Ahmed Ashfaq

Capgemini · Client: Saudi Aramco · Saudi Arabia

Embedded with Saudi Aramco procurement. I sit with the teams, map how the work actually runs, then turn those workflows into production AI: agents, RAG, and automation across supply chain.

• Designed the agent core as a LangGraph multi-agent system, with an intent classifier routing work across procurement domains.
• Built multi-domain workflows for purchase requests, vendor bidding, supplier 360 and scope of work, with multi-hop follow-up so a user can refine a question without starting over.
• Added a Redis memory layer for short-term session context and long-term conversation history.
• Shipped an evaluation loop that scores each answer and regenerates when it falls below a quality threshold.
• Integrated an LLM gateway for token and cost control, routing across Gemini and Aramco's on-prem Qwen.
• Stood up vector search, RAG and document ingestion pipelines, plus MCP servers so agents can call enterprise tools safely.
• Deployed on GCP with FastAPI, observability, and the security controls the environment requires.

MindLink · Denmark · Remote

Led AI product delivery for two US clients: 4xRobots' lead engine, and MyGoat's voice agent that runs outreach calls end to end.

• Built the 4xRobots lead engine to pull thousands of prospects each week from Apollo, the open web, LinkedIn and Twitter, then run a multi-step classifier to keep only ICP-fit accounts.
• Added company research, competitor analysis and C-level contact enrichment, then scored every account with a custom fit algorithm before pushing the best ones into HubSpot with a follow-up sequence attached.
• Shipped MyGoat's AI voice calling system to reach about 500 prospects a day across airports, agriculture and similar site types, answering questions, capturing interest and sending a proposal on the call.
• Handled positive and negative outcomes in the same flow: a 21-day follow-up cascade through the sales cycle, human escalation when needed, and a meeting-invite link when the prospect was ready.

Code Reroute · Canada · Remote

Put machine learning, deep learning and NLP into live products: recommendations, content detection, a support chatbot, sentiment, and the automation around them.

• Shipped personalised Reels recommendations with ML and deep learning, tied to what people actually watched, not a generic feed.
• Built NLP content detection so unsafe, off-policy or low-quality posts could be flagged before they hit the feed.
• Shipped a product chatbot and a sentiment pipeline so support and content teams could act on what users were saying.
• Owned integration: models into the product, then the workflows around them so the team was not babysitting jobs.

Pioneer Ideas · Jeddah, KSA · Remote

Led a team of four on recommendation and inventory forecasting, with MLOps so models could be retrained and shipped without heroics.

• Built recommendation and forecasting models used for inventory decisions, not just a notebook metric.
• Set up training, tracking and CI/CD on AWS so a new model version could reach production on a schedule.
• Coached the team on evaluation and rollout so we were not guessing whether a model was better.

Radical X · United States · Remote

First production ML work: a post recommender and the user-behaviour analysis behind it.

• Improved a TensorFlow recommender by 12%, so people saw posts closer to what they already engaged with.
• Used Pandas and NLP on session behaviour to tighten personalisation instead of guessing features.