Projects

My client work, and the products I build at ProductArc, my AI product company. Each one follows the same outline: the business problem, the tech solution, the results and impact, then the visuals and the tech stack.

Client work

National workforce data, in plain language

UAE Federal Ministry · Client, contract

Business problem: Analysts needed answers from the national workforce database, covering 7 million people, and every question meant writing SQL by hand.
Tech solution: I architected a text-to-SQL assistant in English and Arabic on a federal transformation program, and rebuilt a failed forecasting platform in two weeks, about 70% faster, with row-level data kept out of the LLM.
Results and impact: 80 to 90% accuracy on more than 1,000 evaluated questions, presented to the ministry's leadership.
Visual assets: Screenshots not shown under NDA.
Tech stack: Text-to-SQL RAG Agents Privacy by design

Claims and medical translation AI, in production

GlobeMed, processing nearly $3bn in claims a year across 12 countries · Client, employee then contractor

Business problem: Claims had to be handled faster and medical documents translated accurately between Arabic and English, with no AI team in place.
Tech solution: I built intelligent document parsing and claims case summaries now in production in Bahrain and Saudi Arabia, plus the LLM evaluation framework behind them, and fine-tuned a MarianMT Arabic-English medical translation model in PyTorch on training data I grew from 1K to 15K records.
Results and impact: Running in two new markets on shared engineering standards. It started by pitching a working model to the CEO, was funded in six weeks, and grew into GlobeMed's AI/ML department.
Visual assets: Screenshots not shown under NDA.
Tech stack: Document AI LLM evaluation Fine-tuning MarianMT PyTorch

Products I built at ProductArc

Parlo

AI feedback and analytics platform · Live

Business problem: Long forms get skipped, and the free-text answers that do come back are hard to analyze.
Tech solution: A chat that replaces the form: an LLM asks each question, checks every answer and steers off-topic replies back, built end to end with a private workspace per user.
Results and impact: Live and open to anyone. Answers come back as structured data, with charts and an AI summary.

A respondent answering Parlo questions by chatting — Respondents answer by chatting. Off-topic replies are steered back and invalid answers rejected.

Parlo's AI suggesting questions of every type from a one-line topic — The AI drafts questions of every type from a one-line topic. You pick the ones you want.

Parlo insights: respondents, completion rate and average rating — Insights update as answers come in: respondents, completion rate, ratings and trends.

Tech stack: React FastAPI PostgreSQL OpenAI Docker CI/CD

Try it live Code

SYD

AI health assistant with built-in safety checks · Live

Business problem: A health assistant that guesses is dangerous. People need answers they can trace to a trusted source.
Tech solution: Retrieval over CDC and USPSTF guidance with a relevance cutoff, plus a second AI safety check that reviews each answer before it is shown.
Results and impact: Live. Every answer cites its guideline source, and a question without a trusted source is declined instead of guessed.

SYD answering a flu shot question with a cited CDC source — Every answer cites its guideline source and is safety-checked before it's shown.

SYD answering a screening question with ranked sources — Sources are ranked by relevance and shown alongside the answer.

SYD declining to answer a dosage question without a trusted source — No trusted source? SYD says so instead of guessing.

Tech stack: RAG Guardrails React FastAPI AWS

Try it live Code

TableTalk

Talk to your database with AI agents · Open-source prototype

Business problem: Teams wait on analysts for every data question, because the answers sit behind SQL.
Tech solution: A multi-agent text-to-SQL pipeline: agents plan, write and verify each query before it runs, and RAG builds the table docs and examples for a new schema.
Results and impact: An open-source prototype that sets itself up on a new database and ships with its own training studio and evaluation suite.

Tech stack: Multi-agent RAG Qdrant FastAPI React

Code

More projects

MeetScribe

AI Zoom meeting assistant

Business problem: After a meeting, nobody wants to write the notes, and decisions get lost.
Tech solution: A bot joins the Zoom call, transcribes it speaker by speaker, and an LLM turns the transcript into a recap email, run as background jobs on AWS.
Results and impact: Everyone gets a clear summary of who said what by email, minutes after the call, with no manual note-taking.
Tech stack: Zoom SDK Speech-to-text LLMs FastAPI Celery AWS

CapsLock

AI news agent

Business problem: Following the news on one topic means reading dozens of articles to find the few that matter.
Tech solution: An LLM agent that uses tool calling to gather the news for a topic and a date, rank it and cite every source.
Results and impact: A short, ranked digest with a link behind every point, and automated evals that score the agent's accuracy.
Tech stack: Python FastAPI LLM agents Tool calling Evals

Code

Want something like this built for your product? Book a 30-min call or see how we can work together.