Skip to content
Abstract gold visualization of a layered cognitive architecture on a dark field. Not a medical scan.

Cognitive Brain

Building the

MahaaAi Cognitive Brain.

Perception, language, memory, evidence and safety are drawn as one architecture. Each stage below keeps the maturity the software actually supports.

Digital Nalanda
  1. 01

    Perception

    IN DEVELOPMENT

    Plant Health & Advisory shows observation and missing evidence. It is not a field diagnosis. Camera and voice perception inside the assistant are beta software, not a certified sensor.

  2. 02

    MPE

    RESEARCH

    An internal encoding and session component exists in the repository, including an offline research core and a shadow path. It is not a public model and its algorithm is not described here.

  3. 03

    Working memory / PNDN

    IN DEVELOPMENT

    The assistant can keep session-scoped dialogue memory for a signed-in beta conversation. That is not personal long-term memory. PNDN is named in internal notes and is not a public memory product.

  4. 04

    BIMR reasoning

    RESEARCH

    Named in research work outside this Production page. No public reasoning benchmark is claimed here.

  5. 05

    NPL

    RESEARCH

    A shadow guard can be enabled for a provisioned admin. It does not decide what a visitor hears, and it is off unless that flag is set.

  6. 06

    RAKSHA

    SIMULATION

    Fail-closed authorization before an autonomous physical action. Current evidence is a software safety gate in simulation. Physical validation is planned.

  7. 07

    RAMA-16

    SIMULATION

    Field-oriented target platform in the same concept family as CHANDI-18. Design and simulation stage. Not deployed hardware.

  8. 08

    MAITRI / SDK / robot

    FUTURE VISION

    MAITRI Learning is in development and is not a tutor. A public SDK is not released. Robot motion is not connected to these pages.

What the stack is responsible for

  • Multilingual understanding

    The assistant has language paths under evaluation, including Telugu and English voice copy. That is not a claim of equal quality in every language.

  • Evidence-bound memory

    Session memory is scoped to the conversation. It is not an archive of a person’s life, and these research pages do not read it.

  • Correction and deletion

    Account deletion exists for the product. Research pages do not add a separate memory editor.

  • Local-first computation

    Parts of the stack are designed to run without sending a turn to a paid model. That design is not a promise that every request stays on device.

  • Deterministic reasoning

    Some guards are deterministic checks. The assistant as a whole is not advertised as a deterministic theorem prover.

  • Safety governance

    RAKSHA and Somavati are the names of the fail-closed safety idea. A physical action still requires authorization that these pages cannot grant.

  • Behavioral response

    Conversation behavior is a beta product. It is not a clinical or psychological instrument.

  • Robotics integration

    CHANDI-18 and RAMA-16 are target platforms. Linking them to this architecture is research and simulation, not a factory interface.