AI4OHS-HYBRID - Agentic HSSE/OHS Intelligence System

Overview

AI4OHS-HYBRID is an offline-first OHS intelligence engine that combines:

• Agentic AI (PLAN -> ACT -> REFLECT -> CORRECT -> LEARN)
• RAG (Retrieval-Augmented Generation) with FAISS
• CAG (Compliance-Augmented Guardrails)
• Guarded inference (system prompt + SelfEvaluator/RewriteFlow)
• Llama.cpp local inference
• ML pipelines (risk prediction, incident classification)
• Annual & quarterly automated OHS reporting
• Zeus Voice Interface (offline)

Optimized for:

• Turkish OHS law (6331 + regulations)
• ISO 45001
• OSHA 29 CFR
• WB/IFC ESS1-ESS10
• High-compliance infra projects (e.g., TERRP)

---

Directory Structure (Core Components)

src/
  agentic/
    planner_optimized.py
    self_evaluator.py
    self_eval_rewrite_flow.py
    guarded_inference.py
    task_graphs/
      annual_report_task_graph.py
  agents/
    annual_report_agent.py
  ai_ml/
    incident/incident_classifier.py
    risk_scoring/
      hazard_vectorizer.py
      risk_forecast_model.py
      risk_pipeline.py
    reporting/
      ohs_kpi_dashboard.py
      annual_ohs_report_generator.py
      integrated_kpi_risk_dashboard.py
      advanced_dashboard.py
  governance/
    audit_logger.py
    compliance_heatmap.py
    cag_rules_engine.py
  genai/rag/
    document_loader.py
    ocr_ingest.py
    chunker.py
  zeus_layer/
    listener.py
    startup_tasks.py
scripts/
  run_guarded_llama.py
  run_autonomy_cycle.py

---

Key Components

1. Agentic Task Graph for Annual Reports

src/agentic/task_graphs/annual_report_task_graph.py

• DAG steps: load incidents -> load ESS/6331 -> compute KPI -> heatmap -> Word+Excel -> validate.

from agentic.task_graphs.annual_report_task_graph import AnnualReportTaskGraph
result = AnnualReportTaskGraph().execute()

2. Integrated KPI + Risk Dashboard

• Combines TRIR/LTIFR/Severity + ML risk forecast (integrated_kpi_risk_dashboard.py).

3. Risk Pipeline

UnifiedRiskPipeline merges incident classification + hazard vectorization + ML risk forecast (XGBoost/MLP).

4. Annual Report Generator

Creates Word/Excel and KPI PNG (ai_ml/reporting/annual_ohs_report_generator.py).

5. Zeus Offline Voice Interface

Keyword triggers pipelines (zeus_layer/listener.py).

6. Self-Evaluation Rewrite Loop

RewriteFlow + SelfEvaluator auto-rewrites low-scoring outputs.

7. Guarded Inference (OHS Expert Persona)

• System prompt + guardrails + RAG context: scripts/run_guarded_llama.py
• Self-eval/auto-rewrite: --self-eval --threshold 0.8
• Prompt builder: src/genai/prompting/ohs_prompt_builder.py
• Llama client with ctx/timeout/retry: src/agentic/llama_learning_integration/llama_client.py

8. OCR -> RAG Ingest

• OCR-capable loader for images/PDF/docx: src/genai/rag/document_loader.py
• Direct OCR helper: src/genai/rag/ocr_ingest.py
• Plug screenshots into RAG ingest to “learn” from the screen.

9. Autonomous Cycle (Proposals)

• scripts/run_autonomy_cycle.py: SelfPlanning -> ACE dry-run -> SelfHealing -> SelfEvolving -> Approval queue (AUTONOMY_CYCLE).
• ACE/FERS writes to main repo only after ApprovalManager approval (ACE_ALLOW_AUTO_APPLY env to bypass; not recommended).

---

Data Requirements

data/analytics/ must contain:

• incidents_annual.json
• ess_6331_items.json

---

Output Standards

All outputs must include:

• Safety-first hierarchy
• Compliance mapping (6331 / ISO 45001 / WB ESS)
• Residual risk assessment
• CAPA / Preventive actions
• Strictest rule wins; state uncertainty if data is missing