MHND LABS
The ledger

Field work.

Most of my client work starts the same way: build the thing first, then talk. This page is the ledger. Live demos built to win specific briefs, each computing something real, and case studies from client work, anonymized. No mockups: every demo below is live.

65 live demos5 case studies7 categoriesnewest 2026-10
Field status No mockups is a testable claim. Keep JavaScript on and this line probes every endpoint below from your browser.
Case studiesClient work, anonymized
CASE 01contracted · 5.0★

A card matcher that refuses to guess

for a trading-card arbitrage platform (US) · completed June 2026

Cross-marketplace listing matcher against a live product catalog, audited and verified row by row. The matcher abstains rather than mismatches: it commits only when it is sure and routes everything else to a review queue. I built the labeled sample, verified the client's production matcher against it, then designed and measured the backlog drain.

  • 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
99.4%
precision, priority stratum, verified
CASE 02contracted · ongoing

Making a news site legible to answer engines

for a Singapore business-news publisher · shipped to production June 2026

Full GEO upgrade of a live WordPress news site: speakable and 5W1H structured data on every new article, a machine-readable citation layer, and a site-wide heading-hierarchy fix. Shipped on a live production theme; the markup was validated with schema test scripts and the schema.org validator before it went live. The client then kept me on to measure the results in live AI answers each quarter.

  • 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
3
AI answer surfaces with a recorded citation, June 2026
CASE 03contracted · ongoing

21 gates for a fine-tuning corpus

for a conversational-AI product · 3 of 4 milestones delivered and accepted

Quality system for a supervised fine-tuning dataset: 21 automated checks (20 batch gates plus a cross-batch drift gate) covering schema, voice fingerprint, grounding, brand and distribution, with an LLM-judge review fleet on top. Rebuilt the weakest gate's classifier early on so it recognizes the shapes that actually occur.

  • 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
21
automated quality checks
CASE 04contracted · ongoing

A water-contaminant atlas that runs offline

for a water-quality client (North America) · US atlas delivered June 2026, EU atlas September 2026

Contaminant pipeline across three US regulatory sources (UCMR5, the Water Quality Portal, EPA NPDES discharge reports), 155 analytes with PFAS as the main thread. Every row keeps its source and native record ID, so any figure can be re-derived. Below-detection-limit values get censoring-aware statistics (Kaplan-Meier where the data supports it, the low tail suppressed where censoring is too heavy) instead of half the detection limit. Delivered as an offline single-file atlas by design, so there is no public URL. A second atlas over three European sources followed in September.

  • 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
3 + 3
public sources, two offline atlases
CASE 05contracted · 5.0★

Seven statement formats, one dashboard

for a financial advisory firm · completed April 2026

Commission statements in seven provider formats, mapped into one Excel dashboard with a tab per provider and a summary on top. Built as a Python pipeline wrapped in a Claude Agent Skill and handed over for the client to run from Claude on his own machine.

  • 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
7→1
provider formats, one dashboard
The demo shelfBuilt to win real briefs

Each of these was built for a specific job posting: a working answer to a real brief, not a template. They are pitch demos, presented as exactly that. Prospect-specific data has been checked before listing; a few strong builds stay off this page for confidentiality.

Agents & MCP12 pieces
W-582026-08

Gateway Harness Console

Spec in, loop to green: a model router walking real fallback chains, file-loaded skills, budget caps, an honest stop on a contradictory spec, and knowledge earned only with test provenance. 25 assertions at /api/selftest.

Vercel FunctionsVanilla JSDeterministic engine
↳ an AI gateway harness brief (US)Live ↗
W-332026-07

AI CGO Console

AI Chief Growth Officer console: weekly brief with KB citations, creator re-ranking, human approval queue, Claude/ChatGPT adapter swap, provider-free KB export.

Next.jsTailwind 4
↳ an AI SaaS growth-agent briefLive ↗
W-342026-07

Bilingual Retrieval Workbench

Arabic and English retrieval scored on a fixture eval set: normalisation lifts Arabic recall@1 from 0.688 to 0.938, with adversarial tenant probes and a tokenizer-driven cost model. 31/31 selftest.

Next.jsTypeScriptBM25RRF fusionbyte-level BPE
↳ a bilingual learning-platform RAG briefLive ↗
W-352026-07

Cowork Scholar Plugin

Working Claude Cowork/Code plugin for academic teams: deterministic citation integrity check, reading packs, in-browser selftest on a synthetic corpus.

Claude plugin formatPython (stdlib)Static JS
↳ an academic Cowork plugin briefLive ↗
W-422026-07

MCP Setup & Handoff

Live .mcp.json generator: toggle the stack's MCP servers, get a validated config with per-OS install steps and a credential checklist.

Next.jsMCP
↳ a Claude Code + MCP setup brief (US)Live ↗
W-172026-06

Aegis Agent Console

Self-hosted Claude Agent SDK harness: coordinator plus sub-agent, cross-session memory recall, pause/resume, and a live rubric grader that fails a first draft and passes the revision.

Next.jsClaude Agent SDKOpenRouter
↳ a Claude-agent developer brief (US agency)Live ↗Code ↗
W-182026-06

AI Coding Security Advisor

Deterministic, citation-backed comparison of five AI coding assistants' enterprise security posture, scored against a Bitbucket + AWS context.

Next.jsvendor-doc citations
↳ an AI-security advisory brief (US SaaS)Live ↗
W-202026-06

Company Brain MVP

MCP-native internal-knowledge agent with per-employee memory isolation: two users' stored preferences visibly change the same answer.

Next.jsMCPMem0-style memory
↳ an MCP company-brain brief (US MSP)Live ↗
W-262026-06

O365 Automation Studio

Plain-English process in, working artifact out: picks the right Microsoft 365 surface and writes a verified-shape flow, script or Graph call with deploy steps.

Next.jsClaudePower Automate
↳ a Claude + O365 coaching brief (UK)Live ↗
W-122026-05

Procurement Agent Console

Photometric procurement console on a clearly-labeled mock SAP tenant: one messy RFQ in, a sourced and priced quote out.

Next.jsSAP-shaped connector
↳ a procurement-automation brief (Tel Aviv agency)Live ↗
W-022026-04

CoinQuant Research Engine

Autonomous crypto-backtest sweep: 1,200 backtests over real 2024 BTC and ETH data, natural-language prompt to validated strategy runs.

Agentic backtesterBTC/ETH data
↳ a crypto research-infrastructure brief (Dubai)Live ↗
W-052026-04

NRT: Rust SLM Inference Server

10K-LOC pure-Rust inference server: manifest-driven small-language-model cluster on Candle, 7 crates, 29/29 tests passing. Apache-2.0.

RustCandle
↳ a Rust LLM-infrastructure brief (US)Code ↗
Data engineering22 pieces
W-632026-09

Per-Video Revenue Attribution Console

Audits 500 synthetic video tracking links, dry-runs a bulk rollout with backups, matches four payment rails to leads by email, and credits delayed sales to the originating video. 33 in-browser assertions.

Vanilla JSSVG
↳ a per-video YouTube revenue attribution briefLive ↗
W-642026-09

Continuity

Tracking QA for a multi-client agency: nine checks run in the page across five synthetic clients, with a continuity trace and a client-facing report.

Single-file HTMLno build
↳ a digital marketing agency's tracking SOP and monitoring briefLive ↗
W-542026-08

Catalogue Standardisation Gate

Decodes Magento option IDs, repairs Cyrillic SKUs, normalises units to SI, classifies to ETIM/eCl@ss, generates FR/EN copy, then blocks any row failing six invariants.

Vanilla JSZero-dependency pipeline
↳ an industrial-parts distributor's Magento 2 catalogue standardisation briefLive ↗
W-562026-08

Follower Segmentation Preview

37k-follower Instagram export classified into six lead categories: extraction, confidence flags, in-page selftest.

Static HTMLVanilla JS
↳ an Instagram follower segmentation briefLive ↗
W-602026-08

PostGate: Content Rule Gate

Rules as checkable constraints plus a gate that fails the draft breaking one: 18 drafts, 8 rules, 6 unconstrained drafts caught, 32 assertions.

Vanilla JSVercel Functions
↳ an educational-content workflow briefLive ↗
W-612026-08

Presidio Extension Bench

Five custom Microsoft Presidio classifications benchmarked on a synthetic enterprise corpus: before and after precision, recall, throughput and worker scaling, with a 27-assertion selftest that recomputes in the browser.

Microsoft PresidioPythonspaCyPlaywright
↳ an enterprise data-classification brief (US)Live ↗
W-362026-07

Database decisions for AI builders

A live teaching workbench for founders: answer 4 questions to get a database recommendation, watch normalization decompose a real table, and race a summary table against a raw scan. Every answer framed for AI-assisted building.

HTMLJavaScriptclient-side compute
↳ a data-and-AI teaching communityLive ↗
W-372026-07

Dyadic Cross-Correlation Lab

168 dyadic time-series analyses: 16-lag CCF, mixed-effects inference, APA reporting; recompute all 2,688 coefficients in your browser.

PythonstatsmodelsJupytervanilla JS
↳ an HCI continuous-tracking study briefLive ↗
W-392026-07

Pipeline Audit Workbench

Twice-weekly financial pipeline audit, replicated live in the browser: landing, freshness, row-integrity, schema-drift and sanity checks on synthetic Fed-style series, chained audit log, 28-check selftest.

Vanilla JSVercel
↳ a financial data pipeline maintenance brief (US)Live ↗
W-452026-07

Rollup Workbench

Reproduces the mixed-depth hierarchy rollup bug (understate vs double-count) and proves the closure fix, with 12 invariants recomputed live.

Next.jsTypeScript
↳ an ERP analytics brief from a US manufacturerLive ↗
W-462026-07

Remittance Ledger Console

USD to EGP remittance slice on real Postgres: double-entry ledger, idempotent transfers, sanctions screening and an append-only audit trail, live in the page.

Next.jsPostgres (Neon)double-entry ledger
↳ a fintech backend application (Cairo)Live ↗
W-472026-07

Logbook Pipeline Werkbank

German phpBB threads to RAG-ready English Markdown: shielded translation, per-token fidelity gate, chunk-survival simulator, VLM triage.

Next.jssharp
↳ a UK engineering firm's forum-log transformation briefLive ↗
W-482026-07

P&L-by-Job Workbench

QuickBooks Online line-item job costing: the multi-project bill bug reproduced live, 4 business rules toggleable, a real formatted .xlsx generated on demand, 13-check selftest endpoint.

Next.jsexceljs
↳ a QuickBooks Online job-costing export briefLive ↗
W-492026-07

PPRL Linkage Workbench

Two firms find their shared clients without either seeing the other's list: CLK Bloom-filter encoding, Dice fuzzy matching, live threshold calibration, and a hardening curve priced in recall.

Next.js 16TypeScriptWebCrypto (HMAC-SHA256)
↳ a privacy-preserving record linkage briefLive ↗
W-502026-07

Provenance Sheet Workbench

License-screens 8 real text sources with quoted clauses, recomputes a shingle-dedup run in the browser, and exports a cleared-rows-only provenance CSV.

HTMLJSPlaywrightVercel
↳ an ML training-data sourcing brief (Romania)Live ↗
W-522026-07

Real-time Signal Engine Console

Streaming tick feed to 2-minute OHLCV bars, live indicators, rule evaluation and at-least-once webhook delivery, with fault injection and an acceptance selftest that recomputes every claim per request.

Next.js 16TypeScriptServer-Sent Events
↳ a futures market-data briefLive ↗
W-532026-07

TLCC Workbench

168 dyadic time-series analyses: a 16-point lag sweep per trial, peak-lag extraction, and a random-intercept mixed model, all recomputed in the browser and asserted against scipy and statsmodels.

Next.jsTypeScriptPythonnumpyscipystatsmodels
↳ an HCI time-series analysis briefLive ↗
W-162026-06

Advisor Intelligence Mapping

Pre-meeting producer-intelligence briefings assembled from five messy data sources. Synthetic data, no real PII, stated on the page.

Next.js
↳ an advisor-data intelligence briefLive ↗
W-212026-06

Cloud Data Pipeline Console

Messy operator data in, a clean warehouse fact table out: 6 locations, 2 regions, on-demand pipeline runs with BI on top.

Next.js
↳ a cloud data-engineering briefLive ↗
W-272026-06

Bank-Data Integration Workbench

Real Plaid sandbox flow (link token, Link, exchange, accounts) with a visible client/server secret boundary and an honest security checklist.

Next.jsPlaid SDK
↳ a US fintech integration briefLive ↗
W-292026-06

Tracking-Cockpit (German)

German-language analytics cockpit: GA4 dataLayer events, Consent Mode v2, server-side measurement and a dashboard layer, all clickable.

Next.jsGA4Consent Mode v2
↳ a Stuttgart analytics brief, in GermanLive ↗
W-302026-06

VIN Decoder for Indian EVs

Decodes Indian-market EV VINs across 16 manufacturers with a cited source under every field, a segment visualizer and a 22-assertion selftest.

TypeScript engineNext.js
↳ an EV-data brief (US)Live ↗
GEO & AEO9 pieces
W-652026-10

Pre-rebuild SEO and GEO read of a live Webflow site

315-fetch crawl of a B2B EV-charging site before its redesign: a stale sitemap, canonical loops, a self-cancelling French mirror and two Organization entities, with 27 checks recomputing every count in the browser, in English and Danish.

Pythonraw HTML crawlstatic HTML
↳ a European EV-charging software redesign briefLive ↗
W-592026-08

Answer-Engine Retrievability Audit

Twelve-signal retrievability audit of a live SaaS estate: marketing 71.4 percent, docs 11.3 percent, with 29 assertions recomputing every figure in-browser.

Vanilla JSVercel Functions
↳ an EU enterprise AI-platform vendor briefLive ↗
W-622026-08

AI Search Citability Audit

Scores whether an answer engine can tell WHO wrote a page, computed from raw HTML, with a live SSRF-guarded analyzer for any public URL.

Nodeschema.org JSON-LDrobots.txt parsingVercel Functions
↳ a US public relations agencyLive ↗
W-412026-07

ITAD AI Visibility Audit

13-signal AEO/GEO scorecard for an IT asset disposition vertical; every score recomputed in-browser by a 27-check selftest.

Static HTMLVanilla JSVercel
↳ an ITAD / AI search visibility brief (US)Live ↗
W-192026-06

Answer Engine Readiness Audit

Fetches any URL server-side and scores 13 AEO/GEO signals with cited evidence and an impact-by-effort fix matrix. No keys, no LLM, always live.

Next.jsSSRF-guarded fetcher
↳ a B2B AEO strategy brief (Canada)Live ↗
W-252026-06

Listed-REIT GEO Audit

Vertical GEO audit on a representative listed-REIT entity: real schema generation, live validation, and an AI answer with versus without entity scaffolding.

Next.jsschema.org
↳ a Belgian listed-REIT GEO briefLive ↗
W-112026-05

Finance Link Inventory

A 55-publication link inventory with every metric cited: floors, ceilings and per-site pricing, laid out as a working editorial sample.

Next.js
↳ a London SEO-operations briefLive ↗
W-132026-05

Site Refresh + SEO Migration Plan

SEO migration control room built from a real site crawl: redirect map, verification and rollback sequence to keep all 3,087 referring domains.

Static HTML
↳ an IoT-platform site-refresh briefLive ↗
W-062026-04

Advertorial, Done Properly

Long-form German advertorial, format-faithful: labeled Anzeige, magazine structure and a conversion path in one editorial piece.

Static HTMLGerman copy
↳ a spec advertorial for a Shopify wellness briefLive ↗
Automation & ops10 pieces
W-552026-08

Content Cluster Command Board

Multi-site content ops board: cluster completeness, brief inheritance and publishing runway computed live from one synthetic dataset, with a 17-assertion in-page selftest.

HTMLVanilla JSVercel
↳ a multi-site AI content planning and WordPress publishing brief (AU)Live ↗
W-572026-08

Launch Funnel Router

Simulates a Webflow to Typeform to ActiveCampaign to Attio launch funnel: scored assessment, source routing, consent and suppression, deduped CRM upsert, 24 in-browser assertions.

HTMLCSSJavaScript
↳ a marketing ops launch-build brief (US)Live ↗
W-402026-07

HubSpot Properties Migration Workbench

Portal-to-portal contact properties migration: classified export diff, replacement-property plans for type changes, slot ledger, fresh-export validation, 16-check in-browser selftest.

HTMLJSVercel
↳ an enterprise HubSpot portal migration brief (US)Live ↗
W-442026-07

Multifamily AI Workshop Preview

A 96-unit synthetic deal underwritten with Claude: T-12 normalization, rent roll audit, lease discrepancy sweep, IC memo draft, with a 25-check in-browser recompute.

HTMLJavaScriptPlaywright
↳ a real estate investment firm's AI workshop brief (US)Live ↗
W-512026-07

Research Mentor Workbench

Teach-first mentoring preview: a worked business research prompt chain, a source verification gate that executes in the page, the ChatGPT to Codex graduation, and a 14-check selftest.

HTMLVanilla JSVercel
↳ a ChatGPT and Codex research mentoring postingLive ↗
W-222026-06

Closed-SaaS Webhook Router

Simulates GoHighLevel + Make webhook and API contracts end to end: seven real modules, idempotent retries, and two keyless live round-trips.

Next.jswebhooks
↳ a US insurance-automation briefLive ↗
W-242026-06

Lead Intake → CRM

Real-estate lead flow end to end: intake, qualification scoring (HOT 98 / WARM 58 / COLD 7), CRM record and the auto next-action.

Static HTMLscoring engine
↳ a US real-estate ops briefLive ↗
W-312026-06

Workflow Architect

Paste a documented process, get a before/after complexity scorecard (steps, handoffs, touch time) computed by a real classifier, not a template.

Next.jsClaude
↳ a US ops-architecture briefLive ↗
W-092026-05

AI Blog Production System

Airtable-native blog production line: six Run Script automations move briefs through drafts to review. The page is itself the proof artifact.

Next.jsAirtable
↳ an AI blog-production briefLive ↗
W-102026-05

Marketing Campaign Calendar

Three-altitude campaign calendar backed by a real public Airtable base: 12 campaigns, 5 teams, linked records you can open.

AirtableStatic HTML
↳ an enterprise Airtable brief (US)Live ↗
Documents & forms3 pieces
W-382026-07

Evidence Review Workbench

Legal evidence review on synthetic data: SHA-256 preservation at ingest, original versus derivative separation, duplicate occurrences kept rather than collapsed, deterministic privilege screening that only ever marks a candidate, role-gated counsel notes, an append-only audit log, and a production ZIP whose hash manifest matches the ingest hashes. 31 of 31 invariants asserted live at /api/selftest.

Next.jsTypeScriptWeb CryptoJSZip
↳ a 72-hour confidential document and email review MVP briefLive ↗
W-282026-06

SuccessFactors Integration Console

Reproduces SAP SuccessFactors' OAuth 2.0 SAML-Bearer handshake server-side and reads real OData v2 shapes, with an honest map of what login data actually is.

Next.jsOAuth 2.0OData v2
↳ an enterprise HR-data brief (US SaaS)Live ↗
W-322026-06

XFA → AcroForm Workbench

Converts a dead XFA government form into a working AcroForm in the browser: JSON auto-fill of all 20 fields, live edits, real /Sig signing, clean flatten.

Next.jspdf-lib
↳ a US government-forms briefLive ↗
Vision & media2 pieces
W-432026-07

Messy retailer photos in, product-faithful images out

Segmentation-first cleanup of retailer product photos with a live faithfulness gate: structural similarity plus edge overlap recomputed at an API route, PASS or FAIL per image.

Next.jssharpCV QA
↳ an image-to-3D product brief (US)Live ↗
W-012026-04

Automated Ad Factory

Script in, eight ad variants plus storyboards out: five parametric templates rendered server-side with a multi-modal Claude pipeline.

Next.js 16Claude multi-modal
↳ a direct-response creative brief (US)Live ↗
Product & full-stack7 pieces
W-232026-06

Ghost Multilingual Theme

Faithful static replica of a trilingual Ghost theme: /, /ru/ and /ar/ RTL routes with real hreflang, canonicals and og:locale, plus a live SEO-head inspector.

Ghost themehreflangRTL
↳ a multilingual publishing brief (Beijing)Live ↗
W-142026-05

Physio Training Vitrine (French)

French-language training vitrine with an interactive e-learning to in-person funnel, built from the practice's real brand color.

Next.jsFrench copy
↳ a spec redesign pitched to a French physio-training practiceLive ↗
W-152026-05

Soundscape: Feel the Frequency

Experiential audio-room concept site: three frequency programs and a booking flow. A pure design piece, presented as one.

Next.js
↳ a design-concept piece for an advanced web briefLive ↗
W-032026-04

Arabic Education Platform

Bilingual Arabic-education site with a Stripe subscription flow, English and Arabic content side by side.

Next.jsStripe
↳ an Arabic-education platform briefLive ↗
W-042026-04

Arabic RTL Dashboard

Bilingual EN/Arabic operations console with true RTL: mirrored layout, Arabic script and numerals, and live telemetry views.

Static HTMLArabic RTL
↳ an Arabic IoT-dashboard brief (Saudi Arabia)Live ↗
W-072026-04

Webflow CMS Templates

One template, one CMS collection, four school pages: the structure-to-page architecture as a working proof of concept.

Static HTMLWebflow CMS
↳ a Webflow CMS briefLive ↗
W-082026-04

Patient Digital Identity (FHIR)

Patient digital-identity explorer over 1,053 FHIR R4 resources and 17 LOINC series, with a 5-axis wellness scorer and Claude streaming over AI SDK v6.

Next.jsFHIR R4AI SDK v6
↳ a health-AI brief (US)Live ↗