Eight phases spanning three years — from foundational mathematics to a custom AI chip. Every phase builds on the last; no shortcuts, no skipped steps.
Developing the mathematics of computation — formalizing ternary vector algebra, proving its properties, and building simulations before writing production code.
Formal definition of the ternary vector space — presence (+1), absence (0), negation (−1) — and the fundamental operations: BIND, SUPERPOSE, GATE, PERMUTE.
Proof of associativity, commutativity, and distributivity for every operation. Property-based testing across millions of random vectors.
Formal bounds on dimensionality vs. capacity vs. noise tolerance — the numbers that set hardware requirements for every subsequent phase.
Rules for mapping structured data (records, sequences, graphs) into ternary vectors. Positional encodings and hierarchical composition.
Python implementation of the full algebra for experimental validation of mathematical claims. An executable specification that proves the theory correct.
Coding the core engine — translating the validated mathematics into a high-performance Rust runtime. Ternary vectors map directly onto machine words.
Zero-copy, SIMD-optimized implementation of all four algebraic operations. Ternary vectors packed into a compact bit-level representation.
Structure-of-Arrays layout for batch operations. Cache-line alignment. Benchmarks with linear scaling from 1K to 100K dimensions.
Native Python package via PyO3. NumPy-compatible. Researchers can call bind(a, b) from a Jupyter notebook.
Continuous benchmarks for throughput, memory consumption, and algebraic correctness. Every PR runs 10M tests.
Text modality, the Hardware Abstraction Layer (HAL), and a web interface for the text modality. The first perceptual channel: natural language.
Character-level encoding that maps token sequences into ternary vectors through positional BIND and hierarchical SUPERPOSE. Preserves word order, syntax, and semantic similarity.
Universal interface between raw modalities and the algebraic core. The HAL normalizes inputs from any modality into a single ternary vector space.
Multi-step reasoning over text: premise → GATE → conclusion. Each reasoning step produces a human-readable trace — the system explains why it reached a given answer.
Interactive web app: the user enters text queries and watches the reasoning engine work in real time. Visualization of vector similarity and reasoning traces.
Audio modality — sound as a first-class perceptual input. From raw waveforms to structured ternary representations.
Spectral decomposition into frequency bands, each mapped to positions in the ternary vector. Temporal structure through sequential BIND operations.
Real-time classification of audio streams: voice commands, environmental sounds, speaker identification. With full interpretability.
The first cross-modal demo: describe a sound in text — the system finds matching audio clips. Both modalities share a single vector space through the HAL.
Anomalous-sound detection in industrial and security contexts. The system explains which spectral features triggered the anomaly flag.
Visual modality — the most complex perceptual channel: images, video, spatial relations, and object compositions in ternary vector space.
Hierarchical encoding: local patches → regional features → global composition. Each level uses BIND for position and SUPERPOSE for merging.
Detection of objects and their spatial relations as structured ternary relations. "Red cup ON the table NEAR the window" — a single algebraic expression.
Extension to video by binding frames with temporal positional vectors. Actions and motion patterns become trajectories in ternary space.
Text, audio, and visual data in a single vector space. "Find the video where someone says this phrase near this object" — one similarity query.
Moving into implementation in the physical world — robots, smart homes, sensor networks. The same algebraic engine drives perception and decision-making in the real world.
Fusing data from cameras, microphones, lidars, temperature, and motion sensors into unified ternary representations. The HAL normalizes all physical modalities.
Object recognition, spatial mapping, and task planning on robotic platforms in real time. Every decision is auditable and explainable.
Voice commands, presence detection, sound analysis, and visual scene understanding — all on edge hardware. The home understands context, not just triggers.
The physical world demands safety guarantees. The reasoning engine evaluates actions against encoded constraints before execution — algebraic verification, not heuristics.
A marketplace for knowledge domains — a platform where experts publish, share, and monetize ready-made domain encodings, and users assemble custom reasoning systems from off-the-shelf components.
A standard "knowledge domain" format — an encapsulated set of encoders, reasoning schemas, and reference vectors for a specific subject area. A versioned registry.
Every published domain goes through automated validation: encoding consistency, reasoning correctness, capacity bounds, and conflict detection.
A catalog of knowledge domains with search, ratings, usage statistics, and dependency graphs. Preview a domain's capabilities right in the browser.
Flexible pricing models: free open-source domains, one-time purchases, and subscriptions to regularly updated knowledge bases. Transparent analytics for authors.
Developing a specialized AI chip — custom silicon that executes ternary vector operations natively. The algebra is the hardware. One tape-out, no second chances.
Native machine instructions for BIND, SUPERPOSE, GATE, PERMUTE, and similarity search. Each instruction operates on ternary words in a single cycle.
RTL implementation in Verilog/SystemVerilog. Full functional simulation confirming correctness against the software reference. Timing analysis.
A complete FPGA prototype before committing to ASIC. Real workloads from all previous phases run on the prototype to validate performance.
Final physical design, layout, and tape-out at a leading foundry. Physical verification (DRC/LVS), signal-integrity analysis, and power estimation.
SDK for integrating the MindPrism chip into hardware projects. Drivers, reference boards, documentation, and performance guides.
We're looking for research partners, early adopters, and domain experts. Tell us what interpretable AI should do for your field.