🎙️ SpeechTools: Enterprise Indic Speech SaaS & Autonomous Multi-Agent Loop

---

📌 Executive Summary

SpeechTools is a production-grade, enterprise multi-tenant SaaS platform and autonomous AI agent loop engineered for high-throughput Indic speech annotation, automated boundary correction, intelligent segmentation, transcription, quality control (QC), and self-healing remediation.

The system addresses the critical bottlenecks in large-scale speech dataset annotation across 7 Indian languages: 🇮🇳 Hindi (hi-IN), 🇮🇳 Marathi (mr-IN), 🇮🇳 Bengali (bn-IN), 🇮🇳 Telugu (te-IN), 🇮🇳 Kannada (kn-IN), 🇮🇳 Tamil (ta-IN), and 🇮🇳 English (en-IN).

It operates under a dual-architecture paradigm:

1. Cloud Multi-Tenant SaaS Platform: Asynchronous FastAPI backend, PostgreSQL 16 with Row-Level Security (RLS) forced on all tenant tables, Redis 7 Lua-scripted distributed semaphore, Arq task queue with mid-task crash recovery, Argon2id security pool (<= 256 MiB RAM), and Caddy TLS reverse proxy.
2. Autonomous Multi-Agent Loop (LECS): Supervisory Loop Engineering Control System, Playwright Chromium headless automation, SQLite adaptive rule memory (memory/rules.db) with 110 canonical rules, hardware-fingerprinted zero-knowledge credential vault (security/credential_vault.py), and automated End-of-Day (EOD) reporting.

---

📑 Table of Contents

• 1. Dual-Architecture Overview
• 2. Chronological Engineering Journey (From Start to Finish)
• 3. Core Architecture & Security Models
  • 3.1 Multi-Tenant PostgreSQL Row-Level Security (RLS)
  • 3.2 Concurrency Governance: Distributed Redis Semaphore
  • 3.3 Argon2id Bounded Worker Pool & Memory Budget
  • 3.4 Zero-Knowledge Envelope Encryption Vault
  • 3.5 Adaptive Learning Memory (rules.db)
• 4. Empirical Benchmarks & Measured Experiments (E1 to E9, A4.4)
• 5. Test Traceability & Taxonomy (100% Passing)
• 6. Production Execution & Live Remediation Proofs
  • 6.1 191 Tasks Portal Execution Ledger
  • 6.2 Remediation of 9 QC-Rejected Tasks
  • 6.3 End-of-Day (EOD) Reporting Workflow
• 7. Commercial Monetization & Licensing Engine
• 8. What Was Done vs What Remains (Roadmap / Next Gates)
• 9. Setup, Installation & Verification Guide
• 10. Repository Directory Map

---

1. Dual-Architecture Overview

SpeechTools solves two operational modes simultaneously:

                                  ┌──────────────────────────────────────────────┐
                                  │           SPEECHTOOLS CORE PLATFORM          │
                                  └──────────────────────┬───────────────────────┘
                                                         │
                        ┌────────────────────────────────┴───────────────────────────────┐
                        ▼                                                                ▼
   ┌────────────────────────────────────────┐                      ┌────────────────────────────────────────┐
   │    ENTERPRISE MULTI-TENANT CLOUD SAAS  │                      │   AUTONOMOUS WORKER AGENT & DESKTOP    │
   ├────────────────────────────────────────┤                      ├────────────────────────────────────────┤
   │ • Caddy TLS Reverse Proxy (Trusted XFF)│                      │ • 1-Click Desktop Runner (START.bat)   │
   │ • FastAPI (Async SQLAlchemy + asyncpg) │                      │ • Local Control Plane API (Port 3001)  │
   │ • PostgreSQL 16 RLS (Default Deny)     │                      │ • Playwright Browser Automation Loop   │
   │ • Redis 7 Distributed Semaphore (Cap=8)│                      │ • SQLite Persistent Rule Memory (110)  │
   │ • Arq Async Worker Queue (Auto Requeue)│                      │ • Hardware Zero-Knowledge Vault (AES)  │
   │ • Argon2id Worker Pool (<= 256 MiB RSS)│                      │ • Automated EOD WhatsApp/Slack Ledger  │
   │ • React + TypeScript + Vite Frontend   │                      │ • 27-Task Free Trial & License Guard   │
   └────────────────────────────────────────┘                      └────────────────────────────────────────┘

---

2. Chronological Engineering Journey (From Start to Finish)

Every phase of SpeechTools was developed, empirically measured, and committed through strict gate-based verification:

Phase 0: Multi-Tenant SaaS Skeleton & Environment Setup

• Initialized multi-container orchestration (docker-compose.yml, docker-compose.cloud.yml).
• Configured Caddy reverse proxy with automatic HTTPS and proxy header forwarding.
• Structured least-privilege Docker containers with non-root security contexts (UID 1000).

Part A & A2: Hardening, Database Roles & Readiness Probes

• Established strict PostgreSQL role separation:
  • speechtools_admin: Migration runner with BYPASSRLS.
  • speechtools_app: Runtime application role with NOBYPASSRLS and forced row-level filtering.
• Implemented Kubernetes-style /healthz and /readyz probes verifying deep connectivity to PostgreSQL and Redis.

Phase 1: Authentication, Token Family Rotation & Rate Limiting

• Built state-of-the-art authentication with Argon2id password hashing.
• Implemented secure HTTP-only, SameSite=Lax cookie emission with anti-CSRF double-submit token headers.
• Engineered refresh token rotation with Theft Detection & Family Revocation (replaying a rotated token terminates the entire session).
• Calibrated SlowAPI rate limiting (IP-based) and exponential backoff account lockouts (5 bad attempts -> 15 min lock).
• Verified with empirical benchmarks: Experiments E1, E2, E3, E4, E5.

Phase 2: Tenant Isolation, Row-Level Security & AES-256-GCM Envelope Vault

• Migrated multi-tenant schema with PostgreSQL 16 Row-Level Security (RLS) forced across all tenant tables: vault_credentials, tasks, sessions, and audit_logs.
• Implemented session-level context injection via set_config('app.current_tenant_id', :tenant_id, true).
• Designed two-tier envelope encryption: Random 256-bit DEK per secret, encrypted with host KEK via AES-256-GCM.
• Verified authenticated ciphertext tamper detection and large payload handling (Playwright session state up to 512KB).
• Verified with empirical benchmarks: Experiments E6 & E7.

Phase 3 / Gates 0–5: Concurrency Limits, Worker Isolation & Evidence Protocol

• Gate 0 & 1: Formulated the non-negotiable Testing Constitution (TRUTH > SAFETY > VERIFICATION > RECOVERY > THROUGHPUT) and universal verification harness (scripts/verify_all.py).
• Gate 2: Established test taxonomy markers (@pytest.mark.unit, @pytest.mark.integration, @pytest.mark.e2e) and mapped full requirements in docs/TEST_MATRIX.md.
• Gate 3: Enforced strict test database isolation on speechtools_test with non-bypassable RLS break-it proofs.
• Gate 4: Implemented Redis distributed semaphore (app/services/distributed_semaphore.py) enforcing global atomic limits (MAX_BROWSERS=8, max 4 concurrent tasks per tenant).
• Gate 5: Re-benchmarked Argon2id in a bounded ThreadPoolExecutor(max_workers=4) ensuring peak memory overhead strictly <= 256 MiB, constant-time compare_digest comparisons, and zero user enumeration. Verified via Experiment A4.4.

Portal Automation, QC Remediation & Canonical Rule Engine

• Codified 110 verified rules into memory/rules.db and CANONICAL_GUIDELINES.md.
• Deployed autonomous agent_orchestrator.py and portal_agent.py on the live annotation portal:
  • Processed 191 tasks: 180 submitted, 10 reported (script mismatch), 1 failed.
  • Successfully remediated 9 historically rejected QC tasks with 0 errors on AWS validation gate (REMEDIATED_TASKS_PROOF.md).
• Built 1-Click desktop distribution (START.bat, local_api.py, ui/) with license gating and direct UPI monetization.

---

3. Core Architecture & Security Models

3.1 Multi-Tenant PostgreSQL Row-Level Security (RLS)

All sensitive tables (vault_credentials, tasks, sessions, audit_logs) enforce PostgreSQL Row-Level Security (ALTER TABLE ... FORCE ROW LEVEL SECURITY):

-- RLS Policy Example on vault_credentials
CREATE POLICY tenant_isolation_policy ON vault_credentials
    FOR ALL
    TO speechtools_app
    USING (tenant_id = current_setting('app.current_tenant_id', true)::uuid)
    WITH CHECK (tenant_id = current_setting('app.current_tenant_id', true)::uuid);

• Fail-Closed Default Deny: If a query executes without setting app.current_tenant_id, PostgreSQL returns 0 rows, preventing cross-tenant leakage even if application code omits a WHERE clause.
• Role Boundary: speechtools_app does NOT possess SUPERUSER or BYPASSRLS. Only speechtools_admin can bypass RLS during migrations.

3.2 Concurrency Governance: Distributed Redis Semaphore

To prevent Playwright browser explosions across multiple horizontally scaled worker containers:

• Implemented Lua-scripted atomic slot lease and heartbeat in Redis:
  • Global browser cap: MAX_BROWSERS = 8.
  • Per-tenant queue limit: MAX_TASKS_PER_TENANT = 2 (excess tasks receive HTTP 429).
  • Slot leases expire after 300s if worker crashes, preventing orphaned semaphore deadlocks.

-- Atomic slot acquisition with expiration check
local current = redis.call('ZCARD', KEYS[1])
if current < tonumber(ARGV[1]) then
    redis.call('ZADD', KEYS[1], ARGV[2], ARGV[3])
    return 1
else
    return 0
end

3.3 Argon2id Bounded Worker Pool & Memory Budget

Argon2id parameters calibrated to balance cracking resistance against denial of service:

• time_cost: 4 iterations
• memory_cost: 65,536 KiB (64 MiB)
• parallelism: 4 lanes
• Memory Ceiling: Wrapped in a bounded ThreadPoolExecutor(max_workers=4) gated by asyncio.Semaphore(4). Even under 500 concurrent login attempts, maximum RSS overhead for password hashing is capped at: $$\text{Peak Memory} = 4 \times 64,\text{MiB} = 256,\text{MiB}$$

3.4 Zero-Knowledge Envelope Encryption Vault

Local desktop and cloud credentials use AES-256-GCM envelope encryption:

• Hardware-fingerprinted machine key (SHA-256 derived from motherboard UUID + CPU ID via PBKDF2 with 600,000 iterations).
• Generates a random 256-bit Data Encryption Key (DEK) per secret.
• Plaintext is encrypted with DEK (AES-256-GCM, 12-byte random IV, 16-byte authentication tag).
• DEK is encrypted with Key Encryption Key (KEK).
• Authenticated tag verification detects any bit manipulation or tampering immediately.

3.5 Adaptive Learning Memory (memory/rules.db)

The agent features persistent SQLite memory with an adaptive feedback loop:

• Base Rules: 146 canonical rules parsed directly from CANONICAL_GUIDELINES.md.
• Learned Rules: Automatically extracted from QC rejection comments.
• Pattern Promotion: If 3 or more users encounter the same mistake pattern, the rule is promoted to GLOBAL_CRITICAL status.
• Rules are NEVER forgotten across restarts.

---

4. Empirical Benchmarks & Measured Experiments (E1 to E9, A4.4)

All values below were measured from real code execution on the host system:

Experiment	Focus Area	Metric Measured	Target Threshold	Measured Result	Verdict
E1	Argon2id Calibration	Hashing Latency (ms)	200–350 ms	239.23 ms (StdDev: 24.49 ms)	PASS
E2	Rate Limiting & Lockout	100 IP flood & 20 IP brute	429 on flood; Lock at attempt 5	90/100 HTTP 429; Lockout at attempt 5	PASS
E3	High Concurrency Load	500 concurrent signups/logins	Zero 500 errors	499/500 HTTP 200/201 (0 server errors)	PASS
E4	Token Replay Attack	Replay rotated refresh token	Immediate session revocation	HTTP 401; entire token family revoked	PASS
E5	Postgres Chaos Injection	Kill DB container mid-request	503 during outage; fast healing	Clean 503; self-healed in < 2.94s	PASS
E6	AES-256-GCM Envelope	Crypto Roundtrip Latency (ms)	p95 < 5.0 ms	Mean: 0.123 ms | p95: 0.279 ms	PASS
E7	PostgreSQL RLS Overhead	Direct vs RLS Query (ms)	Overhead < 3.0 ms	Baseline 1.71ms vs RLS 2.70ms (+0.99ms)	PASS
E8	Playwright Process RSS	Memory per Browser Context	< 2.5 GB at max load	1 context: 529MB | 8 contexts: 1,177MB (+91.5MB/ctx)	PASS
E9	Queue Acquisition Latency	Monotonic wait at cap (ms)	Backlog begins immediately	Under cap: 0.01ms | Past cap (15 tasks): 46.90ms mean	PASS
A4.4	Multi-Worker Scaling	40-task burst across 3 workers	Peak concurrency <= 8	Peak: exactly 8 active browsers; 40/40 completed	PASS

Full experimental methodology and raw outputs: docs/EXPERIMENTS.md

---

5. Test Traceability & Taxonomy (100% Passing)

Pytest Backend Suite (77 Tests)

Organized strictly by taxonomy markers:

• @pytest.mark.unit: 14 tests (in-memory settings validator, trace sanitizers, payload caps).
• @pytest.mark.integration: 62 tests (real PostgreSQL 16 with RLS, real Redis 7, Alembic migrations, token rotation).
• @pytest.mark.e2e: 1 test (real Playwright Chromium browser launch, cross-tenant context & cookie isolation).
• Result: 77/77 PASSED (100%).

Commercial System Test Suite (test_full_system.py)

5 consecutive end-to-end verification runs testing:

1. Machine Fingerprint stability (SHA-256 64-char hex).
2. AES-256-GCM encrypt/decrypt roundtrip, tampering, and 10KB payloads.
3. Hardware Credential Vault CRUD and wrong-PIN rejection.
4. Adaptive Rule Memory parsing 146 base rules, learning 5 feedback rules, and persisting across re-opens.
5. Honesty Protocol detecting uncertainty and missing guidelines.
6. License Guard offline handling and trial limit tamper detection.
7. License Server DB user/license/referral verification.
8. Complete integration pipeline simulation.

• Result: 87/87 PASSED across all 5 verification runs (100%).

Multi-Language QC Rules Suite (test_qc_rules.py & test_mock_pipeline.py)

• Verified all 16 QC gates across Hindi, Marathi, Bengali, English, Tamil, Kannada, and Telugu.
• Negative test verified: Script mismatch (e.g. Devanagari in Bengali task) immediately caught and routed to 'Not Target Language' reporting.
• Result: 100% PASSED.

Local API Suite (test_local_api.py)

• Verified all 9 endpoints (/, /api/dashboard, /api/memory/rules, /api/vault/info, /api/task-types, /api/analytics, /api/license/status, /api/memory/feedback, /api/settings).
• Result: 100% PASSED.

---

6. Production Execution & Live Remediation Proofs

6.1 191 Tasks Portal Execution Ledger

The autonomous agent operated on the live speech annotation portal (https://sbs.aidicrowd.com) with the following verified record:

Total Processed : 191 Tasks
✅ Submitted    : 180 Tasks (100% AWS QA Gate Pass Rate)
⚠️ Reported/Skip: 10 Tasks (Non-target language / No auto-transcript)
❌ Failed       : 1 Task

Every submitted task satisfied all mandatory domain rules:

• Segment Duration: Strictly $&lt; 30.0,\text{s}$.
• Boundary Gap: Strictly $\ge 1,\text{ms}$ between consecutive segments (no shared timestamps).
• Zero Hindi Danda: Strictly eliminated all । and |, replacing with sentence-ending punctuation (., ?, !).
• Foreign Spoken Words: Enclosed in <UNKNOWN/> tags.
• Personal Names: Tagged with <PName> ... </PName> with zero artificial letter-splitting.

Detailed task-by-task execution log: TASK_EXECUTION_REPORT.md

6.2 Remediation of 9 QC-Rejected Tasks

9 tasks previously rejected by the human QC team were fed to the autonomous remediation engine. The agent extracted the feedback, updated memory/rules.db, adjusted the timestamps, corrected tags, validated via QA sub-agent, and submitted with photographic proof:

#	Task ID	File Name	QC Rejection Reason	Remediation Applied	Status
1	1787207895405	..._seg3.m4a	English words transliterated; Hindi danda used; boundary overlap	Applied <UNKNOWN/>, replaced । with ., enforced 2ms gaps	✅ SUBMITTED
2	1787207895407	..._seg4.m4a	English words transliterated into Devanagari; punctuation	Tagged English lexicon with <UNKNOWN/>, verified Devanagari script	✅ SUBMITTED
3	1787207895409	..._seg5.m4a	Guidelines violation on profit/trend; sentence endings	Normalized ITN, added proper end punctuation	✅ SUBMITTED
4	1787207895411	..._seg6.m4a	Infrastructure, Capex, Service Sector transliterated	Tagged spoken English terms with <UNKNOWN/>	✅ SUBMITTED
5	1787207895413	..._seg7.m4a	BlackRock transliteration; offensive term handling	Remediated entity spelling, applied acoustic review	✅ SUBMITTED
6	1787207895415	..._seg8.m4a	Ratio, acceleration transliterated; missing tags	Tagged spoken technical terms with <UNKNOWN/>	✅ SUBMITTED
7	1787207895417	..._whole.m4a	Recipe share, unique, my channel transliterated	Applied <UNKNOWN/> to all conversational loan words	✅ SUBMITTED
8	1787207895419	..._whole.m4a	Missing <PName> tags on person names; spelling	Enclosed all names in <PName> ... </PName>	✅ SUBMITTED
9	1787207895421	..._whole.m4a	Missing name tags on historical figures / channels	Identified named entities, tagged <PName>	✅ SUBMITTED

Complete remediation audit and photographic diffs: REMEDIATED_TASKS_PROOF.md

6.3 End-of-Day (EOD) Reporting Workflow

The system maintains an automated ledger of all reported tasks in reported_tasks_eod.json. At the end of each shift, running:

python agent_orchestrator.py --eod-report

outputs the formatted EOD submission block for WhatsApp and Slack groups:

============================================================
📋 EOD REPORT FOR TC TASK ISSUES
============================================================
1. Not Target Language:
   - 029313ad-2540-5f90-bb49-fb772a73c0c6_seg0.m4a
   - 02ac0ca4-cf4f-58fb-9eb5-748532e4eb24_seg1.m4a
2. No Auto Transcription:
   - 1540e5d4-5351-5000-ab96-6f27d1e4610f_seg4.m4a
3. Reco Not in Target Language:
   (Skipped per protocol)
============================================================

---

7. Commercial Monetization & Licensing Engine

SpeechTools includes a built-in commercial distribution model designed for local execution with zero cloud hosting bills:

• 1-Click Portable Distribution: Entire app bundled into a shareable distribution with START.bat (Windows) and START_MAC.command (macOS).
• Free Batch (27 Tasks): Every new machine gets 27 free tasks out-of-the-box.
• Paywall & UPI Payment: On Task #28, the app pauses and displays a direct UPI QR payment modal:
  • Payee: Arsh Malik (HDFC Bank) — 0% payment gateway commission.
  • Pro Plan: ₹2,000/mo (or ₹1,500/mo with referral code SPECIAL500).
• Offline / Local Verification: License validation works with HMAC-SHA256 signature verification tied to the hardware fingerprint. Tamper detection automatically enforces trial limits if local files are manipulated.

---

8. What Was Done vs What Remains (Roadmap / Next Gates)

✅ What Has Been Completed

• Gates 0–5: Foundation, Playwright worker container, test taxonomy, test DB isolation, distributed Redis semaphore, Argon2 bounded pool, and anti-enumeration.
• Multi-Tenant SaaS: PostgreSQL 16 RLS, Alembic migrations, Caddy reverse proxy, JWT token family rotation, SlowAPI rate limiting.
• Local Control Plane & UI: local_api.py, Vite UI dashboard, credentials vault, EOD reporting, learning center.
• 110-Rule Adaptive Memory: SQLite persistent engine (memory/rules.db) with zero-forgetting guarantees.
• Live Portal Execution: 191 tasks executed, 180 submitted, 100% QA pass rate.
• QC Remediation: 9 rejected tasks remediated and verified.
• Comprehensive Benchmarks: Experiments E1 through E9 and A4.4 fully documented with raw data.
• 100% Tests Passing: 77 Pytest tests + 87 System tests + 7 Multi-language tests.

⏳ What Remains (OpenGates Roadmap)

• Gate 6: Coverage enhancement to $\ge 85%$ overall and $\ge 95%$ on critical auth/vault modules, plus Mutation Testing (mutmut $\ge 80%$ mutants killed).
• Gate 7: Phase 3b.1 Mock Portal (mockportal service, side-effect log, fault injection API).
• Gate 8: Phase 3b.2 Formal Loop Engine (PLAN -> ACT -> OBSERVE -> VERIFY -> DECIDE).
• Gate 9: Phase 3b.4 Chaos & Integration tests (lying portal break-it proof, kill-point chaos).
• Gate 10: Phase 3b.5 Experiments E8b, E9b, and E10 (browser lifecycle ADR, queue latency, 300 faults).
• Gate 11: Universal verification run (python scripts/verify_all.py), single consolidated summary, commit & STOP.

---

9. Setup, Installation & Verification Guide

Mode 1: 1-Click Local Desktop Runner (For Annotators & Workers)

# Windows: Double-click START.bat or run:
START.bat

# macOS: Double-click START_MAC.command or run:
bash start_mac.sh

# Open browser to access local UI:
http://localhost:3001

Mode 2: Enterprise Cloud Docker Compose Stack

# 1. Generate secrets (or copy .env.example)
cp .env.example .env

# 2. Build and start containers
docker compose up -d

# 3. Check health of all services
docker compose ps

# 4. Run database migrations
docker exec -it speechtools_backend alembic upgrade head

Mode 3: Running Test Suites

# Run all 77 backend pytest tests:
pytest -v

# Run backend tests categorized by taxonomy:
pytest -m unit
pytest -m integration
pytest -m e2e

# Run commercial system 5x verification suite:
python test_full_system.py

# Run local API endpoint tests:
python test_local_api.py

# Run multi-language QC validation tests:
python test_qc_rules.py
python test_mock_pipeline.py

Mode 4: Headless Task Automation Loop

# Run autonomous portal pipeline:
python agent_orchestrator.py

# Generate End-of-Day report for WhatsApp/Slack:
python agent_orchestrator.py --eod-report

---

10. Repository Directory Map

speech-tools/
├── backend/                  # FastAPI Enterprise Multi-Tenant Backend
│   ├── alembic/              # Database migration versions
│   ├── app/                  # Application core, models, schemas, services, API routes
│   │   ├── core/             # Security, config, settings validator, database session
│   │   ├── models/           # SQLAlchemy 2.0 multi-tenant models with RLS
│   │   ├── services/         # Auth, Vault, Distributed Semaphore, Task Queue
│   │   └── api/              # v1 API routers (auth, vault, tasks, health)
│   └── tests/                # 77 Pytest test files (unit, integration, e2e)
├── frontend/                 # Enterprise Cloud SaaS Web App (React + TS + Vite)
├── ui/                       # Local Desktop Control Plane UI (Port 3001)
├── memory/                   # Adaptive Learning Memory Engine
│   └── rule_memory.py        # SQLite rules engine (110 canonical guidelines)
├── security/                 # Zero-Knowledge Credential Vault & License Guard
│   ├── credential_vault.py   # AES-256-GCM hardware-fingerprinted vault
│   └── license_guard.py      # Trial tracking, paywall gating, HMAC license checks
├── license-server/           # Central License & UPI Verification Server
│   └── main.py               # License issuance, payment webhook, referral system
├── docs/                     # Engineering Evidence & Test Registers
│   ├── EXPERIMENTS.md        # Raw logs for Experiments E1–E9 & A4.4
│   ├── TEST_MATRIX.md        # Traceability matrix across all requirements
│   ├── MICRO_DETAILS.md      # Edge-case register (MD-01 through MD-22)
│   └── evidence/             # Gate execution logs (gate0 through gate10)
├── caddy/                    # Caddy reverse proxy config with HTTPS & trusted XFF
├── scripts/                  # Verification, migration, and benchmarking scripts
│   └── verify_all.py         # Cross-platform single-command verification harness
├── experiments/              # Standalone experimental benchmarks (CSV/JSON output)
├── CANONICAL_GUIDELINES.md   # Official Speech Annotation Guidelines (Source of Truth)
├── TASK_EXECUTION_REPORT.md  # 191 tasks live execution ledger & QA report
├── REMEDIATED_TASKS_PROOF.md # Proof report for 9 remediated QC rejected tasks
├── USER_INSTRUCTION_GUIDE.md # Hindi/English step-by-step user manual
├── agent_orchestrator.py     # Main autonomous multi-agent loop orchestrator
├── portal_agent.py           # Playwright automation engine for annotation portal
├── local_api.py              # Local FastAPI control plane server
├── START.bat                 # 1-Click Windows startup launcher
├── START_MAC.command         # 1-Click macOS startup launcher
└── docker-compose.yml        # Multi-container orchestration definition

---

⚖️ License & Intellectual Property

Copyright © 2026 Arsh Malik. All rights reserved.
Commercial proprietary software. Unauthorized copying, distribution, modification, or commercial exploitation is strictly prohibited without explicit written license from the copyright owner.