InnerEye

**Product and the concrete problem it solves.** InnerEye builds a "brain-computer interface meets computer vision" system that turns the split-second recognition responses of trained human experts into high-quality signal for AI models. The concrete problem is the visual-search bottleneck: intelligence analysts, radiologists, and security screeners must scan enormous volumes of imagery — satellite and aerial ISR frames, medical scans, X-ray baggage feeds — and the human eye/brain remains, for many hard targets, more sensitive than a pure algorithm, but far too slow at conventional viewing speeds and expensive to scale. InnerEye's answer is to present candidate imagery to an expert in a rapid burst while measuring the brain's involuntary "aha" recognition response, then combine that neural signal with a machine-vision score. The commercial packaging as of 2026 centers on two products, **Sense.I** (positioned to deliver higher AI model accuracy, reduced implementation cost, and, per the company, "up to 15x faster model learning speed" by harvesting expert brain responses as labels) and **Sense Plus** (assessing and developing the skills of human experts inside an organization). The through-line is InnerEye's "Minded A.I." thesis: capture tacit expert judgment that never gets written down, and transfer it into models "you can trust."

**Core technology and how it actually works.** The method is documented in InnerEye's granted U.S. patent **US10303971B2, "Image Classification by Brain Computer Interface"** (assignee Innereye Ltd; filed 2 June 2016, granted 28 May 2019; inventors Amir B. Geva, Leon Y. Deouell, Sergey Vaisman, Omri Harish, Ran El Manor, Eitan Netzer, and Shani Shalgi). The pipeline: (1) a computer-vision procedure scans an image and flags candidate regions "suspected as being occupied by a target"; (2) those candidates are shown to a human observer via **Rapid Serial Visual Presentation (RSVP)** — images flashed at roughly 5–10 Hz — while a multi-channel **EEG** headset records brain activity; (3) single-trial classifiers detect the neurophysiological event (a P300-style recognition response) that fires when the brain registers a target; and (4) the biological detection is fused with the machine-vision detection score to decide whether a target is present. The founders' peer-reviewed research (a multimodal neural network combining EEG and image branches, using RSVP on rapid **satellite imagery** to detect target structures) reported strong single-trial performance — on the order of 0.92 AUC and ~85% balanced accuracy for the supervised variant, with a semi-supervised autoencoder variant designed to remain robust to unknown target classes. Classifiers cited include Spatially Weighted Fisher Linear Discriminant and convolutional neural networks. The elegance is that the human never has to click or annotate; the label is read directly from cortical activity at speeds far above manual review.

**Market, customers, and go-to-market.** InnerEye targets three overlapping markets: (1) **defense and intelligence** — geospatial/ISR imagery exploitation, where analysts must find rare, camouflaged, or dispersed targets in oceans of overhead imagery; (2) **homeland security** — checkpoint, baggage, and cargo screening where speed and vigilance decay are operational risks; and (3) **civilian image-intensive verticals**, most prominently medical imaging (radiology triage) and industrial/visual inspection, plus a broader "expert-knowledge-into-AI" data-labeling value proposition that competes with conventional human-in-the-loop annotation. Go-to-market is a mix of direct engagement with defense and government end-users (its investor base and awards point to defense-agency access) and an enterprise-AI framing (Sense.I as a way to build better models faster). As of 2026 InnerEye does not publicly disclose named production customers or contract values, so the commercial footprint should be treated as demonstrated-and-piloted rather than broadly fielded; the company is small (public profiles indicate roughly 11–50 employees, with Startup Nation Finder citing about 27), which is consistent with a specialist deep-tech vendor rather than a scaled platform business.

**Traction, funding, and third-party validation.** InnerEye was founded circa 2013–2014 in Herzliya, Israel (85 Medinat Hayehudim St.), and public databases indicate it raised on the order of **$10M in early/seed-stage funding**; no large recent priced round is publicly documented, which is itself a diligence signal about pace and scale. The most credible third-party validation is a combination of (1) its investor and program roster — Startup Nation Finder lists backers including **The Chartered Group, Macnica, the Israel Ministry of Defense, the DRISHTI (Dual-use Robust India-Israel High-Tech Innovation) program, and the MIT Enterprise Forum of Israel** — which is unusually defense-aligned for a company of its size; (2) recognition as one of Frost & Sullivan's "10 Israeli Startups to Watch in Aerospace, Defense, and Security" and a runner-up in the U.S. Department of Defense / MIT Enterprise Forum "Combating Terrorism Technology Startup Challenge"; and (3) a granted U.S. patent plus peer-reviewed publications co-authored by the founders. The scientific substance is genuine; the commercial-scale evidence is thinner and dated, which is the central calibration point for this record.

**Founders and team background.** InnerEye's differentiation is rooted in an unusually deep neuroscience/signal-processing founding team. **Prof. Leon Y. Deouell** (co-founder, Chief Scientific Officer) is a cognitive neuroscientist at the Hebrew University of Jerusalem's Edmond & Lily Safra Center for Brain Sciences (ELSC), specializing in attention, perception, and EEG. **Prof. Amir B. Geva** (co-founder, CTO) is an expert in machine learning and biomedical signal processing (associated with Ben-Gurion University and Tel Aviv University research) whose lab work underpins the single-trial EEG classification methods. **Uri Antman** is cited as CEO, providing the commercial/operating leadership layer over the scientific core. The team's strength is authentic, publication-grade expertise in exactly the two disciplines the product requires — EEG decoding and computer vision — which is difficult to replicate. The corresponding gap, typical of scientist-founded deep tech, is scaled go-to-market and productization muscle, and the small headcount limits how many parallel deployments the company can support.

**Competitive dynamics.** InnerEye sits at an unusual intersection and therefore faces several partial competitors rather than one direct rival. (1) **Pure computer-vision / automatic target recognition (ATR)** offered by defense primes and ISR-analytics vendors (e.g., image-exploitation stacks used across the intelligence community, and Israeli players such as ImageSat/Cortica-style visual-AI) competes on the premise that algorithms alone will eventually be "good enough," pressuring the human-in-the-loop premium. (2) **BCI/EEG platform companies** (Emotiv, InteraXon, Neurable and similar) share the neural-sensing substrate but generally target consumer/enterprise wellness and interaction rather than defense-grade target detection. (3) **Human-in-the-loop data-labeling and active-learning platforms** (Scale AI, Snorkel-style weak supervision) compete for the same budget with the promise of cheaper high-quality training labels without EEG hardware. (4) **Radiology-AI vendors** contest the medical use case with algorithm-only triage. InnerEye's defensible edges are: (i) a patented, published fusion of RSVP-EEG with computer vision that neither pure-CV nor pure-BCI players replicate; (ii) the ability to extract expert judgment as labels at superhuman review speeds; and (iii) a defense/dual-use pedigree and access that generic AI-labeling vendors lack. The countervailing risk is that improving foundation-model vision narrows the accuracy gap that justifies putting a human (and an EEG cap) in the loop.

**Defense, security, and resilience dual-use relevance.** This is InnerEye's strongest and most credible dimension: the technology is dual-use by design rather than by adjacency. The same RSVP-EEG-plus-CV engine that could triage mammograms is directly applicable to (1) accelerating ISR/geospatial-intelligence exploitation — finding vehicles, launchers, tunnels, or infrastructure in overhead imagery faster and with fewer misses; (2) homeland-security screening at checkpoints, borders, and cargo/baggage inspection, where sustained vigilance is a known human failure mode; and (3) counter-terrorism and force-protection visual search. The defense alignment is not merely notional — the company's public backers reportedly include the Israel Ministry of Defense and the India-Israel dual-use DRISHTI program, and it placed in a U.S. DoD counter-terrorism technology challenge. Calibration matters: the dual-use relevance is inherent and validated by program participation, but publicly disclosed operational deployments and contract values are limited, so this is best read as a proven, defense-aligned capability with demonstrated interest rather than a broadly fielded, disclosed program of record.

**Growth stage, trajectory, and key diligence risks.** InnerEye is best characterized as a long-standing, small, science-led deep-tech company — early-to-mid stage in commercial terms despite roughly a decade of operating history — with strong IP and a distinctive dual-use wedge but modest public evidence of scale. The bull case: a patented, peer-reviewed, genuinely hard-to-copy fusion of neuroscience and computer vision; an authentic defense/homeland-security alignment; and a timely "expert-knowledge-into-AI" narrative as enterprises hunt for high-quality training signal. The bear case should dominate diligence: (1) **commercialization opacity** — no publicly disclosed anchor customers, contract values, or recent large funding round, and dated funding data; (2) **hardware/workflow friction** — requiring trained experts to wear EEG headsets and undergo RSVP sessions is an adoption barrier versus software-only tools; (3) **the closing CV gap** — as algorithm-only vision improves, the marginal value of the human-in-the-loop must be continually re-justified; (4) **scale constraints** — a ~27-person team limits parallel deployments and enterprise support; and (5) **verification gaps** — founding year (2013 vs 2014), exact headcount, funding total, and current customer status vary across public sources and require direct confirmation. Progression signals to watch: a disclosed defense/intelligence program of record, a named healthcare or screening customer, a fresh financing round, and independent benchmarks of accuracy/throughput uplift in operational settings.