My IT agent is the official support channel for 200+ staff. I found the cron job eating 84% of a company's AI bill. Self-taught — from waiter to AI architect.
W3J-Bot is a live LLM trained on my shipped work. Ask it anything, or read the résumé and the code.
What clients and peers say about the work — paraphrased from real feedback.
“A 20-year veteran quoted us $15K to automate our sales data entry and accounting. Jewel built it with Google Sheets and Forms — for free. He calls himself a 'ducttape programmer.' Honestly? It just works. That's rare.”
“Our AI and API costs were spiraling, with usage caps throttling us at the worst times. Jewel re-architected the whole thing — capped us near $2K a month with no usage or rate limits. Same capability, a fraction of the bill.”
“He talks in plain language, knows every tool worth knowing, and always recommends the right stack — even when it's free open-source instead of something he'd profit from. He'll self-host and match the scale of any expensive cloud. Passionate over profit — a dev you can truly rely on.”
Products shipped in production
Testimonials paraphrased from real client & peer feedback; names withheld on request.
Six products shipped and in build under the studio — each one live software with users, not slideware.
WhatsApp · Multi-tenant · Human takeover
A WhatsApp-native AI employee for small businesses. It answers in 5 languages including Manglish from the business's own FAQs, captures leads, books into Cal.com and hands the chat to a human when needed. Live at mybijou.xyz.
Career workspace · Voice interview coach · Human-approved agents
AI career workspace — practise interviews by voice and get scored feedback, discover and rank jobs across ~44 company boards, and generate tailored resume and cover-letter PDFs. Agent actions need your approval. Formerly InterviewOS; relaunching soon.
Prototype · Voice concierge · Gemini Live
A voice maître d' on Gemini Live that adds dishes to your cart through tool calls, plus a text concierge, voice booking and AI-written menu copy. Single-venue prototype.
HR & payroll · Malaysia · Face clock-in
Started in 2023 as my own Sheets + Apps Script system for a 50+ staff café: payroll admin from ~40 hours a month to ~2. Now a Malaysian HR & payroll prototype — face clock-in, EPF/SOCSO/EIS/PCB estimates, PDF payslips.
Resumes · Cover letters · ATS
Turns pasted experience into a one-page, ATS-friendly resume and a matching cover letter, laid out like a professional template.
Developer tooling · Landing pages
Reads a GitHub repository and generates a developer-focused landing page for it.
Smaller things I shipped to learn, test an idea, or win a weekend. Code is public.
Two field deployments and four of my own builds, simplified. Pick one to trace a request from first message to outcome — and see what I'd still fix.
The official IT support channel for 200+ staff. Built for the place people already work, not as a ticket form with a model in front.
Measured in systems people use every day, not in demos. Employers and clients stay unnamed; the work does not.
Instrumented tokens per process instead of per project. One unmonitored cron job was 84% of an organisation's LLM bill; I removed it, then took a further 50–70% off what remained.
An IT service agent in Google Chat, in two languages, that resolves first and raises a Jira ticket only when it should. It grades itself after every ticket; the number comes from that loop.
Azure OpenAI retrieval and classification for high-volume enterprise documents, with an explanation layer operators could interrogate. No compliance findings during the engagement.
Client outcomes, anonymized, live on the company site: logistics compliance · voice sommelier · PunchClock
From the operations floor to enterprise AI architecture — fluent in both the POS and the ML pipeline.
Fully autonomous, event-driven multi-agent state machines. Decoupled agents with strict schema boundaries, async orchestration, programmatic self-healing, and deterministic rollback gates — built to run with zero human in the loop.
Telephony pipelines that wake on inbound, route via state machines, and drive conversations to completed actions. Real-time transcripts, warm-transfer to humans, SDK-driven call control, and CRM mutations on drop.
Multi-tenant SaaS with real users — Bijou AI, MenuMuze, HireOS, PunchClock — built on Next.js, TypeScript, and Supabase. Production traffic, payments, and audit-ready logging.
Containerised agent runners on GCP, AWS, and Railway. MCP tool gateways with inspectable permissions, persistent queues, dead-letter handling, and durable state for long-running autonomous tasks.
PR review agents, on-call issue/bug fix agents, and team-of-agents orchestration. Approval-gated execution, deterministic validator constraints, and a Critic agent in the loop before any state mutation.
Long-memory layers with self-evolving context, vector retrieval feeding grounded responses, and structured memory separate from conversation state. Designed for agents that need to remember across sessions and not just turns.
Technology isn't about disruption. It's about salvation.
When doctors couldn't diagnose my mother's condition, I fed her medical records into AI systems. The AI identified a treatment pathway they had missed. I organized an international medical journey. Today she is alive because of technology. This is why every line of code matters.
Born in Saisanga, Faridganj, Bangladesh — a village without electricity until I was in Class 5. No computer until college. Zero formal CS education. A perfect 5.0/5.0 GPA despite cycling 5–6 miles daily through harsh conditions. Self-taught from zero to building enterprise AI systems.
Won first place in a competitive English course where the prize was a laptop, against students from Dhaka with far better backgrounds. The laptop never came, but I built fluent, neutral-accented English. The real prize is the skill you develop.
In 2010 I was named President Scout by the President of Bangladesh — one of three chosen from about a thousand candidates in my district, after three years in the programme.
Progressed from waiter to Operations Manager — 50+ staff, RM1M+ annual revenue. I lived the industry's pain: clumsy POS, manual inventory, error-prone payroll, and led the digital transformation that fixed them. I don't build features. I solve problems I've personally lived.
I'm still a mixologist and wine sommelier at heart. A drink is judged in one sip, so I hold software to the same bar: if an interface needs explaining, it goes back. The agent does the building; I decide when it's actually done.
The things AI buyers actually ask before they commit.
Hiring for an AI engineering role, or want to compare notes on agents in production? Reach me directly below. For client work, W3J LLC handles scoping and delivery.
Work with W3J LLCBook a free 15-30 min strategy call directly on my calendar
A real LLM with the full portfolio, products, and metrics baked in. Answers in the same control-room voice as the site. For a personal reply, use the form below.
"I build systems that solve real problems, not just technical demos. Let's connect if you want AI agents built, deployed, and managed for you by someone who understands both the business floor and the tech stack."
— Muhammad Nurunnabi Jewel (W3JDEV)