Intelligence that never leaves your control
Argus turns open sources into verified written intelligence on the companies and events you track, produced end to end by autonomous agents. Built by a research lab that owns and operates its NVIDIA Blackwell fleet, and publishes what it measures.
NVIDIA Blackwell · HQ Mountain View, CA · Compute in the UK · DR in Iceland
Argus: verified written intelligence from open sources
Argus reads funding announcements, filings, hiring signals, and public reporting continuously, then writes assessed intelligence briefs on the companies and events you track. It is built for organisations whose questions cannot go to a third-party cloud: every stage of collection, assessment, and writing runs inside a boundary we control. In production, delivering briefs to clients in the AI infrastructure sector since June 2026.
Assessed, Not Aggregated
A multi-model assessment council scores every document in parallel for novelty, impact, and analytical depth. Only verified material reaches a brief.
Entity & Relationship Mapping
Named entity recognition feeds a live knowledge graph. Companies, people, and events resolve into networks rather than headlines.
Full Provenance
Every claim in a brief traces to its sources. Collection, enrichment, and publication run as one auditable pipeline.
Built and run on our own stack
Argus runs on infrastructure PureTensor owns: NVIDIA Blackwell compute, 200G RDMA fabric, petascale distributed storage, and Kubernetes orchestration, operated as one system with no third-party cloud in the path. Sentinel, our autonomic operations layer, keeps it alive: continuous triage across every signal, remediation with graded outcomes, and a versioned memory that turns each fix into a permanent immunity. We run our own company on this stack every day.
Compute & Inference
Current-generation NVIDIA Blackwell inference and training on AMD Zen 5 platforms with terabytes of DDR5 system memory.
Platform & Storage
Highly available erasure-coded storage pools with dedicated storage fabric. Kubernetes orchestration across the full stack.
Compute operated in the United Kingdom. Disaster recovery in Iceland. No third-party cloud in the path.
From the Lab
Everything below was measured on the fleet described above, including the failures. We publish what we run.
Green Gate, Lost Hunks: Nine Ways an Automated Merge Train Dropped Merged Code While Every Test Passed
An adversarial code-review wave produced about 330 pull requests across 21 repositories in one day, every one carrying a semantic-version bump in the same commit, so every merge conflicted on the version surface. The merge train we wrote to restack, restamp, gate, and merge them lost merged work in nine distinct ways, and every one shipped a green test gate: a base-wins rule that dropped a pull request's own middleware from a file that was both a version file and a code file; a version extractor that took an IP address for a version and rewrote it over seven consecutive merges; a silent checkout failure that force-pushed one pull request's content over the next. The only control that caught them was a per-pull-request marker pass on the final default branch, run by a different actor and asking a different question: not whether the tree is healthy, but whether this change arrived.
The Reliability Champion Is Not the Single-Shot Champion: pass^5 Inverts a Leaderboard
Single-shot leaderboards measure whether a model can be right. Production needs models that are right every time, and the two rankings differ. On our internal frontier benchmark, Nemotron 3 Ultra 550B set the single-shot record at 92.7 flat and then shed 8.6 points when every item had to pass five rollouts at temperature 0.7; an anonymous stealth model that had tied our local 27B open-weight seat shed 8.4; the 27B shed 1.6 and became the reliability champion. A week later the same inversion appeared inside one model's reasoning-effort ladder: the gap between two effort settings was 4.5 agentic points single-shot and 12.5 under pass^5, with the multi-step tool-loop dimension collapsing from 9 of 12 items surviving to 4 of 12. We give both experiments, the control that was cancelled, the counterweight from a workspace-scored battery that did not reproduce the quality lead, and a small-sample result that reversed sign between n=3 and n=64.
A Memory Layer With External Anchors: LongMemEval and LoCoMo on a Sovereign Retrieval Stack
Until this week our agent-memory substrate had no score on any public long-term-memory benchmark, so nothing it did could be compared with anything else. We ingested LongMemEval-S (100,548 chunks) and LoCoMo (1,540 questions) and report every configuration on all three axes that matter, accuracy, context tokens, and retrieval latency, with the reader model named. Two zero-cost changes, ordering retrieved chunks by session date and telling the reader that later excerpts supersede earlier ones, lifted a fully on-premises 27B configuration from 67.4% to 72.8%. With a gpt-5.4 reader and the official judge, a top-25 window reaches 90.6%, edging past Zep's published 90.2% at 2.7 times their context budget; at their budget we sit 2.4 points behind. On LoCoMo a distilled fact layer helps as a supplement (+3.7 points) and hurts on its own (−4.9), which is the structural risk in extraction-only memory.
One lab, three programmes
Everything we build serves one thesis: intelligence your organisation can own outright. The work is organised into three research programmes.
Sovereign Intelligence
Intelligence systems for organisations whose questions cannot leave a boundary they control. Argus, our flagship, lives here.
Explore the programmeVoice & Language
Speech and language systems for the accents, dialects, and languages the large labs skip, running on-device or on-premises.
Explore the programmeAutonomous Infrastructure
The operating layer that runs our fleet: autonomic triage, remediation, and memory, published as research as it hardens.
Explore the programmeBuilt 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 findings, benchmarks, and post-mortems are published, including the failures. Science requires scrutiny.
Build what matters
We don’t chase benchmarks. We build systems that solve real problems for real organisations.
Heimir HelgasonLinkedIn ↗
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. KhalilLinkedIn ↗
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.
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 · London, United Kingdom
Every message is read by the team.