All the work
Client results first, then every working demo, grouped by the problem it solves. Each demo was built for one job post and runs live. Client and product names are withheld.
5 client results and 43 working demos in 7 groups.
- Top 10% of Upwork freelancers (Top Rated)
- 100% Job Success
- Every client review: 5.0
Check all three on my Upwork profile ↗, or hire me there.
Client results
-
A matcher runs in production. Can you trust what it commits?
99.4% precision, priority stratum, verified
I verified the client's production matcher against the client's own labeled sample.
for a trading-card arbitrage platform (US), completed June 2026
How this was measured
- 99.4% precision on the priority stratum of a 6,190-listing labeled sample measured on the matcher as shipped (about 96.6% across all strata)
- 2,481 backlog listings committed, each landing on exactly the id the audited dry-run proposed; risky cases held for review
- Read-only against production; every write stayed with the client's engineer
“Muhannad is great to work with and is a clear communicator which I value a lot. Will hire him again.”
Upwork client review, June 2026 -
AI answer engines must be able to read your site and cite it.
3 AI answer surfaces with a recorded citation, June 2026
Full GEO upgrade of a live WordPress news site.
for a Singapore business-news publisher, shipped to production June 2026
How this was measured
- Speakable and 5W1H schema on every new article, plus a citation layer
- Markup validated with schema test scripts and the schema.org validator before shipping
- Shipped on production WordPress (TagDiv) with the site live throughout
- Recorded citations on three AI answer surfaces in a June 2026 capture: Perplexity, Google AI Mode and AI Overviews
“What I particularly liked was that you actually explained what you were doing and you had lots of evidence to show that you were getting the green ticks on all of these things.”
the publisher, at the final sign-off call -
A fine-tuning dataset needs quality checks before it ships.
21 automated quality checks
Quality system for a supervised fine-tuning dataset: 21 automated checks.
for a conversational-AI product, 3 of 4 milestones delivered and accepted
How this was measured
- Client's 3,109-entry source corpus scanned, drift-mapped and calibrated
- Shape classifier coverage raised from 16.4% to 59.5% after the rebuild
- 763 new entries delivered and accepted across 3 milestones
- Every flag lands in a per-batch findings file, fully traceable
“I appreciate you putting the voice issue into actual numbers rather than treating it as subjective taste.”
the client, in a project message, May 2026 -
Public data from several sources must become one dataset, with each figure traceable.
3 + 3 public sources, two offline atlases
Three US regulatory sources unified into one provenance-tracked dataset.
for a water-quality client (North America), US atlas delivered June 2026, EU atlas September 2026
How this was measured
- Three US regulatory sources unified into one provenance-tracked dataset; three EU sources in a second atlas
- Censoring-aware statistics for below-detection-limit values, not substitution
- Ships as one HTML file the client's team opens without a server
- A regression found in the build was fixed before delivery; a 45-check verifier now fails the build loudly if it recurs
“We are really impressed with the work you did here”
the client, after a team review, June 2026 -
Statements arrive in seven formats and must land in one dashboard.
7→1 provider formats, one dashboard
7 provider formats ingested and mapped into one workbook.
for a financial advisory firm, completed April 2026
How this was measured
- 7 provider formats ingested and mapped into one workbook
- Writes only append below existing rows; source files and the client's own workbook are never touched
- Handed over as a Claude Agent Skill for the client to run
“Excellent work and went above and beyond in terms of service and time”
Upwork client review, April 2026
Working demos
This page can check each live demo from your browser.
Our numbers or records are wrong
Data pipelines and record matching. Data cleaned, matched and reconciled, with each error shown.
-
Built for an image-to-3D product brief (US)
Next.js, sharp, CV QA
-
Built for a QuickBooks Online job-costing export brief
Next.js, exceljs
-
Built for a privacy-preserving record linkage brief
Next.js 16, TypeScript, WebCrypto (HMAC-SHA256)
-
Built for an industrial-parts distributor's Magento 2 catalogue standardisation brief
Vanilla JS, Zero-dependency pipeline
-
Built for an Instagram follower segmentation brief
Static HTML, Vanilla JS
-
Built for an enterprise data-classification brief (US)
Microsoft Presidio, Python, spaCy, Playwright
-
Built for an ERP analytics brief from a US manufacturer
Next.js, TypeScript
-
Built for a UK engineering firm's forum-log transformation brief
Next.js, sharp
-
Built for an HCI time-series analysis brief
Next.js, TypeScript, Python, numpy, scipy, statsmodels
-
Built for a cloud data-engineering brief
Next.js
-
Built for an EV-data brief (US)
TypeScript engine, Next.js
We cannot tell if it still works
Audits and automated quality checks. Audits and checks that run again on each load.
-
Built for a per-video YouTube revenue attribution brief
Vanilla JS, SVG
-
Built for a digital marketing agency's tracking SOP and monitoring brief
Single-file HTML, no build
-
Built for a data-quality reconciliation brief
Python, SQL (SQLite), static HTML
-
Built for a financial data pipeline maintenance brief (US)
Vanilla JS, Vercel
-
Built for a futures market-data brief
Next.js 16, TypeScript, Server-Sent Events
We want AI to do real work, and we must trust it
AI agents and retrieval systems. Agents and AI search with rules, tests and stated limits.
-
Built for a DTC story ad brief (US)
Claude skill format, Python (stdlib), Static JS
-
Built for an educational-content workflow brief
Vanilla JS, Vercel Functions
-
Built for a bilingual learning-platform RAG brief
Next.js, TypeScript, BM25, RRF fusion, byte-level BPE
-
Built for a private-markets knowledge-graph brief
Oxigraph (WASM), SPARQL 1.1, OWL/SKOS, SHACL-SPARQL, Vercel Functions, AI Gateway
-
Built for an AI gateway harness brief (US)
Vercel Functions, Vanilla JS, Deterministic engine
-
Built for an academic Cowork plugin brief
Claude plugin format, Python (stdlib), Static JS
-
Built for an MCP company-brain brief (US MSP)
Next.js, MCP, Mem0-style memory
-
Built for a Claude + O365 coaching brief (UK)
Next.js, Claude, Power Automate
-
Built for a procurement-automation brief (Tel Aviv agency)
Next.js, SAP-shaped connector
AI answers do not name us
AI search visibility (AEO and GEO). Visibility in AI search, measured on real answers.
-
Built for a Shopify sports-supplement brief
Perplexity Sonar API, OpenAI web search, Gemini grounding, Vercel AI Gateway
-
Built for an AEO strategy brief
Next.js, SSRF-guarded fetcher
-
Built for a cleaning-business software brief (US)
Vercel AI Gateway, AI SDK 6, Vanilla JS, Python cross-check
Work falls through between our tools
CRM and workflow automation. CRM, lead and content flows that do not drop a record.
-
Built for an enterprise HubSpot portal migration brief (US)
HTML, JS, Vercel
-
Built for a multi-site AI content planning and WordPress publishing brief (AU)
HTML, Vanilla JS, Vercel
-
Built for a launch-funnel brief and an outbound lead-pipeline brief (US)
HTML, CSS, JavaScript
-
Built for a real estate investment firm's AI workshop brief (US)
HTML, JavaScript, Playwright
-
Built for a US insurance-automation brief
Next.js, webhooks
-
Built for a US real-estate ops brief
Static HTML, scoring engine
-
Built for a US fintech integration brief
Next.js, Plaid SDK
-
Built for an enterprise HR-data brief (US SaaS)
Next.js, OAuth 2.0, OData v2
-
Built for an AI blog-production brief
Next.js, Airtable
Our old forms and documents will not process
Document and form processing. Old formats made to work again.
-
Built for a US government-forms brief
Next.js, pdf-lib
-
Built for a 72-hour confidential document and email review MVP brief
Next.js, TypeScript, Web Crypto, JSZip
-
Built for a health-AI brief (US)
Next.js, FHIR R4, AI SDK v6
Our lessons and training do not hold attention
Interactive learning content. Learning tools that listen, branch and score, in English and Arabic.
-
Built for an English self-study programme on Moodle
H5P add-on, Moodle 4.5, wav2vec2 phoneme model (ONNX), espeak-ng WASM, Vercel
-
Built for an Arabic corporate-training brief (Saudi Arabia)
HTML, CSS, JavaScript, Claude Code