Skip to content

ForgedEmir/Parralax

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

PARALLAX

PARALLAX

The autonomous reality reconciliation layer

AI agents can automate digital workflows.
But who verifies the physical world?

Built for Hermes Hackathon TypeScript React MCP License: MIT

#vision-ai #reality-reconciliation #mcp #hermes-agent #automation #warehouse #quality-assurance #computer-vision #deterministic


The problem

Business software records what should have happened.
Cameras see what actually happened.
Nothing connects the two before an agent is allowed to continue.

A wrong item ships. A missing safety seal passes. A job that was never done gets marked complete. PARALLAX is the verification layer that forces an agent to wait until physical reality matches the plan.


Hermes is the brain

PARALLAX does not make decisions autonomously. Hermes Agent is the orchestrator — it discovers the available MCP tools, selects the right protocol for the job, drives the full inspection-repair-release loop, and signs every action it takes.

┌─────────────────────────────────────────────────────────────────┐
│                        HERMES AGENT                             │
│                                                                 │
│   "What protocol applies here?"                                 │
│   "Run the inspection."                                         │
│   "Read the repair directive."                                  │
│   "Verify the corrected evidence."                              │
│   "Release when every invariant passes."                        │
│                                                                 │
│   ↓  7 MCP tools  ↓                                             │
└───────────────────┬─────────────────────────────────────────────┘
                    │
                    ▼
┌─────────────────────────────────────────────────────────────────┐
│                        PARALLAX ENGINE                          │
│                                                                 │
│   Vision model → structured observation                         │
│   Deterministic policy → HELD / RELEASED                        │
│   Idempotent actions → receipts + ledger                        │
└─────────────────────────────────────────────────────────────────┘

Hermes decides what to do. PARALLAX enforces whether it is safe to do it.


How it works

flowchart TD
    A[Hermes Agent] -->|compile_protocol| B[Operation Protocol]
    B --> C{Business Record\nWhat must be true}
    D[Camera / Phone] -->|visual evidence| E[Vision Model\nQwen3-VL · Nemotron]
    E -->|structured observation| F[Deterministic Policy]
    C --> F
    F -->|all invariants pass| G[RELEASED\nActions executed]
    F -->|violation detected| H[HELD\nRepair directive created]
    H -->|Hermes reads directive| A
    H -->|fresh evidence| E
    G --> I[Hash-chained Ledger\nTamper-evident proof]
    G --> J[Idempotent Action Receipts\nInventory · Notify · Release]
Loading

The model perceives. Policy decides. Hermes acts.

Layer Who does it What it does
Perception Vision model (Qwen3-VL / Nemotron) Reads the camera, extracts visible facts
Decision Deterministic code Compares facts to invariants, no LLM vote
Orchestration Hermes Agent via MCP Drives the loop, signs every action
Proof SHA-256 ledger Every transition is independently verifiable

The LLM never decides whether a workflow continues. Deterministic invariants own that gate.


Demo scenarios

Warehouse fulfillment pack-out

A package must contain a red ceramic mug, a USB-C cable, and a warranty card before it ships.

Evidence Vision sees Policy decides
First photo Blue mug · No warranty card HELD — 2 violations
Corrected photo Red mug · Cable · Card RELEASED
Damaged photo Red mug cracked HELD — condition mismatch

HVAC field-service closeout

A technician installs a filter but forgets the safety seal and service label.

Evidence Vision sees Policy decides
First photo Filter only HELD — 2 missing controls
Corrected photo Filter · Seal · Label RELEASED — visit closed

Custom reference scenario (any workflow)

Take a photo of anything. PARALLAX learns the manifest automatically — no predefined objects, no code. Upload a second photo and it detects any deviation in real time.


Quickstart

pnpm install
pnpm dev

Open http://localhost:5173 — the full demo runs in fixture replay mode with no API key.

Enable live vision

Copy .env.example to .env and add one key:

# Option A — Hugging Face (free)
HF_TOKEN=hf_...
HF_VISION_MODEL=Qwen/Qwen3-VL-30B-A3B-Instruct

# Option B — NVIDIA NIM
VISION_BASE_URL=https://integrate.api.nvidia.com/v1
VISION_API_KEY=nvapi-...
VISION_MODEL=nvidia/nemotron-3-nano-omni-30b-a3b-reasoning

Run the Hermes autonomous demo

pnpm demo:hermes

Hermes discovers the 7 MCP tools, drives 4 inspection rounds autonomously, and returns a signed operational report.


MCP tools exposed to Hermes

Tool Description
get_operation_state Read protocol, manifest, status, metrics
compile_protocol Activate a visual operation protocol
inspect_fixture Run a verified visual fixture
get_evidence_history Full inspection history with SHA-256 digests
get_repair_directive Structured correction plan with steps
verify_ledger Independently verify the hash-chained ledger
reset_operation Reset to clean state

Architecture

├── server/
│   ├── index.ts          Express API · SSE broadcast · file upload
│   ├── engine.ts         Deterministic validation · SHA-256 ledger
│   ├── store.ts          Operation state machine
│   ├── vision.ts         HF / NVIDIA NIM / demo adapter
│   ├── actions.ts        Idempotent action executor · webhook outbox
│   ├── protocols.ts      Cross-industry protocol templates
│   ├── fixtures.ts       Verified visual benchmark fixtures
│   └── mcp.ts            7 stdio MCP tools for Hermes
├── src/                  React 19 · Framer Motion · real-time SSE
├── shared/types.ts       Shared TypeScript contracts
└── docs/                 Architecture · Runbook · Use cases

Trust principles

  • Model perceives, policy decides. No LLM controls the release gate.
  • Hermes operates, PARALLAX enforces. The agent cannot bypass invariants.
  • Low confidence = no action. Below threshold → INSUFFICIENT_EVIDENCE.
  • Fixture replay is declared. Never passed off as live inference.
  • Every action has a proof. Idempotent receipts + independently verifiable ledger.

Verify

pnpm verify

Runs type-check · unit tests · MCP smoke test · live vision smoke test.


Contributing

Resource Description
CONTRIBUTING.md Setup, workflow, PR process
CODE_OF_CONDUCT.md Community standards
SECURITY.md Vulnerability reporting
LICENSE MIT — free to use, modify, distribute
HERMES.md Autonomous operator instructions

PARALLAX — Turn any camera into an accountable operations endpoint.

About

Autonomous reality reconciliation layer — AI agents verify the physical world via vision + deterministic policy before taking action. Hermes Agent × MCP × NVIDIA NIM.

Topics

Resources

License

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors