PolyHelper/polyhelper: Trending on GitHub
PolyHelper — Cognitive AI Exoskeleton
Iron Man's suit and Jarvis — at the same time
In 2023, the founder's son was hit by a car — skull fractures, brain hemorrhage, surgery with metal implants. When it became clear that ordinary human capability is not enough, this exoskeleton was born — amplifying what a person can do when it matters most.
The son, a designer who couldn't work after the injury, put on this exoskeleton — and now designs this platform. The father, with no lawyer and no legal training, put on this exoskeleton — and has been standing against the world's top law firms and 10 Big Tech corporations worth $17 trillion, including Google, Apple, and Microsoft, for the past year.
Four corporations, including X.AI, have already settled.
And outpacing them in several AI directions — because this system can think, evolve, and grow on its own. It can pilot every physical exoskeleton model in the world, drive a car, and turn a laptop, smartphone, earbuds, glasses, or a watch into a cognitive AI exoskeleton.
This is what happens when someone builds an AI system because they actually need one.
Read the full story →
Act Beyond Your Limits
An industrial-grade engineering machine that helps people think, decide, and act in high-stakes environments.
10+ frontier AI. Independent governance. Every decision auditable, every outcome reproducible.
PolyHelper wraps around human cognition and supports five abilities: See, Hear, Act, Simplify, Remember. It monitors your mental load in real time and adjusts everything — when AI interrupts, how much autonomy it takes, how it delivers information. You don't configure it. It measures and responds.
Most AI tools answer questions. PolyHelper goes further: it keeps context across sessions, adapts to your state, verifies important decisions through 10+ independent AI models, and preserves a full audit trail. One verified answer — or ten diverse perspectives — or the best artifact from a tournament. Your call.
Three orchestration modes:
Convergent: 10+ AI models synthesize into one best answer for critical decisions.
Divergent: all unique ideas preserved, deduplicated, and ranked — for creative exploration, strategy, and research.
Artifact Council: competing branches generate, critique, debate, and refine documents and code through an 8-phase tournament. Instead of letting one AI produce one draft and hoping it is good enough, the system runs multiple competing versions, has them critique each other, preserves the strongest parts from losing variants, and only then assembles the best final result.
PolyGenesis continuously measures orchestration fitness, proposes controlled mutations, validates them on executable contours, rolls back unsafe changes, and reviews longer-term behavior through Meta-Cognition, Meta-Awareness, and Chairman-gated governance.
PolyGenesis functions as a governed self-evolving digital organism: a vendor-neutral runtime that senses its own weaknesses, mutates in controlled ways, preserves lineage and regret memory, and subjects autonomy to explicit governance rather than unconstrained self-modification.
Like SWIFT doesn't belong to one bank — PolyHelper is an independent arbiter across AI providers, with governed execution, budget-aware routing, and policy enforcement built into the runtime.
What is rare here is the combination: governed self-evolution, multi-model orchestration, user-facing product surfaces, and deployment across high-stakes domains within one vendor-neutral runtime.
The AI Orchestration Standard
Quick Start · Docs · Architecture · Orchestration Hub · Security · MCP
Verified on 2026-06-12
pnpm run build passes for apps/web
pnpm run verify:lint · verify:check · verify:web:release-gates pass
pnpm desktop:build:portable · desktop:verify:portable pass
pnpm run test:frontend passes (1,291/1,291 files, 28,859/28,859 tests)
repository runtime count: 208,483 tests / 8,425 files
mobile Jest runner: 547/547 suites, 5,147/5,147 tests
About this document. This README describes the platform architecture and capabilities. For verified build/test status and current runtime state, see Release Status and Release Artifact.
Contents
Who this is for
What it solves
The exoskeleton: how it works
Flagship domains
Quick Start
Why PolyHelper is different
System-Level Differentiators
Product capabilities
Architecture
Core OS + Domain Packs
Security
Who uses this
Benchmarks & Testing
Who this is for
If you are... Start here
Everyday User Quick Start · FAQ · Voice Setup
Accessibility Expert Accessibility Guide · Architecture
Automotive Partner Driver Solutions · Vehicle Control
Medical / Clinical AI Hospital Platform — 399K LOC, 87 modules, 37 specialty templates, 11 jurisdictions
Developer Developer Onboarding · Architecture
CTO / Architect Architecture · Orchestration · Trusted Execution · Agent Governance
Delegate / Crew Operator Digital Crew Quick Start · Overview
Robotics / Physical AI Robotics Hub · Physical Exoskeleton
What it solves
Enterprise AI has three unsolved problems:
Vendor lock-in — one provider = regulatory and business risk.
No audit trail — impossible to reproduce an AI decision during an incident.
Governance as optional middleware — most frameworks treat policy enforcement as a plugin, not a requirement.
PolyHelper solves all three by design. Governance is embedded in the runtime. The developer cannot skip it.
Use case snapshots: compliance‑ready policy memo · multi‑model code review · driver‑safety escalation · accessibility audit with WCAG evidence
"Hey Poly, ..." ──► AI Models ──► Chairman Synthesis ──► Verified Execution ──► Result ↑ ↓ ↓ ↓ Voice Assistant Diversity Matrix Quality Control Build & Run Loop
Example: "Generate a compliance memo" → 10 models analyze → Chairman synthesizes → verified run → final answer with evidence chain.
The exoskeleton: how it works
The word "exoskeleton" is used in two senses — like how "Google" means both the company and the search engine.
The entire platform is an exoskeleton. It wraps around the person, enhancing their cognitive abilities. It monitors your state (focus, workload, risk), adjusts the behavior of all features, decides when not to distract you, and regulates the autonomy of agents.
Inside the platform, there is a specific exoskeleton subsystem that makes this wrapping possible:
The brain — packages/core-services/cognitive-exoskeleton/ (16 modules): 13 signal sources fused into unified cognitive state, risk assessment, approval gates, feedback loop, metacognitive calibration, affective state detection, scaffolding with autonomy preservation
The muscles — packages/core-services/exoskeleton/ (37 modules): powered exoskeletons (ALICE, AirExo, German Bionic, Ekso, Laevo), 11 BCI adapters (Emotiv, Neuralink, Blackrock, Precision Neuro), FES systems, joint controllers, gait analysis, safety guardian — across 9 deployment profiles (clinical, industrial, field rescue, pediatric, ITAR, warehouse, elderly, sports, research)
The system operates across a full hardware spectrum. Software signals (typing dynamics, calendar pressure, topic-switching entropy, L1/L2 language load) provide immediate cognitive state estimation with zero hardware. Standard devices (microphone, phone sensors) add voice stress analysis. Consumer wearables (smartwatch, fitness band) add biometric signals. Specialized hardware (EEG headbands, smart glasses, haptic suits) adds neural and physiological channels. Medical and industrial hardware (BCI, powered exoskeletons, FES) enables full physical augmentation. Each layer amplifies the system — nothing is required, everything is additive. Every downstream component reacts automatically.
The exoskeleton makes users stronger over time, not dependent. The Cognitive Offloading Guard detects when users delegate tasks they could handle. The Scaffold-Fade Tracker reduces help as competence grows. This is an architectural constraint, not an aspiration.
Full architecture: Exoskeleton Docs · Physical Hardware · System Map
Governed Digital Workforce
The exoskeleton doesn't just help you think — it acts on your behalf.
Soloplex Digital Crew is a governed digital workforce: you define delegates that work across 50+ life and business domains — sales, teaching, research, ops, support, personal life. Each delegate has a persona, communication style, tools, and strict boundaries.
Delegates learn from observing the owner's real behavior (shadow mode), earn autonomy through measured performance, and are protected by 8 independent safety layers — from authority checks to budget limits. They work offline with local models, control physical devices through a 5-layer safety pipeline, carry litigation-grade legal accountability with cryptographic evidence chains, and operate within strict budget governance.
A solo founder running 6 delegates gets the output of a small team — and can stop everything at any moment.
Digital Crew Overview · Quick Start · API Reference · Safety
Flagship domains
Automotive, medical, accessibility, and legal are not four separate products. They are four proofs that the exoskeleton architecture works across high-stakes domains.
Accessibility
Accessibility is a foundational architectural layer, not an add-on. WCAG 2.2 AAA, 1.52M LOC cross-cutting, 60+ accessibility modules. Sign language avatars (ASL, BSL, LSF, DGS). Cognitive-adaptive UI that simplifies itself when the user is overloaded.
Accessibility Guide →
Automotive
Software-defined vehicle orchestration for both the machine and the human. CarPlay/Android Auto, 556K LOC, voice-first driving with 10+ AI consensus, 8-second emergency protocol, SirenSense (deaf driver support), predictive safety, fleet mission policy, simulation-before-commit for physical-world commands.
Two operating modes — the platform works with any vehicle:
Modern Mode — the vehicle has an OS and API (Tesla, BMW, CarPlay, Android Auto). Poly connects to the existing system and provides multi-AI consensus, predictive diagnostics, insurance scoring, AR navigation, and safety orchestration on top of the native stack.
Standalone Mode — the vehicle has no OS. Poly provides a phone-centered control stack: phone = compute, OBD adapter = diagnostics, cameras = vision, AR glasses = display, voice = primary interaction.
One platform — two modes — any vehicle. Two architectural layers: Vehicle Control (system plane) and Driving (driver experience).
Automotive → · Vehicle Control → · Driving →
🚀 Quick Start
V4 is a server-first platform. The backend is the platform. apps/web/ is one of N clients.
pnpm install pnpm start # backend + web UI on :5173 pnpm start:server # backend only (no UI — for headless/CI use)
AI works immediately — no API keys required. The server routes through managed free-tier providers automatically. Add your own keys or Ollama later. See Zero-Setup AI.
Desktop: ./PolyHelper.command (macOS) · ./PolyHelper.sh (Linux) · ./PolyHelper.bat (Windows). Packaged release: pnpm desktop:build:portable.
Mobile (iOS & Android): Full-featured React Native app (Expo SDK 55), ~1,283 screens. Offline-first with seamless cloud switching. Voice assistant, driving mode, biometric auth. Mobile Architecture →
📱 Mobile highlights:
Offline-first — works without internet. Heavy tasks offload to cloud. Seamless switching.
Voice assistant — tap-to-talk, auto-listen, driving mode integration.
Swarm orchestration — execute multi-agent swarms from your phone.
Driving mode — voice-only, 72dp touch targets, emergency button always visible.
Security — FaceID/fingerprint (HIPAA gate), certificate pinning, jailbreak detection.
~556K LOC — 278K platform-specific + 86K shared portable core.
More: Full Quick Start · Developer Onboarding · Cheat Sheet · Mobile Readiness
Self-Evolution (PolyGenesis) — opt-in
V4 includes a self-evolution system that can mutate parameters, generate code, score fitness, and roll back failures — all autonomously. Evolution is opt-in.
cp .env.example .env # add at least one API key (or run Ollama locally) pm2 start ecosystem.config.cjs # start the server (evolution is off by default)
Activate evolution when you're ready — via the "Continue Evolution" button in Settings, the one-time prompt that appears after a few minutes of use, or the API (POST /api/evolution/activation). Full guide: docs/EVOLUTION-SETUP.md
Why PolyHelper is different
Layer Innovation Advantage
Control Plane Swarm + Physical Gates Real-time cost/quality/risk optimization with physical constraint gates
Verification Consensus as Quality Control Divergent Consensus for creativity, Integrity Protocol for high-stakes
Execution Generate → Build → Run → Score Verifies outcomes by running real execution loops
Self-Evolution PolyGenesis Governed Mutation Loop Continuous self-improvement with validation, rollback, lineage, and periodic review governance
Reliability Production-Hardened OS 5-tier resilience, circuit breakers, budget controls, trace-based evals
Security & Trust Enterprise-grade Post-quantum crypto, 7-layer immutable audit, SOC 2 / HIPAA / ISO 27001
Experience Accessibility-First Voice-first, WCAG AAA, cognitive-adaptive UI
Width × Depth × Synergy:
WIDTH (10+ AI models)
┌───┬───┬───┬───┬───┬───┐
│ChatGPT│Claude│Gemini│Grok│...│
├─────────────────────────────────┤
D │ Round 1 │ E ├─────────────────────────────────┤ P │ Round 2 │ T ├─────────────────────────────────┤ H │ ... │ ├─────────────────────────────────┤ (4-10) │ Round 10 │ └───────────────┬─────────────────┘ ▼ ┌─────────────────────────────────┐ │ CHAIRMAN SYNTHESIS │ │ Best from all models │ └─────────────────────────────────┘
Approach Width Depth Synergy
Single AI ❌ ❌ ❌
Router (pick one) ❌ ❌ ❌
Chain-of-Thought ❌ ✅ ❌
Ensemble (voting) ✅ ❌ ❌
PolyHelper ✅ ✅ ✅
System-Level Differentiators
Unified Orchestration Stack — MCP-first tools + latent coordination in one runtime. 4 Power Levels (Solo→Duo→Squad→Full) with cost-aware routing and orchestrator-wide learning.
Verified Reliability — PPE + Chairman consensus + 5-tier resilience (server/hybrid/browser/satellite/local). Every tier is an execution engine, not just a fallback.
Privacy-Grade Autonomy — GUI autonomy with Safe Zones (no-see/no-act regions) + policy-grade guardrails. AI can act on your screen but cannot look at or touch areas you mark as private.
Global AI Governance — Jurisdiction-aware compliance across 10+ regions with Policy-as-Code (Cedar + OPA), 7-layer tamper-evident audit, and automated regulatory documentation.
Documentation-First Architecture — explicit, traceable paths for every layer. No undocumented magic.
How Self-Evolution Works
Hub-activated auto-evolution — PolyGenesis is activated through the hub and runs as part of the runtime, not as a disconnected lab feature.
Closed mutation loop — mutations are generated, captured, validated on executable contours, and rolled back on reject.
Lineage + regret memory — the system retains full ancestry for changes and blocks retries of similar failed approaches.
Weekly/monthly governance — Meta-Cognition, Meta-Awareness, Chairman-gated review, and user outcomes supervise long-horizon change.
Off-peak autonomous improvement — Dream Mode performs nightly self-improvement under safety limits instead of evolving blindly during peak user traffic.
External research — Dream Mode Phase 7 scans academic papers (arXiv, PubMed, Semantic Scholar), monitors AI model releases and benchmarks, associates findings with internal knowledge, and generates hypotheses — all budget-governed and safety-checked.
Cross-domain transfer — successful patterns can migrate across domains through federated genetics.
Living Architecture
PolyGenesis is architected as a governed digital organism — every subsystem maps to a biological function. The organism operates at four scales, just as the human body is simultaneously a mind, a cell factory, an internal ecosystem, and a member of a population:
🧠 Central Nervous System — mind & awareness
Biology System What it does
Brain CaptainOrchestrator + ChairmanConsensus Central nervous system — routes all decisions, synthesizes 10+ model opinions into one verdict
Heart MetaHeartbeatEngine + ChronobiologyEngine Pulse generator and circadian rhythm — drives the Day → Twilight → Night → Dawn lifecycle
Meta-cognition MetaCognitiveLoop Self-awareness: velocity tracking, capability assessment, prediction deltas
Memory EvolvingMemorySystem 4-layer memory (user/task/system/evolution) with consolidation, decay, and hygiene
Sleep & dreams DreamMode Nightly autonomous cycles — consolidates memory, explores mutations during idle
🧬 Genetic Machinery — evolution & protection
Biology System What it does
DNA OrchestrationGenome Versioned configuration genome — snapshots, diffs, rollback, full lineage tree
Epigenetics HyperGenesisEngine Meta-evolution — doesn't change genes, changes how genes are selected and expressed
Mutation EvolutionKernel 6-phase cycle: Observe → Measure → Mutate → Evaluate → Select → Commit/Rollback
Cambrian explosion CambrianExplosionEngine Mass diversification — thousands of mutations, intense selection keeps only the fittest 0.1%
Immune system EvolutionImmuneSystem 8 inviolable rules, antibody generation, SixPointGuard, constitutional firewall
Telomeres & apoptosis ApoptosisEngine Usage extends lifespan; neglect triggers archival to digital permafrost
🦠 Internal Ecosystem — micro-agents & collaboration
Biology System What it does
Microbiome MicrobiomeRegistry 5 micro-agent species: Prompt Bacteria, Token Fungi, Latency Virus, Cache Archaea, Log Decomposer
Symbiosis AutoSymbiosisEngine Multi-model collaborative debugging with shared diagnosis
Autopoiesis CognitiveAutopoiesis Self-creating system — components produce new components in a closed loop, fitness-gated reproduction
🌊 Population Dynamics — cross-instance & diversity
Biology System What it does
Pangenesis PangenomePool + GemmuleProtocol Cross-instance genome exchange via signed fragments with fitness and TTL
Co-evolutionary arms race AdversarialEvolution Red attacks, Blue defends, both evolve under mutual pressure — neither has fixed fitness
Coral reef CoralSearchEngine Multi-agent asynchronous search with shared persistent memory
Niche ecology EcologicalNicheSystem Population-level specialization and niche construction
This is a structurally consistent, mathematically grounded system — not decorative naming. CMA-ES drives continuous parameter evolution, Thompson Sampling selects gene-level strategies, MAP-Elites maintains quality diversity archives, and Bayesian prediction estimates clade potential. The organism senses its own weaknesses, mutates in controlled ways, retains lineage and regret memory, and subjects every change to governance — like biology, but with rollback.
Governed Autonomy Modes
OFF — no autonomous dream evolution.
SAFE_LOCAL — default when local safety prerequisites exist: synthetic validation, budget caps, and rollback.
FULL_GOVERNED — explicit higher-autonomy mode for fully governed operation; not enabled implicitly.
PolyGenesis is a governed self-evolving digital organism, not an unconstrained lifeform: autonomy is bounded by validation, rollback, budgets, and review governance.
PolyHelper focuses on applied domains: consensus orchestration, automotive safety, accessibility, medical AI, legal AI, cognitive exoskeleton, governed AI delegates, and AI governance. The integration of all of them into one system is the bonus, not the thesis.
Product capabilities
Chairman + Multi-AI Synthesis — 191 consensus methods, convergent + divergent modes, 19 life domains. →
Accessibility Stack — 1.52M LOC, WCAG 2.2 AAA, sign language avatars, SirenSense. →
Automotive AI — 556K LOC, voice-first driving, fleet mission policy, simulation-before-commit. →
Native Mobile — 1,283 screens, offline-first, voice assistant, driving mode, biometric auth. →
Governed Digital Workforce — 50+ domain delegates, shadow training, 8 safety layers, offline, legal accountability. →
Deep Deliberation — auto escalation solo → full, Lens-aware routing, budget-aware reasoning. →
Vibe Coding / Creative Factory — NL → product (UI + backend + DB + docs) across 83+ types. →
Code Security — multi-model consensus scanning, SBOM, SARIF, mandatory HITL, CI/CD. →
Selection Intelligence — 10 task profiles, shadow evaluation, learning loop, A/B rollout. →
AI Fluency (Trust Layer) — speed bumps, confidence maps, artifact receipts, fluency scoring. →
Immutable Governance — configurable 7-layer tamper-evident audit: L1 SHA-256 hash chain → checkpoint seals → HMAC signing → Postgres append-only → RFC 3161 TSA timestamps → OpenTimestamps Bitcoin-backed proofs → S3 Object Lock WORM. AI Passport — interoperable model provenance ledger ("who did what, when, and with what") with cross-organization verification. Evidence chain linking: execution → policy → consensus → shadow → outcome. →
Human Intelligence — risk-driven expert review, 5 lanes, RL-based matching, red teaming. →
Agent Governance — registry, DID identity, security graph, kill-switch, verified marketplace. →
Cognitive-Adaptive UI — dashboard tiles auto-simplify at high cognitive load. No market equivalent. →
AI Hospital Platform — 399K LOC, 87 modules, 37 specialty templates, 40 domains, 11 jurisdictions, FHIR/HL7/DICOM. →
PolySkill Plane — 39 slash commands, 11-phase sprint lifecycle, ML security pipeline. →
Token Economics — hard budget control, ROI analysis, safety-preserving downgrades. →
Workflow Engine — self-healing execution: convergence loops (3-5 retries with feedback), adversarial red-team verification, 200+ parallel agents with backpressure, crash recovery (durable persistence), adaptive budget (graceful downgrade, not hard stop), dynamic re-planning, auto disagreement resolution. →
DevLoop AI Engineering — one sentence in → structured issue → work plan → agent writes code → self-corrects on failure → code review → complete PR. 42 I/O-free modules, 71 API endpoints. →
PolyGenesis Self-Evolution — genome-based self-improvement with controlled mutation, executable validation, rollback, lineage/regret memory, Meta-Cognition, Meta-Awareness, Review Council governance, and daily review cadence. Start here → · → · White Paper → · Review Council →
Verdict AI Orchestration — protecting people (medical/legal/disability) and businesses (finance/compliance) via 10+ model consensus, reproducibility fingerprints, CriticAgent veto, Professional Packs (Legal, Finance, HR, Sales, Research), PII redaction, air-gapped local execution. Trusted Execution Spine: shadow execution (dry-run before commit), temporal consensus (multi-round drift detection with policy gates). 5 SKU tiers (Core → Enterprise Plus). Axiom Governance: legal-domain intelligence layer with 4 risk profiles, 19 governance modules, ethical walls, forensic evidence packs, and cross-border diligence workflows. → · Axiom →
Autopilot Fleet Execution — 4 Power Levels: Solo (1 model, $0.01–$0.30) → Duo (3 models + Chairman, $0.10–$1.50) → Squad (6 models + review, $0.50–$8.00) → Full/Autopilot (12+ models + deliberation, $2–$50+). FleetCommander mission orchestration: decompose goal → cost preflight → parallel dispatch via git worktree isolation → health monitoring → result merging → Chairman review → lessons extraction. 4-tier cost-aware routing (Free → Budget → Standard → Premium). LESSONS.md orchestrator-wide learning with Jaccard deduplication.
Sovereign AI Mesh Control Plane — productized national/critical-infrastructure orchestration with 3 profiles: National Critical Infrastructure, Regulated Public Sector, Regional Resilience Mesh. Jurisdiction-bound routing, domestic-only failover, and plan/validate flows. →
Privacy-Grade Autonomy — GUI autonomy with Safe Zones: mark screen regions as no-see/no-act, and AI respects the boundary. Policy-grade guardrails ensure the agent cannot bypass privacy zones even during autonomous workflows. →
Bento Dashboard — 69-tile modular grid with glassmorphic design, 3 density presets (Compact/Standard/Advanced), drag-and-drop reordering, 8 voice commands, mobile swipe carousel — all WCAG 2.2 AAA. The primary UI surface rendering consensus results, provider health, adaptive context, and 10 domain sections in one responsive view. →
Architecture
Nx-managed ESM workspace, five layers. Boundaries enforced by dependency-cruiser.
apps/* → libs/* → packages/server-* → packages/platform-* → packages/core-* → packages/core-os/*
Layer Lines Purpose
sdk/ ~1.11M 8 native SDKs (Go, Java, Kotlin, Python, Rust, Swift, C#, TypeScript)
packages/ ~5.76M Backend engines, platform, server, core-os, packs
libs/ ~1.90M Shared libraries — TypeScript + Svelte
apps/ ~503K Web + mobile + extension + electron clients
Strategic directions (cross-cutting):
Direction Lines Scope
Orchestration ~1.62M Multi-model consensus, chairman synthesis, swarm/fleet, debate, divergent/convergent modes, delegates
Exoskeleton ~643K Cognitive pipeline (state-engine, consciousness bus, scaffolding, affective), physical hardware (BCI, gait, EMG, digital twin), shared control, biomechanical safety
Self-Evolution ~320K PolyGenesis, genome, mutation, fitness, meta-cognition, dream mode, lineage, regret memory
9.41M lines of source code across 10 programming languages. 208,483 tests · 8,425 test files.
API + SDKs: OpenAPI specs, SDKs for 8 languages (1.11M LOC, 65 services, 800+ methods each). SDK Docs →
MCP: OAuth 2.1, REST + JSON-RPC/SSE, managed tool execution. MCP Docs →
PolyGenesis: governed self-evolution in V4 monitors orchestration fitness, mutates safely, validates changes on executable contours, retains lineage and regret memory, and subjects longer-horizon changes to periodic review governance. It operates as a governed self-evolving digital organism inside the runtime rather than as a standalone coding-agent add-on. Start here → · PolyGenesis Docs → · White Paper → · Review Council →
Example: Medical hallucination rate rises to 34% → PolyGenesis detects the deficit → generates a routing mutation (round-robin → quality-first) → validates through SixPointGuard → hallucination rate drops to 11%. Genome v42 → v43. If the mutation had failed, it would be rolled back, an antibody generated, and the mistake recorded in regret memory — never repeated. Full walkthrough →
This makes V4 more than a self-improving coding agent: it is a governed self-evolving orchestration system that operates across real product surfaces and regulated, safety-critical domains.
Full Architecture → · Testing Guide → · Deployment →
Core OS + Domain Packs
Core OS provides AI orchestration, governance, and the cognitive exoskeleton loop. Domain Packs add specialized modules:
🎯 AI Orchestration (Core OS) ♿ Accessibility-First (Pack) 🚗 Driver Safety (Pack) 🔒 Code Security (Pack)
Frontier AI, 191 canonical methods WCAG 2.2 AAA Voice-first, distraction gating Code security pipeline + HITL approvals
Chairman + Polly synthesis 1.52M LOC (cross-cutting) 556K LOC (cross-cutting) Audit trails + DPA gate
Deep deliberation (AUTO 4→10) Haptics/Braille output 10+ dashboard tiles, SirenSense CI/CD SARIF + policy gates
Consensus Engine Accessibility Guide Driver Solutions Security Hub
Domain & Enterprise:
Domain Capability Guide
RAG & Documents Layout-aware ingestion, citation validation, cross-encoder reranking Smart RAG
Finance Portfolio tracking, news sentiment, budget management, fraud detection Finance Agent
Creative Factory 83+ product types, Divergent Consensus, voice wizard Creative Agent
Engineering CAD NL → OpenSCAD → STL/SVG, Design Twin Loop with PPE Engineering CAD
Learning Spaced repetition, adaptive testing, AI tutoring, streaks Learning Agent
Enterprise Admin Agent Control Plane, Tool Gateway, Connectors, Audit Chain Enterprise Admin
Trusted Execution Shadow exec, evidence chain, enterprise packages, /ops admin Trusted Execution
Orchestration Modes (Core OS):
Lens (West/East) — bloc-aware comparisons + agreement/disagreement map + side-by-side. Lens Council →
Deep (5×5 / 10) — higher-reliability orchestration (more diversity, stronger cross-checking).
Compact (3) — fast fallback for speed/cost (3 models + chair).
Security
Governance-by-default: the system will not execute a request without passing the full policy chain (Cedar + OPA pre/in/post-execution gates). The developer cannot skip governance — it is embedded in the runtime.
Post-quantum crypto. 7-layer immutable audit. Automated compliance evidence packs. Mandatory human approval checkpoints.
Consensus engines: Debate engine + Chairman loop + hybrid routing + prompt evolution.
Security Hub → · Trusted Execution →
Who uses this
Audience Why it fits
Everyday Users Voice-first AI with multi-model consensus. BYOS — $0 extra cost.
People with Disabilities WCAG-first UX, 60+ accessibility modules, multimodal assistive outputs.
Drivers Hands-free AI with safety protocols, SirenSense, CarPlay/Android Auto.
Solo Founders & Teams Governed digital workforce — 50+ domain delegates at a fraction of team cost.
Government, Defense & Compliance Sovereign, air-gapped AI with automated compliance evidence. 12+ regulatory frameworks.
Developers & Labs Multi-model experimentation, consensus benchmarks, reproducible runs.
Benchmarks & Testing
208,483 tests · 8,425 files · dual framework (Jest + Vitest) · Playwright E2E
Benchmark Protocol · Performance · Security Benchmark
Testing Guide · UAT Process
SDKs: Python, JavaScript/TypeScript, Java, C#/.NET, Go, Swift, Rust, Kotlin — 1.11M LOC, 65 services, 800+ methods each. SDK Docs →
Deployment: Docker · Kubernetes · Helm · Full Guide
Frontier AI Models: Claude · GPT · Gemini · Grok · DeepSeek · Qwen · Mistral · Kimi · Ernie · GLM · Meta AI Compatible with 1000+ endpoints. Model Catalog →
🏭 🤖 🧠 🧪 ♿ 🔒 🌍
58,453 10+ AI 191 208,483 AA / AAA Post-Quantum 250+ i18n
Code Files Providers Algorithms Tests WCAG Security Languages
Full Documentation → · Contributing · Licensing
Responsible Use: PolyHelper is designed for beneficial applications. We prohibit use for generating harmful content, spreading misinformation, or violating privacy. RESPONSIBLE-USE.md
What You Get for Free
PolyHelper is an open-core project. Most of the platform is free and open source.
Open Source (Apache 2.0) — 7.75M lines of code, free for any use, including commercial:
Accessibility platform — 1.52M lines of code, WCAG 2.2 AAA, sign language avatars, haptic feedback, braille output, 60+ modules
8 Native SDKs — Go, Java, Kotlin, Python, Rust, Swift, C#, TypeScript (1.11M LOC, 65 services, 800+ methods each)
MCP integration — OAuth 2.1, REST + JSON-RPC/SSE, managed tool execution
Internationalization — 250+ languages, RTL support, cultural formatting
Desktop application — Electron for macOS, Linux, Windows
Mobile application — React Native for iOS and Android, offline-first, 1,283 screens
Bento Dashboard — 69 tiles, glassmorphic design, drag-and-drop, voice commands, WCAG AAA
PolySkill command framework — extensible skill system
FHIR / HL7 / DICOM interoperability layer — healthcare data standards
Post-quantum cryptography — lattice-based primitives, hybrid key exchange
Basic consensus runner and workflow executor
Basic memory layer and audit logging
Basic RAG pipeline
Demo applications and examples
Full public documentation, architecture guides, and contributor docs
Source-Available (BSL 1.1) — 1.11M lines of code, free for non-commercial, educational, and research use:
Full orchestration engine (191 consensus methods, Chairman synthesis)
Selection Intelligence, FleetCommander, Deep Deliberation
Workflow Engine, Evolving Memory, AI Fluency and Trust Layer
Human Intelligence Layer, 7-Layer Immutable Audit Trail
Lens Council, Coral Search Engine
Code Security Pipeline, Enterprise Admin, Token Economics
Converts to Apache 2.0 on June 13, 2033. See LICENSING.md for full details.
Academic and Research Access
We provide free access to advanced modules for universities, research labs, and non-profit organizations.
Orchestration engine internals and consensus algorithms
Self-evolution (PolyGenesis) architecture
Safety governance frameworks
Any module needed for your research
No complex legal process — just a simple research agreement. Cite PolyHelper in your publications.
Research Access → · Contact: info@polyhelper.ai
