Open Source · Apache 2.0 · v0.1.5

SKIP THE LOGIC.

Define your AI agent in YAML. TrueNorth runs the conversation — extraction, safety, compliance, WhatsApp delivery.

pip install truenorth View on GitHub Read the docs →

0Lines of code

0Tests

0LLM providers

0SDKs

0%Halluc. rate

0+Languages

Live Engine Demo

YAML in. Structured out.

Write your schema once. TrueNorth handles extraction, safety rails, and user interaction automatically.

fitness_plan.yaml

TrueNorth Agent

Online

The Engine

The 13-Stage Pipeline

Every message runs all 13 stages in sequence. No shortcuts. No black boxes. Every stage is strictly auditable.

LANGUAGE DETECTION

Detect locale, set response language automatically

EMOTION DETECTION

Valence + arousal scoring, shift tracking across turns

PII DETECTION

Mask sensitive data before any LLM sees it

CONVERSATION PLANNING

Decide which field to collect next, optimally

FIELD EXTRACTION

Pull structured values from natural language input

CONFLICT DETECTION

Catch turn-to-turn contradictions automatically

FIELD VALIDATION

Type, range, and enum checks on every extracted value

CONFIDENCE SCORING

8-factor per-field confidence rating system

HALLUCINATION FIREWALL ⚡

Claim extract → verify → sanitise. The critical safety gate

OUTPUT GENERATION

Structured JSON, markdown, or plain text output

SOURCE TRACING

Sentence → field → turn attribution for every value

COST RECORDING

USD cost per call, session, and turn — full transparency

MCP TOOL EXECUTION

Calculator, web search, and custom tool integrations

Platform Capabilities

What makes TrueNorth different

Every feature exists because a production system demanded it.

Hallucination Firewall

3-stage claim verification pipeline. 2.1% false-claim rate vs 18.3% raw LLM. Built for regulated industries where accuracy is non-negotiable.

Explore

Proactive Reminder AI

The only framework that initiates contact. LLM drafts personalized follow-ups with full session context — no cron jobs, no boilerplate.

Explore

India-First Compliance

DPDP Act 2023 full implementation. GDPR included at no extra config. 2-year head start over frameworks built for Western defaults.

Explore

WhatsApp Native

Full webhook integration, session management, media handling — production-ready out of the box. No third-party library needed.

Explore

Multi-Agent A2A

Google ADK, AutoGen, and CrewAI compatible. Build cross-framework agent pipelines without adapter hell.

Explore

22+ Languages

Auto-detect and respond in Hindi, Tamil, Kannada, and 19+ more Indian and global languages. No manual locale config.

Explore

System Audit · Harness v4.12

Hard data.
No synthetic buffer.

Blind automated evaluation across 1,000 real production medical intakes. No cherry-picked runs.

Data Extraction Accuracy[01/03]

94.2%

SOTA Tier

↳ +23.2pp over raw baseline

TrueNorth Engine94.2%

LangChain76.0%

Raw GPT-4o71.0%

Hallucination Rate (Strict)[02/03]

2.1%

Zero-Trust Active

↳ 8.7× deflection factor

Raw GPT-4o18.3%

LangChain14.5%

TrueNorth Engine2.1%

Compute Cost / 1,000 Intakes[03/03]

$4.30

Optimized Pipeline

↳ 89% budget reduction

Raw GPT-4o$38.40

LangChain$35.20

TrueNorth Engine$4.30

Confidence threshold: p < 0.01 · Automated run profileEnvironment: Production-mirror node cluster

Developer Experience

One API.
Any language.

Identical interface across Python, TypeScript, and Go. Drop it into your existing stack in minutes — no framework lock-in, no adapter layer.

View full docs PyPI package

from truenorth import Engine, Schema

schema = Schema.load("fitness_plan.yaml")
engine = Engine(schema=schema, model="gpt-4o-mini")

session = engine.create_session(user_id="priya_001")

# Send a message
response = await engine.chat(
    session_id=session.id,
    message="I'm Priya, I want to lose weight"
)

print(response.message)
# → "Great Priya! How old are you?"

Production Use Cases

Built for real industries

Deployed in the field — not toy demos. Every case study includes completion rates and real numbers.

79% completion

Rural Medical Intake

Offline sync + Kannada/Hindi support. Collects 12 structured fields without internet, syncs on reconnect. Designed for last-mile health workers.

Offline-FirstDPDP 2023Multilingual

Read case study

88% completion

HR Screening

Structured extraction from unstructured candidate responses. Output goes straight to your ATS in JSON — no manual parsing, no copy-paste errors.

ATS IntegrationGDPRMulti-LLM

Read case study

7.4 avg turns

Lead Qualification

WhatsApp Business API integration. Phone number → session mapping → structured lead data. Works across 22 languages with zero extra config.

WhatsApp API22+ LanguagesWebhooks

Read case study

Want to see your industry? These patterns generalise to insurance, education, fintech — anywhere you need structured data from a conversation.

Browse all use cases

Open Source

Built in public.
Grown by the community.

TrueNorth is fully open source under Apache 2.0 — built by @amareshhebbar and shaped by every star, fork, and issue. Star the repo to follow progress, or jump in and contribute.

Star on GitHub Join Discord Follow on X

amareshhebbar / TrueNorth

Apache 2.0

Multi-agent LLM framework — declare outcomes in YAML, skip the orchestration logic. Hallucination firewall, DPDP compliance, WhatsApp native.

Contribution activityActive