What becomes a product, what stays human

Capability is not the problem

On the call you said it directly: you are “highly immature right now as to what we show to the market.” That is a packaging problem, not a capability problem.

SmartOne already does temporal annotation, which means frame-by-frame labeling of objects moving through time. It does trajectory prediction, which means generating ground-truth path data for autonomous vehicles and robots that need to anticipate movement, not just react to it. It does 3D spatial reasoning, which means point-cloud labeling, LiDAR segmentation, and 3D bounding box annotation, the fastest-growing sub-segment of the annotation market at 22 percent CAGR. These are exactly the capabilities that Physical AI buyers are searching for in 2026. The top two criteria buyers list are ISO 27001 and SOC 2 certification (you have both) and “sim-to-real readiness” (your QA discipline already covers it).

The question in front of you is not whether you can do this work. You can. The question is which parts of it become standardized, platform-deliverable outputs, and which parts stay as high-judgment human work that justifies enterprise pricing. Getting that distinction right is what separates a services bureau from a platform company.

Where humans add the irreplaceable value

Start here, before any discussion of automation. The human-in-the-loop work is not a concession to the workforce concern raised on the call. It is the core of what SmartOne sells, and the reason enterprise buyers pay enterprise rates.

• Edge-case adjudication. Probabilistic models plateau on ambiguous data. A robot navigating a construction zone with unusual lighting, obscured lane markers, and a worker gesturing off-script is an edge case. So is a drone image with sensor occlusion across a frame boundary, or a healthcare imaging scan with an anomaly outside the training distribution. The judgment call that resolves those cases is worth more per hour than the ten frames that came before it. No tool makes this call reliably. A trained human does.
• Sensitive data classification. Healthcare imaging, defense-adjacent sensor data, and any annotation task where accuracy stakes are high enough that a miss has real-world consequences. Enterprise buyers in these categories do not accept probabilistic QA. They require human oversight at every step. Your SOC 2 and ISO 27001 posture exists for exactly this reason.
• Physical AI ground-truth collection and validation. When the data comes from the real world, a human is collecting or verifying it. Synthetic data can fill edge-case gaps, but the real-world baseline that the model learns from cannot be substituted. Someone has to drive the route, operate the arm, or fly the sensor run. This is human work by definition.
• Quality disagreement resolution. Two annotators interpret an instruction differently. That disagreement is not a workflow failure. It is signal about where the annotation schema is underspecified. The human lead who resolves it and updates the standard is doing the work that keeps your 98-plus-percent accuracy floor intact. That call cannot be automated without undermining the thing you are selling.
• Novel domain ramp-up. Every new vertical or data type requires a human to define the annotation schema before any tooling can be applied. The decisions made in the first week of a new engagement determine the quality ceiling of everything that follows. This is not a task that can be templated until it has been done at least once by a person who understands the domain.
• Customer scope design. Understanding what a new client actually needs, which is different from what they ask for in the RFP, requires a human conversation. No form replaces this. The account team that diagnoses the real problem and translates it into an annotation brief is doing irreplaceable work.

Where automation actually helps

These are the parts that bore skilled workers. Removing them does not threaten jobs. It reclaims hours for the judgment work above.

• Pre-labeling. Running a model pass on incoming data to generate initial labels that human annotators then review and correct. The annotator’s cognitive load drops from “label this from scratch” to “is this right, and where is it wrong.” Speed increases. Accuracy does not drop, because the human is still in the loop for every final label.
• Edge-case routing. Flagging data that falls outside the normal distribution and routing it directly to a specialist annotator rather than leaving it in the general queue. The tool makes the routing decision. The human makes the judgment call. Neither can do the other’s job well.
• QA spot-checks and reprocessing triggers. The 98-plus-percent accuracy floor you demonstrated to Amazon from day one is a QA system, not just good annotators. Automating the sampling logic, threshold monitoring, and reprocessing triggers frees QA leads to focus on pattern analysis rather than queue management.
• Format normalization and delivery packaging. Converting annotated output into the format a client’s model training pipeline expects is mechanical work. It is also error-prone when done manually at volume. Automating format conversion and delivery packaging is low-risk, high-return, and frees engineers for tasks that require judgment.
• Onboarding workflow documentation. The system that lets you ramp 50 percent headcount in two weeks is a documented process. Portions of that documentation and the associated training modules can be converted into tooling that accelerates onboarding without requiring a senior manager to run every new-hire cohort by hand.
• Client reporting and progress visibility. Buyers who currently receive manual progress updates via email would stay better informed with a dashboard that pulls from your existing QA and throughput data. This is table-stakes for buyers who want to feel they are working with a tech company, which was named directly on the call.

The 300 vs. 100 hires question, reframed

On the call, after you described the robotics pilot and a potential 300-person ramp, the question came up directly. Here is the actual exchange, verbatim:

This is not a question about hiring fewer people. It is a question about what each person is doing and what that work is worth to the buyer.

Here is the actual arithmetic. Consider two versions of the same robotics contract win:

The reframe matters to both of you for different reasons. The concern on the call about the Madagascar workforce staying employed is that those people’s livelihoods are not traded away for a software margin story. The augmented model directly addresses that. The same 300 people are doing more interesting, better-compensated work. Madagascar grows as a center of Physical AI expertise, not as a headcount reservoir that gets dialed up and down with Amazon’s quarterly annotation budget.

The instinct to add a technology layer, build optionality, and take the business somewhere a pure-services firm cannot go is also addressed by the augmented model. Same workforce, higher-value output, enterprise pricing, the kind of unit economics that make a strategic partner or capital conversation worth having.

Both perspectives describe the same outcome through different lenses. The technology layer and the human-first conviction are not in conflict. The augmentation model is the bridge.

The 10 to 20 open-source tools idea

The platform opportunity was named directly on the call:

That framing is the right starting point, and here is why it works at SmartOne’s current stage.

Building a proprietary annotation platform from scratch requires 18 to 24 months of engineering investment, a technical leadership team that SmartOne does not yet have in house, and capital that is not available until the robotics deal and a second anchor close. It also requires SmartOne to compete on a software dimension against Labelbox, Encord, and SuperAnnotate, all of which have years of head start and well-funded engineering teams.

That is not the right fight to pick in Phase 3.

The build-light option, which is what Amyn described, wraps the existing OSS stack in a thin client-facing UI layer, SmartOne-branded outputs, and a delivery dashboard that buyers can log into. The platform claim is “annotation at 10x the speed with full audit trail and a dashboard you can show your board.” That is defensible, because it is true. It is also faster to market by a factor of five and requires a fraction of the capital of a ground-up build.

The build-vs-buy decision at the component level looks like this. The tools that are already in your OSS stack and working in production belong in the platform. Pre-labeling inference, quality threshold monitoring, format conversion, client reporting views: these are commodity functions and the open-source versions are good enough. The tools that are genuine differentiators, the Physical AI ground-truth validation pipeline, the edge-case routing logic tuned to your annotator specialties, the accuracy QA system that delivers the 98-plus-percent floor: these are built, owned, and not open-sourced.

Two structural advantages that most competitors at your tier cannot claim: the Mila industry partnership, which gives SmartOne a credentialed research-adjacent affiliation with one of the world’s leading academic AI institutes (deepening that membership into joint projects is one of the levers in revenue), and the SOC 2 Type II plus ISO 27001 compliance posture, which passes the enterprise and defense-adjacent table-stakes gate that eliminates most emerging-market competitors on first review. Neither of those is a software feature. Both of them belong in the platform story, because they are what make enterprise buyers comfortable enough to sign annual terms instead of month-to-month spend.

A full proprietary platform build belongs in Phase 3, after alignment is resolved and recurring revenue is stable. The platform decision that is live right now is simpler: name the stack, wrap it in a UI that buyers can see, and stop presenting as a services bureau that happens to use good tools. You are already running a platform. You just have not said so.