Offline CI/CD Mock Generation (Zero token expenditure)

🎯 Architectural Purpose

Synthesizing formal PMO deliverables using Large Language Models (LLMs) requires much more than naive zero-shot chat prompts. Naive generation frequently produces: 1. Hallucinated structure: Inconsistent markdown headings, missing document control blocks, and fabricated compliance codes. 2. Schema non-conformance: Numerical dates formatted improperly, missing mandatory project fields, or inverted risk scoring. 3. Stylistic dissonance: Conversational, non-authoritative prose unfit for executive and steering committee review.

The Tasleemat LLM Synthesis Pipeline solves this through Three-Tier Grounding: - Layer 1: Structural Scaffolding: Grounded against pre-engineered system prompts and template parameters (forms/en/.../*.md and forms/ar/.../*.md). - Layer 2: Authentic Few-Shot Exemplars: Conditioned on 228 verified real-world project deliverables in examples/en/ and examples/ar/. - Layer 3: Programmatic Schema Enforcement: Validated against companion JSON Schemas via tasleemat.validate_artifact_data prior to persistent storage.

flowchart TD
    NOTES["Raw Input: Meeting Notes / Transcript"] --> PROMPT_ENG["In-Context Prompt Assembler"]
    TEMPLATE["Template Markdown & Parameter Block"] --> PROMPT_ENG
    EXEMPLAR["Few-Shot Exemplar (examples/en/*.md)"] --> PROMPT_ENG
    PROMPT_ENG --> LLM["LLM Engine (Gemini / OpenAI / Claude / Ollama)"]
    LLM --> RAW_MD["Synthesized Markdown Deliverable"]
    RAW_MD --> EXTRACT["Extract Structured JSON Payload"]
    EXTRACT --> VALIDATE{"tasleemat.validate_artifact_data()"}
    VALIDATE -->|Valid| FINAL["Approved Publication-Ready Deliverable"]
    VALIDATE -->|Errors| FIX["Self-Correction Loop"]
    FIX --> LLM

🚀 1. The Unified Python AI Client (`tasleemat.ai.AIClient`)

Tasleemat includes a built-in multi-provider LLM client designed specifically for enterprise document drafting. It automatically resolves credentials from ~/.tasleemat/config.json or standard environment variables (GEMINI_API_KEY, OPENAI_API_KEY, ANTHROPIC_API_KEY).

Supported Providers & Models

Provider Code	Default Model	Custom Endpoints Supported	Environment Variable
`gemini`	`gemini-2.5-flash`	No	`GEMINI_API_KEY`
`openai`	`gpt-4o`	Yes (`api_base`)	`OPENAI_API_KEY`
`anthropic`	`claude-3-5-sonnet`	No	`ANTHROPIC_API_KEY`
`ollama`	`llama3.3:latest`	Yes (`http://localhost:11434`)	None (Local execution)
`mock`	`offline-deterministic`	Local	None (Testing & CI pipelines)

Programmatic Python Synthesis Example

import os
import tasleemat
from tasleemat.ai import AIClient

# 1. Initialize client with automatic failover support
ai_client = AIClient(provider="gemini", model="gemini-2.5-flash")

# 2. Retrieve authoritative template instructions for Project Charter (PMO-03.01)
charter_def = tasleemat.get_artifact_definition("PMO-03.01")
charter_prompt_template = tasleemat.get_artifact_asset("PMO-03.01", role="prompt", lang="en")

# 3. Assemble Grounded In-Context Prompt
user_meeting_notes = """
Project: Enterprise Data Mesh Migration
Sponsor: Dr. Khalid Al-Mansour, VP of Digital Technology
Budget Envelope: $3,200,000 USD
Timeline: 9 months, kickoff on 2026-11-01.
Primary Objective: Transition 14 legacy departmental databases into a unified, decentralized data mesh architecture with real-time governance.
Key Risks: Resistance from legacy database administrators, data quality disparities.
"""

system_instruction = (
    "You are a Senior Enterprise PMO Director. Generate a formal, publication-ready "
    "Project Charter strictly adhering to the template format below. "
    "Do not include conversational conversational introductions or concluding remarks."
)

full_prompt = f"""
{charter_prompt_template}

### INPUT PROJECT NOTES:
{user_meeting_notes}
"""

# 4. Synthesize Deliverable
generated_charter = ai_client.generate(
    prompt=full_prompt,
    system_instruction=system_instruction
)

print(f"Synthesized Charter ({len(generated_charter)} characters generated)")

🛠️ 2. CLI Automated Synthesis (`tasleemat generate`)

For automated shell scripts and local developer workflows, the CLI provides the generate command:

# Basic generation using notes file
tasleemat generate \
  --form PMO-03.01 \
  --notes ./raw_transcripts/kickoff_notes.txt \
  --out ./my_project/03_01_Project_Charter.md

# Generate in Arabic using OpenAI GPT-4o
tasleemat generate \
  --form PMO-04.08.02 \
  --notes ./workshop_notes_ar.txt \
  --lang ar \
  --provider openai \
  --model gpt-4o \
  --out ./my_project/04_08_02_سجل_المخاطر.md

# Offline CI/CD Mock Generation (Zero token expenditure)
tasleemat generate \
  --form PMO-03.01 \
  --notes "Unit test payload" \
  --mock \
  --out /tmp/mock_charter.md

🔒 3. Defensive Prompt Engineering: Preventing Hallucination

When prompting models to write PMO deliverables, apply these four architectural guardrails:

Guardrail 1: The `[ Requires Human Input ]` Directive

Explicitly instruct the LLM in system instructions:

"Whenever specific project data (such as exact financial amounts, vendor names, or milestone dates) is absent from the provided notes, output [ Requires Human Input: <Description> ]. Never invent metrics or financial figures."

Guardrail 2: Enforcing Document Control Headers

Ensure the model formats the metadata header with standardized variable slots:

| Document Attribute | Specification |
| :--- | :--- |
| **Project Code** | PRJ-2026-MESH-01 |
| **Document ID** | PMO-03.01-CHARTER |
| **Author** | Tasleemat AI Assistant (Supervised by PM Lead) |
| **Version** | 1.0 (Draft) |
| **Security Classification** | Internal Confidential |

Guardrail 3: Cryptographic Audit Stamping

Tasleemat templates embed an audit placeholder block. Direct the LLM to preserve the placeholder:

<!-- AUDIT_BLOCK_START -->
**Cryptographic Integrity Stamp:** `{{Audit_Trail_SHA256_Hash}}`
<!-- AUDIT_BLOCK_END -->

Prior to gate submission, your pipeline replaces this placeholder with the SHA-256 hash of the sanitized text body, ensuring immutable change detection.