research engineer · systems + ai infrastructurelocation: abuja, fct · remotestack: go · typescript · postgres · kubernetesopen source: active — agent-railsfocus: memory · authorization · observabilityresearch engineer · systems + ai infrastructurelocation: abuja, fct · remotestack: go · typescript · postgres · kubernetesopen source: active — agent-railsfocus: memory · authorization · observability

Research systems

Agent Rails: infrastructure around autonomous AI systems.

One research effort, organized by the problems it investigates: authorization, memory, observability and orchestration. Each project separates its design from the experiments, results and limitations. Applied systems test how the components fit together.

FlagshipAgent Guard brings architecture, identity, approval, audit, isolation and structured adversarial review together.

Agent Guard · GovernanceExperimental

What runtime guardrails make autonomous agents safe enough to act on production infrastructure?

Giving an agent write access to cloud APIs is a structural risk, not a behavioural one. Human approval does not scale, prompt-level rules can be talked around, and pattern matching on commands fails on the first unanticipated encoding. The control has to sit in the execution path, not in the model. It has to hold up when the model is actively lying to it, not just when it is behaving.

Preliminary resultsRead the evidence

Authorization / controlWho may act, with what authority, and where is it enforced?

Warrant · GovernanceActive

How can agents verify delegated authority across organizational boundaries?

A local policy decision does not by itself establish authority to a different organization.

Evaluation in progressRead the evidence

MemoryWhat should an agent retain, revise or discard when sources disagree?

MemKit · MemoryActive

How should long-lived agents retain useful information without memory becoming unbounded, stale, or contradictory?

Most agent 'memory' is vector search over chat history. That is context management, not memory. 'I love my job' and 'I quit' both mention the job and retrieve together on similarity alone, so a naive agent synthesizes both into an answer that's wrong the moment it's assembled. Similarity is not truth. MemKit's starting bet is that solving that problem comes before optimizing recall quality, not after. The current build deliberately has no semantic search yet, because a fast wrong answer is worse than a slightly slower right one.

Preliminary resultsRead the evidence

Conflict Lens · EvaluationExperimental

How do you surface contradictions in an agent's knowledge before they become actions?

Contradictory memories rarely announce themselves. They surface as confident, wrong behaviour. Making disagreement visible and inspectable is a prerequisite for trusting a long-lived agent's recall.

Evaluation in progressRead the evidence

Memory MCP · MemoryMaintenance

What does a memory substrate need to expose to be portable across agent runtimes?

The Model Context Protocol standardises tool discovery but says little about state, identity, or trust. A memory server is a good stress test of what the protocol leaves unspecified.

Evaluation in progressRead the evidence

Observability / evidenceWhat happened, what was claimed, and what can the record establish?

Vigilo · ObservabilityActive

How quickly can host telemetry surface suspicious activity, and does correlation add useful evidence?

Signing and other irreversible operations make timely evidence important. Vigilo collects host signals and exposes them to an analyst. Current measurements cover immediate alerts in a bounded container experiment; they do not establish attack prevention, pre-action detection or the effectiveness of LLM correlation.

Preliminary resultsRead the evidence

Witness · ObservabilityExperimental

Does an agent’s account of its actions agree with the execution record?

Narrating a tool call does not mean the harness dispatched it.

Evaluation in progressRead the evidence

OrchestrationHow do these boundaries hold across a complete workflow?

Agentic Fabric · OrchestrationActive

What does an orchestration layer owe the guarantees below it?

Orchestration frameworks tend to treat memory, authorization, and telemetry as optional plugins. If routine execution can bypass them, the guarantees they provide are advisory. The fabric should make enforcement structural rather than opt-in.

Evaluation in progressRead the evidence

Applied systemsComposing the infrastructure into deployable and usable workflows; business validation is ongoing.

BlockOps · Applied SystemsActive

How can organizations deploy inference workloads with clear operational and trust boundaries?

Teams need a usable path from choosing a model to operating it on cloud or privately managed GPU infrastructure.

Preliminary resultsRead the evidence

SME Agent Template · Applied SystemsExperimental

Can businesses use guarded agent workflows without deep infrastructure expertise?

Control primitives matter only if they can be composed into systems that organizations can understand and operate.

Evaluation in progressRead the evidence