Disha
DISHA
"Direction" in Sanskrit — A self-evolving, multi-agent AGI platform for intelligence, cognitive reasoning, defense, and discovery.
What is DISHA?
DISHA is a production-grade AGI platform built from first principles — combining a 7-layer intelligent architecture, real-time multi-agent reasoning, 3-layer memory, quantum-inspired decision making, and a full-stack observability dashboard.
It is not a wrapper around a single LLM. It is a cognitive system designed to perceive, reason, deliberate, act, reflect, and learn — autonomously, in a loop.
Input → Perceive → Attend → Reason → Deliberate → Act → Reflect → Consolidate → Output
↑ |
└──────────── Memory & Learning ──────────┘
Architecture Overview
DISHA is organized into 7 layers, each independently deployable and CI-tested.
| Layer | Module | Purpose |
|---|---|---|
| 1 | src/ |
Core CLI Engine (TypeScript + Bun + React/Ink) |
| 2 | ai-platform/ |
Multi-Agent Intelligence Platform (FastAPI + Next.js) |
| 3 | cognitive-engine/ |
DISHA-MIND: 7-Stage Cognitive Loop |
| 4 | decision-engine/ |
4-Agent Reasoning Framework |
| 5 | cyber-defense/ |
Honeypot Network + ML Threat Detection |
| 6 | quantum-physics/ |
Quantum Circuit Simulator + Physics Engines |
| 7 | historical-strategy/ |
AI Military & Strategic Intelligence |
Layer 3 — DISHA-MIND Cognitive Engine
The cognitive engine is the intelligence core of DISHA. It processes every input through 7 stages, maintaining persistent memory across sessions.
7-Stage Cognitive Loop
Stage 1 PERCEIVE Intent classification, entity extraction, uncertainty estimation
Stage 2 ATTEND 3-layer memory retrieval, working memory decay management
Stage 3 REASON Parallel deductive / inductive / abductive hypothesis generation
Stage 4 DELIBERATE Multi-agent consensus (Planner + Executor + Critic) + dissent preservation
Stage 5 ACT Confidence-gated action selection or clarification request
Stage 6 REFLECT Quality scoring, metacognitive analysis, learning trigger detection
Stage 7 CONSOLIDATE Episodic storage, concept extraction, semantic graph update
3-Layer Memory Architecture
| Layer | Type | Capacity | Persistence |
|---|---|---|---|
| Working | Volatile attention buffer | 8 slots | In-process |
| Episodic | Time-stamped event log | Unlimited | JSON on disk |
| Semantic | Concept relationship graph | Unlimited | JSON on disk |
Multi-Agent Deliberation
Three independent agents reason in parallel:
- Planner — Strategic recommendation
- Executor — Tactical recommendation
- Critic — Challenge and dissent
Consensus is computed by confidence-weighted voting. Dissenting views are preserved, not discarded — they surface in the reflection stage and influence learning.
Iterative consensus: if inter-agent agreement < 0.4, bottom-50% agents re-deliberate (up to 3 rounds).
Layer 2 — AI Intelligence Platform
Backend (FastAPI)
| Component | Description |
|---|---|
| Orchestrator | 5-phase investigation pipeline |
| OSINT Agent | Passive DNS, IP intel, threat feeds |
| Crypto Agent | Blockchain address analysis |
| Detection Agent | Anomaly detection from entities |
| Graph Agent | Neo4j knowledge graph (UNWIND batch) |
| Reasoning Agent | LLM-based chain-of-thought analysis |
| Vision Agent | GPT-4o / LLaVA image analysis |
| Audio Agent | Whisper transcription + analysis |
| RL Engine | PPO (12-dim state, 8 actions, prioritized replay) |
| GNN | GCN encoder + link predictor + graph classifier |
| Vector Store | ChromaDB (async, non-blocking) |
| Ranking | PageRank + temporal decay |
| Prompt Optimizer | Evolutionary optimization |
Frontend (Next.js + Tailwind)
14 visualization panels including:
- Real-time alerts feed (4 severity levels)
- Knowledge graph canvas
- RL training metrics
- Quantum physics interface
- Geographic threat map
- Cognitive Loop visualizer — live 7-stage pipeline with working memory bars, hypothesis panel, agent deliberations, reflection quality gauge
Layer 1 — Core CLI Engine
Built with TypeScript, Bun, and React/Ink for terminal rendering.
- 40+ tools — file I/O, bash execution, web search, LSP, MCP, agent spawning
- 100+ commands — git, code review, configuration, model management
- Model Context Protocol (MCP) — server with STDIO/HTTP/SSE transports
- IDE bridge — VS Code and JetBrains integration
- OpenTelemetry — traces, metrics, logs out of the box
Security Hardening (v4.0.0)
| Issue | Fix |
|---|---|
| WebSocket unauthenticated | JWT validation via ?token= query param; closes with 4001/4003 on failure |
| Empty SECRET_KEY signed valid JWTs | field_validator auto-generates secure 32-byte key in dev; enforces explicit setting in production |
| Hardcoded CORS origins | CORS_ORIGINS env var with get_cors_origins() method |
| Blocking async event loop | ChromaDB calls wrapped in asyncio.to_thread() |
| N+1 Neo4j writes | Single UNWIND $rows batch query replaces per-entity MERGE loop |
/context no input validation |
Query(min_length=1, max_length=500, ge=1, le=20) guards |
CI/CD
10 GitHub Actions pipelines, all green:
| Pipeline | Trigger | Checks |
|---|---|---|
ci.yml |
push/PR to main | Biome lint + Bun tests |
cognitive-engine-ci.yml |
cognitive-engine/** |
flake8 + pytest (≥20% coverage) |
ai-platform-ci.yml |
ai-platform/** |
flake8 + pytest |
decision-engine-ci.yml |
decision-engine/** |
flake8 + pytest |
cyber-defense-ci.yml |
cyber-defense/** |
flake8 + pytest |
sentinel-ci.yml |
scripts/sentinel/** |
flake8 + pytest |
codeql.yml |
Scheduled | SAST (TypeScript + Python) |
continuous-training.yml |
Daily 2 AM UTC | RL + GNN + Decision Engine training |
disha-mythos.yml |
Scheduled | Learning agent execution |
modules-ci.yml |
push to main | Module-level testing |
Knowledge Base
8 domains, continuously expanding:
| Domain | Coverage |
|---|---|
| Physics | Quantum mechanics, relativity, unified field theory |
| Mathematics | Number theory, calculus, probability, discrete math |
| Computing | Algorithms, systems, CS fundamentals |
| Chemistry | All 118 elements of the periodic table |
| Law & Politics | Constitutional law, case law, legal frameworks |
| Cybersecurity | Threat taxonomy, OSINT, defense patterns |
| Innovation | Emerging tech, future systems |
| Historical Strategy | 32+ military conflicts, scenario simulation |
Quick Start
Prerequisites
- Bun ≥ 1.1.0
- Python ≥ 3.11
- Docker + Docker Compose
- Node.js ≥ 18 (for frontend)
CLI
git clone https://github.com/Tashima-Tarsh/Disha.git
cd Disha
bun install
bun run build
./dist/cli.mjs
AI Platform (Full Stack)
cd ai-platform/docker
docker compose up -d
Services start at:
- Frontend:
http://localhost:3000 - Backend API:
http://localhost:8000 - Cognitive Engine:
http://localhost:8001 - Neo4j:
http://localhost:7474 - ChromaDB:
http://localhost:8002
Cognitive Engine (Standalone)
cd cognitive-engine
pip install -r requirements.txt
python -m pytest tests/ -v
from cognitive_engine import CognitiveEngine
engine = CognitiveEngine()
state = await engine.process("Analyze domain evil.io for threats", session_id="s1")
print(state.action)
Project Stats
| Metric | Count |
|---|---|
| Files | 3,700+ |
| Lines of Code | 452,000+ |
| Python Modules | 80+ |
| TypeScript/TSX Files | 100+ |
| AI Agents | 7 |
| Cognitive Stages | 7 |
| Memory Layers | 3 |
| API Endpoints | 49+ |
| Tools | 40+ |
| Commands | 100+ |
| CI Pipelines | 10 |
| Dockerfiles | 19 |
| Knowledge Domains | 8 |
| Test Files | 13+ |
| Historical Scenarios | 32+ |
About the Creator
Tashima Tarsh — AGI researcher, full-stack engineer, and system architect. DISHA is an independent research project exploring cognitive architectures, multi-agent systems, and autonomous intelligence.
- GitHub: @Tashima-Tarsh
Links
| Resource | Link |
|---|---|
| Wiki | WIKI.md |
| Changelog | CHANGELOG.md |
| Contributing | CONTRIBUTING.md |
| Usage Guide | USAGE_GUIDE.md |
| Issues | GitHub Issues |
DISHA — दिशा — Direction. Built with intention.