Building production-grade multi-agent AI systems, bi-temporal cognitive memory engines,
cross-platform swarm orchestration, and Arabic-native LLM intelligence.
I'm a Systems Architect and Software Engineer passionate about agentic AI architectures, autonomous swarms, cognitive memory systems, and multi-platform software.
- 🧠 Creator of Kazma: Designed and built the Kazma Multi-Agent Framework from the ground up (~159K LOC, 4,360+ tests, solo-architected).
- 🧬 Pure V2 Cognitive Memory: Architected Kazma's single-source-of-truth memory engine — featuring bi-temporal belief stores (
valid_from/valid_until), Local Ego-Graph Personalized PageRank (PPR), Laplace-smoothed procedural action DAGs, and prompt-fenced per-turn context injection. - 🐝 Dynamic Swarm Autoscaler: Built dynamic worker templates and best-model-per-task auto-routing (coding, reasoning, vision) for zero-configuration swarm execution.
- 🛡️ Safety & Reliability: Engineered triple-wired Human-In-The-Loop (HITL) safety gates, prompt-fenced Soul evolution deltas, and self-healing circuit breakers.
- 🌐 Arabic-Native AI: Spearheaded dialect formalization, custom tokenization, and cultural protocol alignment (Kuwaiti & Gulf dialects) for enterprise LLMs.
Production-Grade Autonomous Agents with Deep Multilingual Intelligence
🔗 Repository: github.com/Mubder/kazma
| Metric | Details |
|---|---|
| Codebase Size | ~159,000 Lines of Python, JS/CSS, & Shell |
| Test Suite | 4,360+ Automated Tests Passing (Zero Regressions) |
| Architecture | LangGraph Supervisor + Swarm Autoscaler + Pure V2 Memory + Dual Web/TUI IDE |
| Multi-Platform | Web UI, TUI, CLI, Telegram, Discord, Slack |
-
Pure V2 Cognitive Memory Engine: Fully migrated to Memory V2 as the single source of truth. Features bi-temporal belief tracking, RRF episode retrieval, Local Ego-Graph PPR, and Laplace confidence-smoothed action DAGs
$C(d) = \frac{S+1}{N+2}$ . -
Dynamic Swarm Autoscaler: Auto-spawns specialist workers from templates (
coder,researcher,generalist) with automatic best-model-per-task prompt classification. -
Pluggable Scraping Proxy Engine: Rotating residential/mobile proxy provider (
anyip.io) with automatic 429/403 backoff retries and user-agent pool rotation. -
Prompt-Fenced Soul Engine: System-prompt refinements (Soul deltas) are persisted to
ConfigStoreand wrapped in<kazma:data untrusted>fences to prevent prompt injection. -
Time-Travel Replay & Branching: Snapshot recorder captures supervisor states to SQLite WAL (
snapshots.db), allowing in-place rewinds (/replay) and thread branching (/fork). - Vision Capability Router: Intelligently routes vision tasks based on model capabilities, protecting text-only models (like DeepSeek) and auto-selecting vision models.
-
Nightly Automated Memory Backups: Background schedulers run 24-hour native
sqlite3.backup()routines and JSONL / GraphML snapshot dumps intomemory_ops.db.
| Domain | Technologies |
|---|---|
| Languages & Runtimes | Python 3.11+ • TypeScript • JavaScript (ES6+) • SQL • PowerShell / Bash |
| AI & Agentic Frameworks | LangGraph • OpenAI API • Anthropic Claude • Google Gemini • DeepSeek • MCP |
| Memory & Search Engines | SQLite-vec • FTS5 • Bi-Temporal Graph • PPR Graph • Reciprocal Rank Fusion |
| Backend & Infrastructure | FastAPI • Uvicorn • AsyncIO • PostgreSQL • SQLite WAL • Docker • uv |
| Frontend & Interfaces | Vanilla CSS (Modern/Dark) • Alpine.js • HTML5 Canvas • Textual TUI |
- 🐙 GitHub Profile: @Mubder
- 📦 Kazma Repository: Mubder/kazma
- 💬 Ask me about: Agentic AI architecture, cognitive memory graphs, LangGraph, Python performance & multi-agent swarms.
Designed & engineered by Mubder Alfaris — Built in Kuwait 🇰🇼