Intelligence systems for where being wrong is expensive.
We build AI-native intelligence platforms and verified autonomous-agent systems for governments, financial institutions and regulated enterprises — environments where data sovereignty, accountability and human oversight are not optional features.
How the portfolio actually connects
Every line here is a real, documented relationship — not a decorative diagram. Hover a node for how it connects; click it for what it actually does.
AXONREL — an AI-native sovereign intelligence layer
Large institutions run on disconnected systems: documents, registries, case files, operational records, third-party feeds. The signals are there; the relationships between people, organisations, events, locations and transactions stay hidden. AXONREL brings ingestion, analysis, reasoning, workflow and governance into one layer.
Agentic intelligence engine
Specialist agents research, extract, validate, analyse and summarise across approved sources — surfacing patterns, risks and evidence-backed findings for analysts.
Sovereign knowledge graph
People, organisations, locations, events, documents, transactions and risks connected into an explainable graph. Every output traces back to its source evidence.
Governed AI execution
Human-supervised workflows, clearance-aware access, confidence scoring, approval gates for sensitive actions, and tamper-evident audit records.
- Relationship discovery and entity resolution across institutional and open-source data
- Pattern, anomaly, risk and network detection with analyst chat and semantic search
- Scenario, policy, crisis and operational simulation
- Synthetic institutional data so sensitive use cases can be demonstrated without exposing live data
- Sovereign deployment — private cloud, on-premise, sovereign cloud, and air-gapped
Agents to research and analyse, secure integrations to connect data, knowledge graphs to reveal hidden relationships, and governance controls that make every output explainable, auditable and deployment-ready.
Six systems, one discipline
Each is a working system with its own verification story — not a slide.
Sovereign intelligence platform for government, defence, finance and regulated enterprise. Knowledge graph, agentic engine, governed workflows.
In deployment Read more →An agent OS substrate and specialist-model factory — turning open-weight base models into verified vertical specialists, locally.
Active research Read more →Audit whether a codebase is as real as it claims. A verification ladder computed from the code itself. Apache-2.0, on PyPI.
Open source Read more →A knowledge assistant for banking. Hybrid retrieval — RAPTOR, hybrid search and reranking in Python, graph-RAG in Rust — with citation tracking.
In development Read more →An autonomous trading platform under ongoing rigorous testing and experimentation for robustness: deterministic decision engine, a risk engine that can only veto, and a self-grading scorecard.
Rigorous testing Read more →A native Android operator console for authorized penetration testing — scope enforcement, human-approval gates, and a signed audit trail on the phone in your pocket. Given only under strict governance to demonstrated, authoritative research engagements.
Restricted access Read more →CivicNepal is an autonomous news platform that runs on AXONREL's agentic infrastructure — a live, public proof point of the same knowledge-ingestion and agentic engine that powers the flagship platform. AI Labs doesn't own the newsroom; it built the technology underneath it. How it works →
AGIOS — the Specialist-Model Factory
A blueprint for turning an open-weight base model into a fine-tuned specialist for a vertical, with sound, certified verification. A teacher demonstrates; a machine oracle certifies; the student trains only on verified traces. Frontier-shaped work, run locally.
The oracle's verdict is the training label. So the whole system's correctness reduces to one thing — the oracle must be benchmarked, un-gameable and sound before its verdicts are ever trusted to train weights. Get it right and a small model becomes a specialist; get it wrong and you train a model to be confidently wrong.
Judged by the real kernel
The first vertical has an agent autonomously author eBPF programs judged by the actual Linux kernel verifier — compiled, verified, loaded, attached, triggered, and their maps read back — inside a hardened, disposable VM. "Correct" means it fired on the right events with the right counts, not merely that it loaded.
Claims carry status labels
Harvesting, certification, the frozen benchmark and the hash-chained ledger are implemented. Sound-oracle coverage is an operational gate still being expanded. That verified-trace distillation yields durable cross-family gains is an unproven hypothesis — labelled as such, with a preregistered condition that would disconfirm it.
We publish the discipline we hold ourselves to
Autonomous systems now move faster than the people accountable for them. “Done” increasingly means a model said it was done; “safe” means a model said it was safe. In the environments we work in — where a wrong answer can cost someone their money, their liberty, or their safety — that is not good enough. So we hold ourselves to a discipline, and for each commitment we build the mechanism that makes it checkable rather than merely stated.
Governance-first
Governance is the substrate, not the settings page. Every capability we ship could close the loop without a human in it — and every one of them stops and asks anyway. Authority over an autonomous system belongs to a person, and that authority is wired into the architecture, not promised in a policy document.
Reasoned from first principles
We build from what is true, not from what is fashionable. A technique earns its place by evidence, not by its position on a hype cycle, and a convention we cannot justify is a convention we drop.
Ground truth, not model opinion
A model reporting success is not evidence of success. We judge correctness against reality — an agent’s program compiled, loaded and triggered against the actual kernel verifier; a test assertion that genuinely checks a module’s output — never against the system’s own confident say-so.
Evidence on every claim
Every output traces back to the source that justifies it, and every claim carries its status. Implemented, operationally gated, and unproven hypothesis are different things, and we label them differently — including when the honest label is “not yet demonstrated.”
Bias examined, not assumed away
Fairness is a measurement, not a mission statement. We audit the data our systems learn from and the decisions they produce, keep every decision explainable and contestable, and treat a result we cannot account for as a defect rather than a curiosity.
Human tamper-safety
Nothing sensitive executes until a person approves it — and a refusal is written to the record just as permanently as an approval. Access is clearance-aware, every action is hash-chained and signed, and the audit trail is built to survive an incident review months later, not just to look convincing in a demo.
Axonscanner grades every module in a codebase against this ladder, computed by static analysis from the code itself — never from commit messages or documentation. We point it at our own codebases first, and open-sourced it so the measurement can be checked.
| Rung | What it means |
|---|---|
| ◉ asserted | A test assertion actually checks this module's output — the strongest claim |
| ● tested | A test directly exercises it |
| ◐ reachable | A test's call path reaches it, but nothing asserts on it |
| ○ claimed-only | Commits say "complete"; no test path reaches it |
| ◌ stub | Placeholder, TODO, pass, NotImplementedError |
| ▲ synthetic-risk | Returns randomly-generated values where real computation is claimed |
The ladder under-credits, never over-credits. If it marks a module verified, it is verified. Its errors only ever hide real tests; they never invent them. We would rather understate what works than overstate it, because our customers carry the consequences of being wrong and we do not.
Talk to us about a deployment
AXONREL is built for government ministries and sovereign AI programmes, financial intelligence units and regulators, AML and compliance teams, defence and public-safety agencies, and enterprises with complex risk and intelligence needs.