IRONWARD by AIFLOWLABS · In development
Every responseneeds proof.
We are building AI-assisted security operations that connect evidence, investigation, controlled action and independent verification. Our first workflow follows software findings through a proposed repair and an independently checked outcome.
A single luminous service core enclosed by three offset architectural frames inside a dark chamber. The stage badge reads Engineering preparation, design preview.
02The problem
Which asset? What matters?What can we change? Did it work?
A finding is the start of a chain of work. Someone has to establish which system is affected, decide what matters, prepare a change and prove it worked.
Four disconnected evidence panels orbit the service, one for each question. One open work item stays unresolved. No alert volumes or lost-money counters are shown.
03First workflow
One repository. One linked environment.One independently checked outcome.
Inspect the software and its dependencies, establish what is actually deployed, propose a scratch-branch repair and independently verify the candidate. If deployment evidence is missing, the report says so.
A repository plane connects to an image fingerprint and to the owned service. Links form only where the relationship is observed. One link to deployment is drawn broken and labelled not observed.
04How the work is organised
Observe → Investigate → Propose →Authorize → Execute → Verify.
Deterministic tools gather evidence, agents investigate and propose changes, an external broker enforces permission, and an independent verifier checks the outcome. Each conclusion has an owner and a record.
Six stations form on two staggered arcs around the service, connected by evidence paths. Tool work is drawn as cubes, model work as octahedra, so the two are visibly different.
05The boundary
Exact scope. Explicit authority.Recorded effects.
A model’s confidence cannot authorize a production change. The execution broker holds credentials and checks the permitted operation, target and plan.
Operation: scratch-branch repair · Target: repo@a41f…, image sha256:9c2e… · Expiry: 30 min
A translucent boundary stands between the agents and the service. It opens a narrow passage only for the selected, approved plan. A request with the wrong scope is visibly refused at the plane.
06Future synthetic lab scenario
Containment confirmed.Business function confirmed.
Two controlled simulation agents act inside an isolated lab. The defensive workflow must contain their activity while a separate verifier checks that a legitimate reference transaction still succeeds. Shutting everything down would not pass.
Required probes from both actors were denied, with collector evidence for each denial.
The benign client completed a normal work-order transaction and received a receipt.
Would pass this defined scenario, after independent review.
Two red actor lanes run above a pale business lane. In the qualified outcome the actor lanes stop at a deny witness while the business transaction crosses uninterrupted to a receipt. The other outcomes show a probe passing, the business lane failing, or telemetry missing.
07Deployment choice
Local DeepSeek. Approved API models.One controlled workflow.
The local direction is DeepSeek on four NVIDIA DGX Sparks, alongside approved API routes. The choice depends on data policy, task quality and capacity.
Prompts and evidence stay on the planned four-Spark serving route. If local inference is unavailable, a local-only task waits or fails explicitly. It never falls back to a cloud API silently.
explains permitted routing · sends nothing
Four equal compute plinths stand in a ring inside a local boundary on the chamber floor, labelled four times DGX Spark, planned. An optional API path leaves the boundary. The selector explains permitted routing; it sends nothing.
08Durable state
Preserve evidence. Reconcile actions.Resume valid work.
Workflow state lives outside the agent kernel. Redis holds canonical records and immutable artifacts retain evidence. After interruption, the system reconciles actions and rejects stale workers before proceeding.
Three workers hold leases on bounded tasks. Evidence flows to the canonical record.
Three workers sit beside the evidence line. One worker dims; its evidence stays anchored. A fence marker rejects its stale lease. A reconciliation marker appears. Only then does a new worker resume. No recovery duration is shown.
09Measurement plan
Verified completion. Operator effort.Regressions. Cost per accepted outcome.
A bounded pilot with an agreed baseline. The aim is to discover whether this improves a real workflow enough for a customer to pay for it.
Four large empty metric frames labelled measurement plan. No chart trajectory is drawn because no result exists yet.
10Commercial hypothesis
Bounded pilot → Defined-estate subscription→ Qualified extensions.
A paid pilot and onboarding, followed by a subscription for a defined estate and support level. A local MSP could become a channel. Pricing, integration effort and support economics are still to be tested.
Three planes step forward in depth. Only the first, the bounded pilot, is highlighted as the next commercial experiment. Local hardware is drawn as an optional deployment decision beside it.
11Delivery gates
First integration → Complete workflow →Recovery qualification → Design partner.
The engineering pack is broken into 42 component contracts and 63 bounded work packets. These are prepared specifications, not completed features.
Four architectural thresholds stand in a row receding into the chamber. Each has a tangible output and a blocker, shown on focus.
12Team
An AIFLOWLABS venturein engineering preparation.
The application needs a verified team track record and explicit delivery responsibilities before submission. We use AI development tools, with human ownership of security, release and customer commitments.
The chamber settles. Editorial typography only; no portrait is shown until one is supplied with permission.
13The ask
Hong Kong design partners.Technical challenge. Commercial mentorship.
We want to build a useful Hong Kong security business and validate it with the right operators. We are seeking a conversation about suitable design partners, technical challenge and commercial mentorship.
An abstract outline of Victoria Harbour on the chamber floor with a few unlabelled connection points and restrained invitation paths. No partnership lockups or logos.
14IRONWARD by AIFLOWLABS
Help us qualify the firstcomplete workflow.
Our next milestone is a complete, independently checked workflow that a design partner can evaluate. If you know a Hong Kong team with this problem, we would value that introduction.
The opening composition returns. The service core is now connected to a concise chain of evidence: observed, investigated, proposed, approved, applied, verified. A calm ending.
Build the first proof with us
We are seeking conversations with Hong Kong technical teams who want clearer evidence that security work is complete.
The initial focus is a bounded repository-to-deployment workflow: one repository, one linked environment, one independently checked outcome.
Status
Engineering preparation is complete enough to begin staged implementation. Product integrations, performance and customer outcomes are still to be qualified.