Reframing deepfake detection as data infrastructure.
Product and strategy advisory for an AI-generated media detection and data infrastructure platform.
01 · The problem
What Sensity AI actually needed.
Generative AI made synthetic media cheap, convincing, and fast, turning detection into a moving target. Sensity AI works in AI-generated media detection (deepfakes) and the data infrastructure beneath it. The hard problem isn't running one classifier on one image. It's building a platform that keeps detecting as generators evolve, across modalities, at real customer volume. In a category crowded with demos and one-off scores, the strategic risk is being seen as a feature: a checkbox "deepfake score" rather than the trust-and-safety infrastructure layer that enterprises, platforms, and governments build on.
02 · Context and insight
The reframe that set the direction.
This is my flagship AI/ML advisory engagement, and the depth is the point. A decade across machine learning (search ranking at Preply and Letgo), a founded AI startup (Assist AI), and consumer web3 gives an unusual vantage: I read detection-model behavior the way an ML PM does, then frame it for buyers the way a consumer product leader does. My reframe: deepfake detection is not a score, it's an infrastructure problem. A score decays the moment a new generator ships. Infrastructure makes detection trustworthy over time and defensible as a business: robust data pipelines, multi-modal coverage, continuous refresh, clean integration points.
03, The approach
The decisions that mattered.
Reframe the category: detection as data infrastructure
The first move was narrative. Instead of leading with model accuracy (a number every competitor claims and every new generator erodes), I positioned Sensity as the data infrastructure layer for AI-generated media. Detection is the visible surface; the durable value is the pipeline, coverage, and continuous refresh underneath. This reframe shapes who the buyer is (platform and trust-and-safety teams, not analysts running one-off checks) and what the moat is (a system you build on, not a benchmark you cite).
Anchor positioning to the model's real failure modes
Positioning only holds if it's honest about where detection breaks. Drawing on ML PM experience, I grounded the narrative in actual model behavior rather than headline accuracy: adversarial robustness, generalization to unseen generators, false-positive cost in high-stakes contexts. The explicit tradeoff was optimizing the story for durable trust over a single benchmark number that looks good in a deck but ages badly.
Sequence the platform roadmap around defensibility
With the infrastructure thesis set, the platform strategy followed: build what makes detection sticky and hard to rip out. The lens was defensibility and integration depth over feature breadth. Data and pipeline investments compound; API and workflow surfaces embed Sensity into customer operations; broad features are easy to copy. The senior tradeoff: resist chasing every new deepfake type as a standalone feature, and instead invest in infrastructure that absorbs new threats as they emerge.
04 · How it's built
Close to the stack, not above it.
As an advisory engagement, this was strategy and product direction rather than hands-on shipping: positioning, platform roadmap, and narrative. The leverage came from pairing real ML literacy (reading detection-model behavior, failure modes, and refresh dynamics) with the ability to package it into a category story and a buildable roadmap.
The engagement reframed how Sensity positions itself, moving from a deepfake-detection point tool toward an AI-generated media detection and data infrastructure platform, and aligned the platform roadmap to that thesis. The relationship is still active, and the clearest signal of value is continuity: ongoing work, not a one-off audit.
What I’d carry forward
In deep-tech AI, positioning isn't decoration. It decides whether you compete on a benchmark you'll lose to the next model release or on infrastructure that compounds. The hardest part of advising an ML product isn't the model. It's keeping the narrative both ambitious and truthful about failure modes, because overclaiming in a trust-and-safety category is its own existential risk.