Merna Hesham

I built and deployed three production AI applications so prospective clients could test real systems rather than watch a recording: a financial report analyzer that answers grounded questions over uploaded financial documents, a pharmacy copilot for medication and stock queries, and a sales lead prioritizer that scores and ranks inbound leads against a qualification rubric.

Each runs as an isolated containerized service on Google Cloud Run in europe-west1, keeping processing inside the EU. I designed the access model so no endpoint is publicly callable: the Next.js front end issues a short-lived session token and every service validates it before doing any work, so a leaked back-end URL is useless on its own.

Retrieval uses embeddings stored in Supabase with pgvector, and prompts and grounding rules are written so that answers cite the source document rather than improvising. I owned the full stack, including retrieval design, prompt engineering, containers, deployment, and UI.

I architected an internal knowledge system that turns a growing library of business and technical documents into queryable, structured memory. Two inference engines run against the same pipeline: a cloud engine on the Anthropic API for high-quality concept extraction and reasoning, and a local engine running qwen2.5 through Ollama for anything that should never leave the machine.

Documents are chunked and embedded with nomic-embed-text, stored in Supabase with pgvector, and extracted concepts are linked into a knowledge graph that surfaces how decisions, projects, and constraints connect. On top of it, I built an internal Next.js admin dashboard with a table browser, a graph view, a content pipeline, and a scheduling module.

The design principle was routing over lock-in: any document can be processed locally or in the cloud based on sensitivity, and the retrieval interface is identical either way, so no vendor sits on the critical path.

I built a transaction-screening engine that assesses crypto wallets and transactions for money-laundering risk before funds are accepted. The core is a deterministic risk pipeline: wallet footprint analysis, exposure tracing through counterparties, and sanctions/darknet-entity screening, combined into an explainable risk score with the specific evidence behind every flag, so that a compliance officer can see why a wallet was rejected, not just that it was.

Correctness was the hard requirement, so I engineered it like a safety-critical system: 200+ automated tests covering the full pipeline, and every risk rule verified against live blockchain data rather than fixtures. A key design principle was to make completeness explicit: every screening stage reports what it checked and what it could not check, so a data-source outage degrades to a documented gap rather than a silent pass.

I owned the entire build: risk model design, data source integration, the scoring engine, and the test infrastructure that proves it behaves as specified.