Welcome to my digital systems hub. I bring 15+ years of hands-on operations leadership in telecom and supply chain management, paired with modern Agentic AI Engineering (LangGraph, RAG, WebSockets, deterministic safety guardrails).
Below are the direct engineering deep-dives, architecture logs, and production-grade case studies available in the DigiSuite Portfolio Hub:
- 📗 Engineering Log 01: Guardrails & Hallucination Prevention
Building a deterministic safety layer around an LLM agent with live SQL tool-calling and strict citation verification. - 📙 Engineering Log 02: Multilingual Semantic Search & Shamela Integration
Scaling Ahmad-AI to 113,842 indexed Islamic records using two-stage retrieval (MiniLM-L12+GATE-Reranker-V1). - 🎙️ Engineering Log 03: Real-Time Voice AI Agent System
Full-duplex continuous voice assistant built on WebSockets, Deepgram Nova-3, GPT-4o-mini, and Aura-2 TTS. - 📈 Engineering Log 04: Production Demand Forecasting Engine
Multi-model wholesale ERP demand forecasting (Naive, Seasonal Naive, ETS) evaluated across 84,741 real order lines. - 📖 Engineering Log 05: Building a Quran Retrieval Benchmark From Scratch
68 hand-labeled questions, three rejected datasets, and the counter-intuitive finding that translating the corpus beats translating the query for cross-lingual retrieval. - 🧪 Engineering Log 06: Scaling Ahmad-Eval to 196 Hand-Reviewed Questions
Growing the benchmark 3x, hand-verifying every gold verse against the Quranic text, fixing a qid collision bug, and re-evaluating the retriever on the cleaned gold. - 🔍 Engineering Log 07: The Reranker Was Hurting, and the Metric Was Wrong
A corpus-wide indexing bug found by a diagnostic built for another purpose, a 12-point reranker regression hidden behind an auto-selector, and a 40-minute failure read showing the 19% metric was measuring gold-set specificity, not retriever quality. - 🚪 Engineering Log 08: The LLM Selector Failed; the Gate Ships
Four local models tested as Quranic abstainers. Zero correct abstentions across 24 draws. A two-layer scope-routing gate replaced the failed hypothesis and shipped at the pre-registered threshold: 8/10 out-of-scope catches, 0/4 answerable false positives. - 📕 Engineering Log 09: The Pipeline Does Not Generalize — and That's the Finding
196 calibration questions pass. 36 fresh questions fail the pre-registered thresholds on both conditions. The 232-question benchmark is published on Hugging Face with both the passing and the failing numbers, and the four-tier promise from Chapter 03 is checked against what was actually built. - 📊 Case Study: Wholesale ERP Demand Forecasting Validation
12-month backtested evaluation, WAPE analysis, lumpy order detection, and automated Excel worklist generation.
I build and optimize operational ecosystems that work — combining 15+ years of senior operations leadership with production-grade software development.
- Ericsson BSA (Saudi Arabia): Founding operational lead for one of the largest telecom managed services contracts in KSA (STC). Built vendor ecosystems from scratch, engineered SLA/WLA governance frameworks saving SAR 100,000+/month, and led zero-disruption warehouse migrations.
- Blue Light LLC (Pakistan): Restructured manual supply chain operations, cut FedEx logistics delivery cycle time by 50%, reduced shipment costs by 20%, and managed a $10M+ logistics portfolio.
| System | Description | Primary Tech Stack | Links |
|---|---|---|---|
| GrowEasy ERP | Full-stack SaaS ERP for SMEs (Inventory, Invoicing, Finance) with live SCM agent data layer. | Python, Flask, Postgres, Redis, Docker | GrowEasy Site |
| Demand Forecasting Engine | Multi-model B2B demand forecasting engine analyzing 84,741 order lines & 4,307 SKUs. | Python, Pandas, Statsmodels, Streamlit | Read Log 04 | Case Study |
| Ahmad-AI (RAG Agent) | Offline multi-turn research agent with hybrid cross-encoder reranking over 113k records; 196-question hand-reviewed Quran retrieval benchmark; per-bucket routing; two-layer scope-routing gate (LLM intent classifier + keyword backstop) that catches 8/10 out-of-scope questions without showing a wrong verse. | LangGraph, ChromaDB, GATE-Reranker, Ollama, camel-tools, BGE-reranker-v2-m3, command-r7b-arabic | Log 01 | Log 02 | Log 05 | Log 06 | Log 07 | Log 08 |https://emadaed.github.io/digisuite/engineering-log-9.html |
| Serverless WMS | Enterprise fulfillment & order processing platform running inside Google Workspace. | Google Apps Script, Google Sheets | Source Code | Live Demo |
| Learn With Misham | Scalable CRM & Learning Management System actively used by education providers. | JavaScript, Google Apps Script, LocalStorage | Live Platform |
- AI Systems & Agents: LangGraph, RAG (ChromaDB), Cross-Encoder Reranking (
GATE-Reranker-V1), WebSockets, Speech-to-Text (Nova-3), Text-to-Speech (Aura-2), Deterministic Guardrails. - Backend & Systems: Python, Flask, FastAPI, PostgreSQL, Redis, Celery, Docker, Google Apps Script.
- Supply Chain & Management: Digital Transformation, Demand Forecasting (ETS/Naive), Warehouse Management (WMS), SLA/WLA Governance, PMP®, RMP®.
Open to global roles and consulting in Agentic AI Systems Engineering, Digital Supply Chain Architecture, and Operations Automation.
- 📧 Email: muhammad.a.ahmed@gmail.com
- 💼 LinkedIn: linkedin.com/in/muhammad-ahmad-scm
- 💬 WhatsApp: +92 309 5222202
- 🌐 Portfolio Hub: https://emadaed.github.io/digisuite/
License: MIT — Feel free to explore, learn from, or build upon the open-source architecture components.