Bioadaptive AI — somatic interfaces

    AI that adapts to the person — not only the prompt.

    Since 2014, Hannes Bend has developed biofeedback artworks and research prototypes that led to granted adaptive-interface patents and a portfolio of ventures. The work explores how measured signals, context and explicit feedback can adapt screens, audio and responsive environments.

    Experience the Application Lab Explore the evidenceDiscuss an application

    Patent drawings

    What the patents actually disclose

    Examples disclosed in US 10,423,893 B2.

    The patent describes intended system behaviour and technical scope; disclosure does not show that the described outcomes were demonstrated, clinically validated or effective in a product.

    Outputs — Figures 14–16

    Brightness, volume and odorant-level personalization

    US 10,423,893 B2, Figure 14: patent flowchart depicting disclosed brightness-level personalization from user data through an adaptive interface to display output.
    FIG. 14 — Brightness-level personalization
    US 10,423,893 B2, Figure 15: patent flowchart depicting disclosed volume-level personalization from user data through an adaptive interface to audio output.
    FIG. 15 — Volume-level personalization
    US 10,423,893 B2, Figure 16: patent flowchart depicting disclosed odorant-level personalization from user data through an adaptive interface to olfactory output.
    FIG. 16 — Odorant-level personalization

    Devices — Figures 17–19

    Mobile interface, headphones and artificial olfaction

    US 10,423,893 B2, Figure 17: patent drawing of four mobile-interface examples beneath a system diagram for customizing output based on user data.
    FIG. 17 — Mobile-interface examples
    US 10,423,893 B2, Figure 18: patent drawing of four headphone outputs with waveform and volume symbols beneath an adaptive-interface system diagram.
    FIG. 18 — Headphone outputs
    US 10,423,893 B2, Figure 19: patent drawing of four artificial-olfaction output examples beneath an adaptive-interface system diagram.
    FIG. 19 — Artificial-olfaction outputs

    Loop — Figures 20–22

    Camera capture, analysis and continuously adapted output

    US 10,423,893 B2, Figure 20: patent drawing of camera capture, facial region-of-interest selection, RGB channels and signal-processing steps.
    FIG. 20 — Capture and signal processing
    US 10,423,893 B2, Figure 21: patent drawing of signal plots and listed stages for signal extraction, heart-rate and respiratory-rate estimation, adaptive modeling and adaptive implementation.
    FIG. 21 — Analysis of captured data
    US 10,423,893 B2, Figure 22: patent drawing of a sequence of mobile-interface outputs connected to adaptive-interface, programming and computing components.
    FIG. 22 — Continuously adapted output

    Thumbnails are reduced for page weight. Select any drawing to open the full-resolution plate, where the figure numbers and the US 10,423,893 B2 identifier are legible.

    The full record is available at US 10,423,893 B2.

    Architecture

    How the loop works — and where measurement can fail

    Closed-loop diagram of a bioadaptive interface with a signal-quality and uncertainty gateA consented signal and its context enter a signal-quality and uncertainty gate. On the USABLE branch the loop continues to a personal baseline and context, then a bounded reversible adaptation, then observing the response and requesting feedback, and finally learning, reverting or doing nothing, which returns to the start. On the UNCERTAIN branch the loop terminates in doing nothing or asking, so no adaptation is applied when confidence is low.1Consented signal + contextA measured signal — for example apulse-rate estimate2Signal-quality and uncertainty gateIs this measurement good enough to act on?USABLEUNCERTAIN3Personalbaseline + contextCompared with this person,not a population averageDo nothing or askThe loop stops here.No adaptation is applied.4Bounded reversible adaptationOne small change, inside fixed limits,always undoable5Observe response + request feedbackKeep this change, or restore the previous state?6Learn, revert or do nothingThe personal baseline updates either wayThe updated baseline re-enters the loop

    A bioadaptive interface should not jump from a signal to a conclusion. It checks signal quality, makes one bounded and reversible change, and asks whether that change should be kept.

    Four words are used identically here and in the Application Lab, with the same colours: Signal, Quality check, Response, Feedback.

    A pulse-rate estimate is a signal. Stress, focus or emotion are higher-level interpretations requiring separate validation. When confidence is low, the system should pause rather than guess: the gate has two branches, and one of them does nothing.

    A measurement problem worth naming

    Illumination and feedback

    Camera-based estimation depends on the light reaching the skin. Screen-driven illumination changes, mixed or flickering light sources and low light all degrade the recovered signal — and an adapting screen can itself alter the illumination it is measuring under.

    Motion

    Head movement, speech and posture shifts introduce artefacts that are difficult to separate from the underlying pulse waveform.

    Camera pipeline

    Compression, auto-exposure, auto-white-balance, frame-rate variability and sensor differences all change what reaches the algorithm. The same person on two devices is not the same input.

    Demographic and dataset limits

    Published evaluations report performance differences across skin tone and demographic groups, and widely used public datasets are themselves demographically unbalanced. Performance measured on one population does not transfer automatically to another.

    • Liu, Wang and Wang (2014) — “The Effect of Light Conditions on Photoplethysmographic Image Acquisition Using a Commercial Camera,” IEEE Journal of Translational Engineering in Health and Medicine 2, 1–11 DOI · Open access
    • Nowara, McDuff and Veeraraghavan (CVPRW 2020) — skin tone and gender meta-analysis CVF paper · DOI
    • Dasari et al. (npj Digital Medicine 2021) — evaluation of bias in rPPG measurements Nature · Open access
    • Bondarenko, Menon and Elgendi (npj Digital Medicine 2025) — demographic bias in public rPPG datasets Nature

    Boundary: Higher-level inferences are treated as uncertain and never diagnostic.

    See this loop in action in the Application Lab — directly below.

    Application Lab

    Screen, voice and room — made concrete

    Interactive simulation

    Watch an interface reduce friction automatically.

    Choose a simulated signal scenario. The Lab checks its quality, makes one small reversible response across screen, voice or room, and asks whether to keep it. Nothing is measured from you.

    The Lab illustrates the closed loop described in the granted adaptive-interface patent family and developed through the research since 2014. It is an illustration of the loop, not an implementation of the patents.

    Simulation only · No sensing

    Interactive simulation

    Application Lab

    1. SignalSteady and unremarkable during focused work74 bpm · 12 breaths/min · simulated · Focused work
    2. Quality checkGood enough to act onadaptation permitted
    3. ResponseThe full reply is shown, with everything at onceBalanced lines · balanced density
    4. FeedbackWas that right?keep it or undo it — your choice
    Before—

    Can you walk me through what I need to decide today?

    There are three decisions open right now.

    The subscription renewal closes today. The vendor review is due Thursday, and the team's travel budget needs a decision before the end of the month.

    You can confirm the renewal, defer it by a week, ask for the vendor comparison, or reassign the travel decision.

    Confirm renewalDefer a weekSend comparisonReassign travel

    Nothing has changed yet. Start the simulation to watch one response apply itself.

    Explore the controls

    Optional. The demonstration above happens without touching anything.

    Simulated signal scenario

    Inputs are simulated pulse, simulated breathing, the context you select above and your explicit feedback. Light, sound and scent are responses, never measured inputs. Nothing is measured from you. One illustrative response—not a prediction of what a person needs.

    Ventures

    Four ventures from one bioadaptive research trajectory

    TransformingAI

    Research and commercialization work around adaptive-interface systems.

    Breathing AI

    Guided breathing, biofeedback and screen-mediated product work.

    VitalSign AI

    Camera-based physiological signal work and its developer-facing surface.

    OptimizingAI

    Applied work on interfaces and environments that respond to measured context.

    See all ventures

    Evidence ledger

    What is disclosed, what is researched, what is validated

    The rows below concern US 10,423,893 B2. The patent describes intended system behaviour and technical scope; disclosure does not show that the described outcomes were demonstrated, clinically validated or effective in a product.

    Application-level evidence matrix separating patent disclosure, independent evidence and product validation for each bioadaptive application.
    ApplicationPatent disclosureIndependent evidenceProduct validation / safe claim
    Screen and typographyUS 10,423,893 B2 discloses changes to font, brightness, contrast, colour and screen-interface output. FIGS. 14 and 17 illustrate brightness and screen-interface adaptation. This establishes disclosed technical scope, not demonstrated efficacy.AdaptiFont found reading-speed gains from individualized font adaptation in a non-biometric study. This supports personalization, not physiology-triggered typography.No product validation has been supplied for physiology-triggered typography. Individualized and user-controlled readability are supported research directions; biometric triggering remains experimental.
    Voice and audioUS 10,423,893 B2 discloses changes to volume and other audio characteristics, including pitch, tone, bass and treble. FIGS. 15 and 18 illustrate audio and headphone adaptations. This establishes disclosed technical scope, not demonstrated efficacy.AdaptiveVoice reports task-specific effects from adapting voice characteristics in a non-biometric study. It does not validate pulse- or breathing-triggered voice adaptation.No product validation has been supplied for pulse- or breathing-triggered voice delivery. A biometric extension remains experimental; a demonstrated user benefit has not been established.
    LightingUS 10,423,893 B2 discusses brightness and illuminance in lux, and FIG. 14 illustrates brightness adaptation. The record does not establish product validation for adaptive room lighting.The CIE supports integrative lighting recommendations that account for time, context and user needs. This is not validation of physiology-triggered lighting.No product validation has been supplied for biometric lighting. Current evidence is stronger for adaptation based on time, task, accessibility, explicit preference and user control.
    OlfactoryUS 10,423,893 B2 discloses odorant output and artificial olfaction. FIGS. 16 and 19 illustrate those applications. This establishes disclosed technical scope, not demonstrated efficacy.Holloman and Crawford review olfactory-display technology, use cases and constraints. Their review does not establish efficacy for a bioadaptive olfactory product.No product-specific outcome validation has been supplied. Any implementation remains experimental, opt-in, exposure-aware, immediately stoppable and off by default.
    rPPG pulse estimateUS 10,423,893 B2 depicts image capture, analysis and continuously adapted output in FIGS. 20–22, and its specification discusses pulse and respiration sensing. This establishes disclosed scope, not a universal measurement claim.Camera-based pulse estimation is supported as a measurement method under constrained conditions. Published studies document sensitivity to motion, lighting, camera pipelines, skin presentation and dataset demographics. Nowara et al. (2020) · Dasari et al. (2021) · Bondarenko et al. (2025)Validation is device-, lighting-, motion-, compression- and population-specific. No single accuracy figure transfers across settings. No product-level validation package has been supplied for publication on this page.
    Emotion, discrete stress and blood pressureUS 10,423,893 B2 mentions stress level and blood pressure among possible measurements or inferences. Patent disclosure describes intended system behaviour; it is not evidence of validated emotion, stress or blood-pressure measurement.Siegel et al. (Open-access copy) and Barrett et al. (Open-access copy) document limits on inferring discrete emotion categories from physiological signals. Their findings come from an overlapping research programme and do not constitute independent convergence from unrelated programmes. Blood-pressure accuracy requires separate product- and population-specific validation.This work does not claim detection of emotion categories or discrete stress states, and it does not claim blood-pressure measurement. Higher-level inferences are treated as uncertain and non-diagnostic.

    Research and IP

    Primary records

    A related adaptive-interface patent portfolio originating from 2017 priority work includes two US grants and one Chinese grant. Hannes Bendfeldt is named as inventor. Patent status and assignee information should be confirmed in the relevant official registers for legal or licensing decisions.

    Lineage

    It started in the artwork

    The sensing and adaptation logic came out of installations and immersive works made with participants in museums and public space, where the body was always the input. See the artwork archive or the chronology.

    Definition

    Somatic AI, defined after the evidence

    Somatic AI names a category of systems whose primary input is the physiological and behavioural state of the human body, and whose primary output is a bounded adaptive response calibrated to that state. Bioadaptive is used here as an adjective: the work adapts screens, audio and environments. The category argument lives on its own page — what is Somatic AI? — alongside the mechanism explainer at bioadaptive interfaces.

    Next

    Build, license or investigate an application

    Advisory and consulting

    Design, claim review and technical direction for bioadaptive systems.

    Research collaboration

    Joint studies on sensing, uncertainty handling and bounded adaptation.

    Investment and ventures

    Conversations about the venture portfolio and its research base.

    Speaking

    Keynotes and workshops on adaptive interfaces and honest claims.

    Art and exhibitions

    Exhibition, commission and collection enquiries for the artwork lineage.

    Start a conversation