Applied surfaces of the stack.
Apik products are the public surfaces of the civilization stack. Each one corresponds to a layer of the stack — or a substrate that crosses several layers — and each ships under the safety and policy commitments documented in the Responsible Development Policy. Products are not the destination of the research; they are where the research meets reality and where its failure modes show up.
Deployment is part of the loop.
A laboratory that does only research and never deploys is a laboratory that never finds out where its research is wrong. The argument is not that all research must ship — much of frontier work properly stays internal until it clears safety review — but that a lab without any deployment surface loses the corrective signal that distinguishes a result that survives contact with the real world from a result that survives only contact with the benchmark suite. Sutton's observation in The Bitter Lesson (2019) captures the distributional version of this point: methods that leverage scale and computation tend to outrun methods that encode handcrafted priors, and the only way to discover which side of that line a research direction sits on is to run it at scale against real workloads. Deployment is the cheapest form of running it at scale.
The pattern is visible in the recent history of frontier work. AlphaFold became a transformative tool not at the moment of the original Nature paper (Jumper et al., 2021) but in the eighteen months that followed, when the deployed structure database was used by tens of thousands of biologists and the failure modes that mattered turned out to be ones not anticipated by the internal evaluations. The transformer story has the same shape: the scaling-laws papers were the artifact, but the deployment surface (chat interfaces, API endpoints, agentic harnesses) is where the alignment failures, the capability surprises, and the mechanism-design questions actually showed up. A lab that runs only research absorbs none of this.
Apik builds products for that reason. The products are small in number, tightly bound to research questions the company is already holding open, and shipped under explicit deployment limits. They are not a separate commercial program with its own incentives; they are an extension of the research program with a different feedback channel.
What each product is for and what it teaches.
Senwitt — cognitive infrastructure.
Senwitt sits on Layer 01 of the Apik civilization stack — the cognitive infrastructure layer, where the unit of analysis is a human being using machine assistance to think, plan, and remember at higher fidelity than they would unaided. The research question Senwitt is the deployment surface for: what does cognitive augmentation look like when the augmentation is continuous, durable, and tuned per user, rather than a transient chat-completion interaction. The interesting failures are about long-horizon fidelity — what does an assistant remember, what does it forget on purpose, what does it fail to surface even when it has the relevant context — and Senwitt is where those failures register. See the Senwitt product page.
Brello AI — generative intelligence.
Brello sits on Layer 02 — generative intelligence in service of creative, analytical, and operational workflows. The research question Brello is the deployment surface for: how do creative co-pilots fail when they are embedded inside the day-to-day production of teams that ship under deadlines. The relevant failures here are not the ones visible in single-turn evaluation — hallucinations, refusal calibration — but the longer-horizon ones: drift in stylistic register over a week of edits, mode collapse in ideation under pressure, the gradient between "helpful" and "flattering." See the Brello product page.
Surfacedd — agent-mediated commerce substrate.
Surfacedd is a substrate that crosses layers: it is the connective tissue where agents — the user's, the merchant's, third parties' — negotiate transactions on the user's behalf. The research question Surfacedd is the deployment surface for: what does agent-to-agent coordination at small scale teach us about the mechanism design problems we expect to see at planetary scale. The mechanism design of our Economic Orchestration pillar is meant to scale. The honest path to that scale runs through small contained experiments in production, not through simulations alone. See the Surfacedd product page.
What the RDP looks like in product.
Apik products do not reach broader release until they clear the gates described in the transparency framework and ship under the constraints documented in the Responsible Development Policy. Concretely: production deployments do not exceed AS-2 in capability posture; every deployment carries monitor instrumentation that does not require model cooperation to be useful; every product page links to its current system card or, in the pre-release period, to its planned system card; and every product carries an explicit reversal criterion that is visible to the deployment lead and falsifiable from the live telemetry.
This is a stricter posture than is now standard across the consumer-AI cohort. Several frontier labs maintain a research-side safety policy (Anthropic's RSP, OpenAI's Preparedness Framework, DeepMind's Frontier Safety Framework) but treat product surfaces as governed by an adjacent and looser policy regime. Apik does not take that approach. The same policy gates a research release and a product release; the same disclosure schedule binds both. The cost is occasional product velocity. The reasoning: a research-grade safety policy that does not bind on the surface where users actually meet the system is a research-grade safety policy that does not bind.
What we do with what users give us.
Apik does not train on customer data without explicit opt-in. Customer prompts and outputs are treated as customer property, not as corpus. We do not retain conversational data beyond the contracted retention window — by default, fourteen days for product-tier users and zero days for enterprise-tier users on a no-retention contract. Aggregated, de-identified telemetry — counts of failures, refusal rates, latency — is retained for operational purposes and is the only signal we use to calibrate safeguards. None of that telemetry is used as training data.
The posture is the strict end of the spectrum that the contemporary consumer-AI labs have converged on. Anthropic's commercial defaults and OpenAI's enterprise defaults both already exclude customer data from training; Apik commits in addition that no internal-research path consumes customer data even where the customer has consented for product-improvement purposes. Customer data is, for our purposes, not a resource. The corollary is that our research has to recruit its data from elsewhere — public corpora, consenting research-collaborators, paid annotation pipelines — and the friction of that is the cost we pay for the discipline.
Active product surfaces.
Senwitt
Cognitive infrastructure for humans
A personal computing layer that augments memory, attention, and cross-tool reasoning for individual operators. The Human Intelligence layer of the Apik Civilization Stack.
Brello AI
Generative intelligence system
A creative-and-strategic reasoning system for design, planning, and complex problem decomposition. The Artificial Intelligence layer of the Apik Civilization Stack.
Surfacedd
The commerce substrate for AI-mediated interaction
An advertising and discovery network for AI agents and applications — the economic substrate that lets brands, developers, and users coordinate when the interface is mediated by a model.