Farhan Ali
AI Engineer · Riyadh, Saudi Arabia
8+ years in data and AI engineering. 10+ systems designed, built and delivered. Local + Cloud LLMs open-weight models on-prem, Claude, GPT and Gemini in the cloud.
About
I'm Farhan Ali, an AI Engineer in Riyadh with 8+ years in data and AI. I take AI systems from the first sketch to production and into the daily work of operations teams.
My focus is agentic AI and LLM systems: RAG assistants that cite their sources, text-to-SQL agents that check and repair their own queries, and decision engines where code owns every number and the model only explains it. I deliver on Claude, GPT and Gemini when the data can travel, and on open-weight models on local GPUs when it cannot.
Analytics runs through everything I build: Python, SQL and Power BI, from the data model to the forecasting pipelines and the dashboards people open every morning. I have delivered across telecom, construction and healthcare supply, working directly with the people who use the systems.
Experience
The route so far
Ascend Solutions
AI Engineer, Apr 2026 to Present, Riyadh
Building a family of AI systems for national medical supply, on local open-weight models because the data has to stay inside: a document assistant, a SQL agent, a stock-transfer engine, predictive models and the portal that ties them together.
- Private RAG assistant on local LLMs
- Text to SQL agent with a semantic layer
- STO Recommendation Engine
- AI Portal for every system
- Demand and availability forecasting
Freyssinet Saudi Arabia
AI Engineer, Sept 2024 to Mar 2026, Riyadh
Led design and delivery of production AI across departments on cloud models: retrieval platforms on Azure OpenAI, agentic chatbots on Gemini, predictive pipelines, automated reporting and the dashboards leadership decided from.
- Azure RAG chatbots, 90%+ retrieval relevance
- Employee evaluation system, review cycle down 50%
- Duplicate billing detection
- ARIMA and SARIMAX demand forecasting
- Power BI for finance and operations
Special Communication Organization
Data Analyst / AI Engineer, Sept 2018 to Apr 2024, Pakistan
Over five years turning telecom network data into decisions: analytics, dashboards and predictive maintenance for a national operator.
- Predictive maintenance, unplanned outages down 30%
- Coverage initiative, regional revenue up 15%
- Power BI network and energy dashboards
- SQL and Python data quality pipelines
Selected work
Systems in
production
No real data or logos are shown. Every use case below runs on its real interface, with synthetic data only.
SQL Agent
Agentic text to SQL, fully offline
Ask the database a question in plain English and get a verified answer in seconds: no SQL, no analyst queue, and no data leaving the building.
- A LangGraph agent routes the question, writes the SQL, guards it as read-only, runs it and repairs its own errors.
- A semantic layer and a 789-KPI dictionary lifted answer accuracy from 33% to between 85% and 95%.
- Four modes in one workspace: free-text SQL, dashboards, KPI analysis with backtested forecasts, and an Action Center.
LangGraph · Qwen2.5-Coder-14B · llama.cpp · FastAPI · Next.js · SQL Server · ARIMA / ETS
AI Assistant
Private RAG over company documents
A private assistant for every employee. It answers from the company's own documents, shows the source of every line, and says so when it does not know.
- Six local models work as one: generation, embeddings, reranking, OCR and speech, with nothing leaving the network.
- Hybrid retrieval with reranking finds an exact file name or KPI code as reliably as a concept.
- Citations are verified after generation, so a source appears only when it truly supports the line.
LangChain · LangGraph · FAISS · BM25 · Reranker · OCR · Whisper · FastAPI · Next.js
STO Recommendation Engine
Decision engine + local LLM assistant
A decision engine that finds stock sitting idle in one warehouse while another runs dry, and proposes the exact transfer that fixes it.
- Every proposal follows six fixed business rules and can be traced back to the rule behind it.
- Planners see the impact before they act: availability before and after, on a live map of real road routes.
- A local LLM explains the plan in plain language, and every number it writes is verified against the engine.
Python · Rule engine · llama.cpp · Qwen3.8 27B · SQL Server · SVG map · Vanilla JS
Clinical Catalogue Search
Matching agent with LLM verification
Describe a clinical product in any wording and get the right catalogue code in seconds, with a confidence score and the reason behind it.
- Embedding search narrows about 55,000 items to the three closest matches.
- A local LLM verifies salt, strength and form, and explains what matches and what differs.
- Works on one item at a time or on a whole spreadsheet in batch.
Embeddings · Vector search · LLM verification · Batch CSV / XLSX · On-prem
Predictive ML Dashboards
Machine learning models, delivered in Power BI
Three machine learning models that warn before a problem lands, each delivered as a live Power BI dashboard built from custom HTML visuals.
- Proactive Fulfillment Model: scores every inbound shipment for delay risk and shows which vendors to chase first.
- Future Availability AI Insights: predicts zero-stock items one to three months ahead, the signal the STO engine plans against.
- Future Availability Outbound Demand: forecasts demand for the next 30, 60 and 90 days by region, category and material.
Python · scikit-learn · Random Forest · statsmodels · Power BI · DAX · HTML / CSS / SVG visuals
AI Portal
One front door, with a pulse
One address for every AI system and dashboard on the network, each with live status, response time and uptime.
- Built on the Python standard library alone: nothing to install, license or patch.
- Health checks with retries tell a slow service from a dead one, and adding a system is one line of config.
Python stdlib · Health probes · Config-driven · Vanilla JS
Crypto JARVIS
Market intelligence assistant
A personal market-intelligence desk built on one rule: code computes every signal, score and risk, and the model only explains a verdict it cannot change.
- A probabilistic BTC outlook built from options-implied and realised volatility.
- Twenty-plus live data sources, refreshed every fifteen minutes.
- Buy and sell questions answered from the whole portfolio, backed by 166 automated tests.
Next.js · TypeScript · Redis · Scheduled ingestion · Gemini · Vercel
Index
More work
Operational Reporting and Power BI
Ascend Solutions, 2026
Automated stakeholder reporting: data ingestion and consolidation pipelines, availability reporting, and dashboard visuals that replaced manual report compilation.
Azure RAG Chatbots
Freyssinet Saudi Arabia, 2024 to 2025
Several retrieval chatbots on Microsoft Azure using Azure OpenAI Service, Azure AI Search and vector stores (FAISS, Pinecone), reaching over 90% retrieval relevance.
SmartDoc Analyzer
Freyssinet Saudi Arabia, 2024 to 2025
A document assistant for construction records, built up over several versions. It started as an agentic LangChain, LangGraph and Gemini chatbot for large-context reasoning over PDFs, Excel and IFC files. The current pipeline runs scanned PDFs through OCR, expands each question into five variations, fuses BM25 and semantic search with RRF, and puts inline citations on every answer, at 90%+ retrieval accuracy. One build paired the vector store with a Neo4j knowledge graph.
Employee Evaluation Platform
Freyssinet Saudi Arabia, 2024 to 2025
A full-stack platform on Google Cloud with OAuth 2.0 that automates six-dimensional performance calculations and approvals. Review cycle time fell by 50%.
Duplicate Billing Detection
Freyssinet Saudi Arabia, 2024 to 2025
A multi-level validation pipeline that finds supplier billing anomalies across thousands of records, helping management recover duplicate charges.
BIM to Robot Orchestration
Freyssinet Saudi Arabia, 2024 to 2025
An end-to-end Azure pipeline from IFC and Revit models through Azure Digital Twins and Azure OpenAI task planning to ROS2 and Gazebo simulation.
TradePilot
Personal, 2026
A multi-agent spot trading bot: five analysis agents vote on entries, with fee-aware exits, paper and testnet modes, and a drawdown kill switch.
Investment Tracker and Wealth Bank
Personal, 2026
A web investment tracker for crypto, stocks and gold with allocations, FIFO profit and loss and multi-currency performance, and Wealth Bank, its Android app in React Native with PIN unlock and price alerts.
Trading Analyzer Pro
Personal, 2025
Upload a chart screenshot and get a full technical read. Gemini vision reads the chart while Python calculates RSI, EMAs, MACD, Bollinger Bands and ATR from market data, then a weighted confluence score and a trade setup with entry, stop and three targets are drawn back onto the chart. A later version splits the work across four specialist agents on Claude, Gemini and plain Python.
Backpack Cycle Trading Bot
Personal, 2026
A multi-user bot for the Backpack exchange. It places maker buy orders below the price and sells each fill at a small spread above fees, cycling across SOL, ETH and BNB. Every user gets their own account, encrypted API keys and bot instance, with rate limiting and stale-order clean-up.
Predictive Maintenance for a Telecom Network
Special Communication Organization, 2018 to 2024
Automated alerting and proactive maintenance scheduling from outage history cut unplanned outages by 30%. A Wi-Fi coverage proposal for low-population areas raised regional revenue by 15%.
Education
B.E. Electrical (Telecommunication) Engineering
National University of Sciences and Technology (NUST), 2014 to 2018
Certifications
- Microsoft Certified: Azure AI Fundamentals
- Databricks Accredited: Generative AI Fundamentals
- PL-300: Microsoft Power BI Data Analyst
- Data Analysis with Python: Pandas and NumPy
- Making Data Driven Decisions (Google)
- Intermediate DAX in Power BI (DataCamp)
- Data Manipulation in SQL (DataCamp)
Toolbox
Agents and RAG: LangGraph, LangChain, Tool calling, Text to SQL, Hybrid retrieval, Rerankers, FAISS, Pinecone, Azure AI Search, BM25. Cloud models: Claude, GPT, Gemini, Azure OpenAI. Local models: Qwen, Gemma, gpt-oss, llama.cpp, Ollama, LM Studio, Whisper. Build: Python, FastAPI, Flask, Streamlit, Next.js, React Native, Docker, Git, n8n, Claude Code, Codex. Data and ML: SQL Server, pandas, scikit-learn, statsmodels, Power BI + DAX, Redis, Firebase, Neo4j. Cloud: Microsoft Azure, Google Cloud Platform, Vercel.
Contact
FARHAN ALI
Have a system that needs to work in the real world? I would like to hear about it.