zAInMuhammad Zain Ahmad

AI Engineer
Riyadh, Saudi Arabia

Muhammad Zain AhmadAI Engineer · Riyadh
Let’s talk

Portfolio

I build your AI. And make it work for you.

I’m Zain. I build tools that make answers traceable, agent behavior inspectable, and everyday work easier.

Explore the work

Selected work

Built to be
examined.

Two working tools. One tests AI. The other keeps track of research and its sources.

Concept artwork: lime streamlines passing through a graphite wind tunnel, bending around a disruption.
Agent evaluation Concept artwork

Test the behavior

Agent
Wind Tunnel

A stress test for AI assistants.

Give an AI assistant the same task twice: once with working tools, and once with a deliberate problem. A side-by-side report shows how its response changed, helping you spot weaknesses.

Python · CLI · JSON · HTML reports
How it works

Each run starts with fresh agent state. A fault is inserted at a specified tool call; the trace records requests, observations, final answers, and whether that fault was reached.

Test cases
Transient tool errors, stale results, and untrusted instructions inside a tool response.
Output
JSON traces, a paired summary, and an HTML report. The pass rate counts only valid comparisons.

The offline demo uses scripted decisions and fixture data. An optional OpenAI adapter connects real tool calls to the same workflow.

Research memory History intact

Keep the evidence

Memory
Atlas

Research notes that remember their sources and corrections.

Save research notes, find relevant passages, and link facts to their sources. When you correct a fact, its earlier version stays in the history. You can see what changed and where each version came from.

Python · SQLite · CLI · Docker
How it works

Imports text and Markdown into SQLite, stores source versions and passage hashes, and retrieves matching passages through lexical search. Corrections supersede current facts while preserving earlier values.

Workflow
Ingest a source, inspect passages, record a cited fact, then ask a question or review its correction chain.
Evidence
Passage citations, source paths, versions, and hashes. Optional model answers expose only validated citations.

Offline search and fact history remain available. The optional OpenAI path returns clear no-evidence or rejected states when needed.

Approach

A useful answer
needs a clear
trail of evidence.

I work across Python, model APIs, retrieval, and structured data. I care about what happens after a promising demo: whether an answer can be traced to a source and a failure can be reproduced.

Start with a clear question.

Define the task and what a useful result looks like before choosing the tools.

Make failure inspectable.

Record the steps, introduce controlled faults, and compare what actually happened.

Keep the source in view.

Connect answers to passages and preserve the history when evidence changes.

PythonSQLLLM APIsRetrievalDockerSQLiteAgent evaluation

Experience

From workflow
to outcome.

Internal AI applications, data work, and the engineering that connects them.

Apr 2025 to Jul 2026

Resident AI Engineer /
Software Engineer

Graphene Communication · Lahore, Pakistan

Built and locally deployed an internal lead-generation system using model APIs, retrieval, and historical records. Packaged the application and supporting services with Docker.

180%improvement in lead-collection efficiency
38%increase in conversions

Dec 2024 to Mar 2025

Technology Intern

ILM Fund, University of Management and Technology · Lahore, Pakistan

Used ChatGPT’s data-analysis tools to clean and organize records, streamline data entry, and explore ways to extend the fund’s reach. Supported everyday technology and social media work.

Contact

Let’s make
useful things.

Interested in practical AI systems, evaluation, or the details behind these projects? Let’s talk.

LinkedInGitHub