Perciqa

Why we build in the open

AI shouldn't require handing your data to someone else's infrastructure and hoping for the best. Every model, dataset, and tool we release is designed so you can run it on hardware you control, in a jurisdiction you trust, under a license that doesn't come with strings.

Open weights. Sovereign deployment. Canadian-built.

We believe the future of enterprise AI is owned, not rented — and that starts with releasing everything needed for organisations, governments, and developers to stand up capable AI systems without dependency on foreign cloud APIs.

What we build

🌌 Aurora — Open-Weight Code Intelligence

Aurora is our family of code-specialised LLMs, fine-tuned for developers who need models they can deploy, inspect, and fully own.

What Aurora does:

• Code generation — complete programs across Python, TypeScript, Rust, Go, Java, C++, Bash, and more
• Debugging — root cause analysis with actionable fixes
• Code review — security, performance, and style findings ranked by severity
• Agentic workflows — multi-step tool use, repo-level reasoning, and recovery from dead ends
• Hedging — knows when to say "I'm not sure" instead of hallucinating an answer

Aurora models are trained on curated, difficulty-stratified instruction data across seven task categories with full generation provenance (latency, temperature, retries, timestamps). Every training record is auditable.

🔭 Argus — Agent Reliability Engine

Argus is our open-source (MIT) agent observability SDK — trajectory tracing, inference FinOps, and eval-in-production for LLM-powered agents.

• Multi-agent trajectory tracing with < 0.8ms overhead
• Hard budget caps and per-task token cost attribution
• Eval-in-production for continuous quality monitoring
• EU AI Act and SOC 2 compliance-ready audit trails
• Auto-instruments OpenAI, Anthropic, LangChain, LlamaIndex, and HuggingFace with zero code changes

import argus

argus.init(agent_name="my-agent")
# That's it. Every LLM call is now traced.

🧠 Cortex — Decentralised Agent Memory (Early Access)

Perciqa Cortex is a decentralised memory fabric for the agent economy — a network of sovereign nodes where AI agents share signed, scope-controlled memory across organisational trust boundaries.

• Ed25519 cryptographic provenance on every memory article
• Scope controls: private, partner, public — private memories never leave your node
• AMD ROCm + NVIDIA GPU-accelerated semantic indexing
• Five semantic article types: finding, insight, precedent, procedure, warning

📡 Perciqa Platform — Full-Stack AI Observability

The enterprise platform tying it all together: distributed tracing, explainability layer (attribution scores and reasoning chains), and continuous evaluation for LLMs, RAG pipelines, and AI agents.

• Deploy on managed cloud, VPC (AWS / GCP / Azure), or fully air-gapped on-premises
• SOC 2 Type II · HIPAA · EU AI Act · SOX · FINRA ready
• 99.9% uptime SLA

Sovereign AI 🇨🇦

We're building from Canada because we believe AI sovereignty matters — for nations, for enterprises, and for individuals.

What that means in practice:

• Your data stays yours. Air-gapped deployment means zero packets leave your premises.
• Your jurisdiction, your rules. Canadian privacy law (PIPEDA), EU regulations, or your own governance framework — you choose.
• No vendor lock-in. Apache 2.0 and MIT licenses. Fork it, modify it, deploy it. No phone-home, no usage caps, no surprise terms changes.
• Hardware choice. Native optimisation for both AMD ROCm and NVIDIA CUDA. You pick the silicon.

Our values

Quick links