day.