Sthala
Your AI's place.
Most AI deployments let LLMs touch everything — computation, decisions, execution. That's the failure mode. Sthala is built on one hard constraint: LLMs narrate, code executes. Always. Enforced at build time.
An open-source reference framework for governed AI runtime placement on commodity x86 hardware. Airgapped by default. Built for SMBs, CA firms, clinics, schools — any organisation where data movement must stay under explicit boundary control.
Sthala answers a placement question that applies at any scale. This repository is the reference profile for one end of it; larger estates apply the same questions to cloud, hybrid and edge placement. It is a reference implementation: it applies ContextOps and ContextBoundary.
"Sthala" (स्थल / ஸ்தலம்) — place, site, ground. Where your AI actually runs.
The Problem
Modern AI systems let interpretation, decision-making, and execution share one trust boundary. That lets probabilistic systems directly own deterministic consequences — the root cause of prompt injection, unsafe automation, context contamination, and authority confusion.
This project is part of the context-stack, which separates interpretation from authority: intelligence proposes, governance validates, execution authorizes. Intelligence can suggest anything. Authority stays deterministic.
Full doctrine: https://github.com/kannanokannan/context-stack
Why Sthala
- Cloud AI is expensive — 3-year TCO cloud vs Sthala appliance: ~6× savings
- Compliance is tightening — EU AI Act (Aug 2026), India DPDP Act (Stage 2: Nov 2026), MCA backup mandates
- LLMs are commodities — the moat is the pipeline, not the model
- Refurb hardware works — a ₹1.5L node serves 50–200 concurrent SMB users on 7B–34B models
Sthala is not a product. It is a documented pattern. Read it, fork it, deploy it, adapt it.
What You Get
Boot a commodity PC → drop your data → get AI-powered analysis
Nothing crosses the approved boundary unless you explicitly approve it
- Immutable Linux base (no drift, atomic updates)
- Sarvam-30B default (Apache 2.0, Indic-optimised, 24GB VRAM Q6_K)
- Multi-model verifier pipeline (LLM extracts → code computes → LLM narrates)
- Egress consent gateway (you see what leaves, you approve every send)
- Vertical recipes: CA firm, distributor, school, clinic
Quick Start
curl -fsSL https://raw.githubusercontent.com/kannanokannan/sthala/main/install.sh | bash
Or with a profile:
curl -fsSL .../install.sh | bash -s -- --profile ca-firm
Repository Structure
sthala/
├── SPEC.md ← Framework specification
├── AGENTS.md ← Machine-readable build contract for AI coding agents
├── CLAUDE.md ← Context file for Claude / AI assistants
├── CONTRIBUTING.md ← Contribution guide
├── install.sh ← Idempotent bootstrap script
├── profiles/ ← Vertical presets
│ ├── generic.yaml
│ ├── accounting-firm.yaml
│ ├── clinic.yaml
│ ├── distributor.yaml
│ └── school.yaml
├── hardware/
│ └── detect.sh ← Auto-detect CPU/GPU/RAM, recommend tier
├── stack/
│ ├── docker-compose.yml ← Full service stack
│ ├── inference/ ← Inference services
│ ├── ingestion/ ← Ingestion services
│ ├── observability/ ← Metrics and tracing
│ └── rag/ ← Retrieval services
├── recipes/
│ ├── _template/ ← Six-stage recipe template
│ ├── in/ ← India recipes
│ ├── eu/ ← EU recipes
│ ├── us/ ← US recipes
│ └── global/ ← Global recipes
├── research/
│ ├── 2026-05-market.md ← India SMB AI market research
│ └── 2026-05-standards.md ← Standards & academic landscape
└── docs/
├── architecture.md ← Architecture deep-dive
├── compliance.md ← Compliance notes
├── egress.md ← Egress guidance
├── hardware.md ← Hardware guidance
└── models.md ← Curated model catalog with HuggingFace links
Design Principles
- LLM-as-component, not LLM-as-magic — LLM extracts, code computes, LLM narrates
- Deterministic compute, fuzzy understanding — numbers never come from an LLM
- Egress-by-consent — nothing leaves without explicit human approval
- Commodity hardware first — if it needs a $30K GPU, it's out of scope
- Vertical recipes over generic stacks — a CA firm is not a chatbot
Standards Alignment
- NIST AI RMF 1.0 + GenAI Profile
- ISO/IEC 42001 (AI management systems)
- EU AI Act (non-high-risk tier by default)
- India DPDP Act 2023 compliance posture
- 16-Factor App for AI (Google, 2024)
- OpenTelemetry GenAI semantic conventions
- CycloneDX 1.7 AI BOM auto-generated
License
Apache 2.0 — use it, fork it, ship it, build on it.
Author
Kannan Okannan — Chennai, India
Part of the Context Stack.
Part of the Stack
ContextOps and ContextBoundary are the specification layer. contextboundary-gw, Sthala and Griha are reference implementations that apply it. The composition is defined once, in COMPOSITION.md.
Sthala is a reference implementation. It applies ContextOps and ContextBoundary to governed runtime placement.
Canonical terminology and cross-project decisions: context-stack
