Ahmad Yusuf R.

AI is easy to demo and hard to trust. I'm an AI engineer. I build AI agents, automations and the software around them, and I make them check their own output before anything is saved. WHAT I BUILD • AI agents & business automation: orders, shipping, CRM, outreach and admin moving between your tools (n8n, Make, Python), with Claude only in the steps that need judgment. Every path has validation checks, retries and logs. • LLM apps & APIs: web apps and APIs (Next.js, TypeScript, Python) with Claude inside. Structured JSON output, tool calling, evidence checks and token cost control. • Claude Code agents: multi-step agents with validation checks, retry rules and project instructions, so they follow your rules every time. • MCP servers & connectors: Claude connected to your CRM, database, Google Sheets, QuickBooks or internal API. Clear permissions, read-only by default, and an approval step before anything is written. • Claude Skills & Cowork setup: the processes your team repeats every week written as Claude Skills and tested on real examples, so everyone gets the same result. WHO I HELP • Software teams and founders who need AI inside their product: LLM features, AI agents, or taking an app built with Claude Code to production. • Small businesses and agencies that run on connected tools: orders and shipping, CRM and outreach, support, and daily admin that can be automated. • Finance and operations teams, where one wrong number is expensive: invoices, ledgers and month-end checks. If your case is a little different, send it anyway. I will tell you honestly whether I can help. HOW I MAKE AI TRUSTWORTHY On one Claude structured-output pipeline, my own validation-error message contained a number. Claude read it back on retry and filled a field with it. Nine fabricated orders passed validation. I fixed it structurally: every value now needs verbatim evidence from the source document, or the record fails. Result: zero fabricated values across 50 test cases, and it held against 10 adversarial injection attempts. RECENT WORK • Storeganise → ShipStation (Make), client project. New valet orders arrive by webhook and are created in ShipStation automatically, with duplicate checks and a retry on failure. A second scenario pulls new invoices and line items and emails a clean summary. Rated 5.0. • AI PR & media outreach engine (n8n), client project. Scraping, LLM enrichment, personalised pitches and state tracking, so no contact is pitched twice. • Ledgerproof (Claude + MCP), reference build on synthetic ledgers. Claude reviews a ledger and every finding must quote the exact row, or it goes back once and then to a person. A web queue and a Claude Desktop MCP server handle approve and reject, with a dry run before saving. On 8 test ledgers, 3 runs each: 97.4% precision, 93.8% recall, 0 decoy rows flagged. Every miss is listed in the public repo. • Books Checker (Claude Skill), reference build on synthetic ledgers. A script runs the exact month-end checks, Claude judges the rest, and every finding must quote its ledger row. On 8 test ledgers run twice: 53 of 54 planted errors found and 0 false alarms (plain Claude: 6). • AI invoice → QuickBooks agent (Claude + n8n), reference build. Claude reads PDF invoices, code checks every value against the PDF and the math, then the bill is created in a QuickBooks sandbox and read back. On a 12-invoice test set: 120 of 120 fields extracted correctly, 0 wrong values written, and every invoice that failed a check went to review with the right reason. • Guarded Postgres MCP server, reference build. Claude can read and write a Postgres database safely, and all limits live in one policy file. In a test of 19 attack attempts (injection, data leaks, DROP TABLE), 0 succeeded and the data stayed unchanged. HOW I START 1. A written audit. You send me how the work runs today, and I send back a plan: what to build first, what it needs, and where a person should stay in charge. 2. A build in stages, with logging, retries and alerts, so failures are visible instead of silent. 3. Handover with notes your team can follow. You own everything I build. If a fixed rule solves your problem, I will tell you that you don't need AI. It is cheaper and more reliable. Certified by Anthropic: Claude Code in Action, Introduction to Agent Skills, Introduction to Model Context Protocol, Building with the Claude API, Claude 101, Claude Platform 101, Claude Code 101. Also Make Foundation and n8n Quickstart. Send me the process your team repeats, or the app that needs to get ready, and I will tell you what I would build first. AI Engineer | AI Agents | AI Automation | Claude | Claude Code | MCP | LLM Apps | n8n | Make | Python | Next.js | TypeScript | API Integration | CRM Automation | QuickBooks | Shopify