Blog

Long-form notes, build logs, and product diaries.

01 / Feb 17, 2026the journey of building loqlA look at building LoQL during college and the lessons learned.Open article02 / Feb 15, 2026my thesis on ai-generated image verification gatewaythe democratization of generative ai has driven the marginal cost of creating hyper-realistic synthetic media to zero, dismantling the "human verification" layer that global commerce and security rely on. this white paper argues that the solution is not a better detector, but a new infrastructure layer: the ai integrity gateway. by fusing deterministic provenance (c2pa) with probabilistic forensic ensembles, this middleware architecture serves as the necessary "trust os" for the modern internet, transforming reality verification from a manual bottleneck into an automated, scalable api. 1. the crisis: the "zero marginal cost" of deception for the last twenty years, digital trust was predicated on a simple heuristic: proof of work. creating a convincing fake image, a forged passport scan, or a synthetic voice required time, technical skill (e.g., photoshop expertise), and expense. this friction acted as a natural firewall against fraud at scale. generative ai has removed this friction. in 2026, the marginal cost of generating a photorealistic image or cloning a ceo's voice is effectively zero. • scale: bad actors can now generate thousands of unique, synthetic identity documents per hour for kyc (know your customer) fraud attacks. • sophistication: we have moved beyond "deepfakes" with visible artifacts. modern diffusion models produce imagery that is statistically identical to reality in the spatial domain, often bypassing traditional visual inspection. • the "liar's dividend": the mere existence of high-quality fakes allows bad actors to dismiss real incriminating evidence as ai-generated, creating a dual crisis of fraud and plausible deniability.    the result is a collapse of the "human firewall." humans are evolutionarily conditioned to trust their eyes, identifying deepfakes with accuracy rates hovering near 50-60%—functionally equivalent to a coin toss. we cannot hire enough humans to solve this problem, nor can we trust their judgment. 2. the architectural thesis: why "detectors" fail but "gateways" succeed the market is currently flooded with single-point "deepfake detectors"—simple classifiers trained to spot specific artifacts. these are destined to fail because detection is an adversarial game; as soon as a detector is released, generators are trained to bypass it. to build a defensible unicorn in 2026, we must stop building "detectors" and start building infrastructure. the concept: the ai integrity gateway our proposed solution is a middleware gateway—an intelligent "router" for digital media that sits between the outside world (users, internet) and the enterprise core. just as an api gateway manages traffic and authentication, the ai integrity gateway manages truth. this gateway enforces a "defense in depth" strategy with three distinct verification layers: layer 1: deterministic trust (the provenance check) before running expensive gpu compute, the gateway checks for cryptographic signatures. using the c2pa (coalition for content provenance and authenticity) standard, we verify the "chain of custody." if an image was taken by a trusted device (e.g., a leica m11-p or a specialized banking app) and signed, it is whitelisted. this is deterministic verification—math, not guessing.    layer 2: probabilistic forensics (the ensemble engine) if provenance is missing (which applies to 99% of current web content), the media is routed to an ensemble of experts. relying on a single model is dangerous. our architecture orchestrates multiple detection methodologies simultaneously: • dire (diffusion reconstruction error): reverses the image diffusion process to spot inconsistencies in "noise" that the human eye cannot see. • frequency analysis (dct/fft): analyzes the image in the frequency domain to find the "grid-like" fingerprints left by upsampling algorithms. • biological liveness (ppg): for video, analyzing sub-perceptual changes in skin color caused by a beating heart—something generative models struggle to simulate accurately. layer 3: semantic reasoning (the "why" layer) we utilize multimodal llms (mllms) as reasoning agents. unlike pixel-based detectors, these models analyze context. they answer questions like: "is the lighting on the subject's face consistent with the background sun position?" or "does the text on the street sign contain the gibberish characteristic of ai hallucinations?" this layer provides explainability, ensuring the user understands why an image was flagged. 3. the market opportunity: regulatory tailwinds and financial fear the catalyst for this startup is not just technology; it is liability. • the financial stick: in 2024, a single deepfake video call cost a multinational firm $25 million.. cfos are realizing that "reality security" is now a balance sheet item. • the regulatory hammer: the eu ai act (article 50 & 52) and california's sb 942 mandate transparency and labeling for ai-generated content.. enterprises must now prove they have taken "reasonable measures" to detect and label synthetic media or face massive fines. • the adoption curve: we are transitioning from "early adopters" (media/news) to "early majority" (banking, insurance, identity). the deepfake ai market is projected to grow at a 42.8% cagr, reaching $7.2 billion by 2031. 4. the path to capital: why we are investable investors in 2026 are skeptical of "wrapper" startups. they are looking for systemic trust. • the data moat: we are not just training on public datasets (like genimage or artifact).. we are building a proprietary feedback loop of real-world attacks from high-risk sectors (fintech/insurance), creating a dataset of "adversarial fakes" that academic models never see. • the unit economics: by using a "gateway" approach, we filter low-risk traffic cheaply (via c2pa or lightweight hashing) and only spend high-cost gpu inference credits (h100s) on high-risk anomalies. this preserves margins in a way that brute-force detection api competitors cannot. • the vision: we are not building a tool to "catch fakes." we are building the ssl for reality. just as the green padlock icon became the standard for secure web browsing, the ai integrity gateway will become the standard for secure media consumption. 5. conclusion the era of "passive trust" is over. we are entering an era of "zero trust media." the organizations that survive this transition will be those that treat media authenticity as a core infrastructure requirement, not a feature. we are building the infrastructure to secure that future.Open article