Eleazar.

AI automation engineer.

I’m Eleazar. I build AI agents and real software with Claude Code that take the daily checks off your team.

Chapter One5 builds

Claude Code

Real builds, made with Claude Code and held to written checks.

  • ThisIsNoted, my meeting-notes app
  • A launch video made from code
  • Fact-checked AI tool reviews
  • 14 websites on automatic deploys
  • Cart-filling browser extension

Claude Code is my engineering partner. I set the rules and write the checks, and the work is not done until it passes them.

Some of the checks behind these builds: 1,327 tests on my own meeting-notes app, 600 checks on a used-book sourcing engine, 219 checks on an Instagram AI agent, and 208 quoted facts re‑found by a script before an AI tool review could ship. All 45 commits on a seller dashboard were written with Claude Code.

How I build with Claude Code

ThisIsNoted, my own meeting-notes app

Meeting notes from your calendar, tidied by AI. You type in your own shorthand during the meeting; it drafts the summary, decisions and to-dos, and every point shows where it came from. Not released yet: the Windows version is nearly ready.

ThisIsNoted's Today screen: meetings from the calendar with a Take notes button, and a card asking you to promise, once, that you will tell everyone in each meeting that it is transcribed.

Swipe sideways to see the whole screen.

Today: meetings from your calendar, and a one-time promise to tell everyone in each meeting that it is transcribed. Sample data.

Quick notes in, professional notes out.

The idea
AI notes you can check. Every AI point shows where it came from, and plain code (not the AI) flags names, numbers and dates that are not in your notes, the meeting invite or what was said on the call. Your typed text is never changed.
Meetings
From a calendar file, a private Google or Outlook calendar link, or a pasted invite. The next meeting in a series opens with its open to-dos, decisions and questions.
Live notes
Optional transcription with your own ElevenLabs or Soniox key, after a one-time promise to tell everyone in each meeting. ThisIsNoted keeps no audio.
Your AI
Your own Claude or OpenRouter key, a practice mode labelled “not real AI”, or off. Keys are encrypted with Windows’ own protection.
Checks
1,327 automated tests and a 187-step click-through of the real app, passing on 26 October 2026, and an evaluation kit of 11 made-up meetings with answer keys.
Built with
Claude Code, Electron, React and TypeScript.
Status
My own product, not released yet. The Windows version comes first.

A launch video, made from code

A 36-second launch video for a made-up app, built entirely in code with Remotion and Claude Code: no editing app, no stock footage, and music generated by code on the same beat as the picture (shown here without sound).

Remotion · React · Claude Code · wide and vertical

Fact-checked AI tool reviews

A Claude Code project that tests AI tools hands-on, and will not let a review page out until every vendor fact in it is re-found in a saved copy of the vendor’s page.

For
A client building a US review site about AI tools.
How
Every vendor page a review relies on is saved. A script re-finds each quoted fact in the saved page and records the page, the time and the result.
The gate
Since 17 October, a Claude Code hook checks each commit Claude makes against the parts of the written standard a script can test. If a finished page fails, or one of its quoted facts cannot be re-found, the commit is blocked.
So far
4 review pages and 208 facts, all re-found. 175 commits between 16 October and 1 November 2026.
Status
Client build, in delivery.
Read the case study
{
  "row": "Pricing: Free $0, 10,000 credits a month",
  "quote": "Free $0 (10,000 credits)",
  "file": "evidence/audit/V27us-elevenlabs-pricing.txt",
  "vendorHost": "elevenlabs.io",
  "check": {
    "status": "OK",
    "header": {
      "status": "200",
      "finalUrl": "https://elevenlabs.io/pricing",
      "utc": "2026-10-27T12:47:20.566Z",
      "title": "ElevenLabs Pricing for Creators & Businesses of All Sizes"
    }
  }
}
One of the 73 checked facts behind the ElevenLabs review, as the checker left it (empty fields removed).

14 websites on automatic deploys

For
A US membership business that teaches business owners to use AI.
What I did
Moved 13 sites from drag-and-drop uploads to GitHub with automatic deploys, and launched a 14th the same way. Logged test deploys went live in 7 to 14 seconds.
Checks
Every move followed a written checklist (live status, security headers, deploy source, hardcoded paths, assets, domain), with the results logged. All 14 sites are set up with 5 security headers.
Hand-over
A one-page guide, written so no developer background is needed, that teaches the team to change their own sites by asking Claude Code, in 4 steps.
Status
Delivered, May 2026.

Cart-filling browser extension

What it does
Loads a used-book seller’s morning picks into his shop baskets. After he pays, it reads his orders and logs each one into his inventory system.
Safety
It never buys, never sees a card and stores no password: it works inside his own signed-in browser.
Proof
Tested live: baskets filled on both shops, and 3 real orders from one shop logged, with 13 duplicates refused.
When unsure
An order it cannot match without doubt is refused, with the reason. It never guesses.
Status
Built and published on the Chrome Web Store.

Chapter Two3 builds

AI agents

Agents that read a question, find the answer and hand the work on.

  • IT troubleshooting engine
  • Multi-agent support system
  • Problem identification agent

Three parts of one IT troubleshooting engine, built for one client.

How I build AI agents
The DeductiveLabs intelligence dashboard: session totals, average confidence, knowledge base coverage and a problem category heatmap
web app

DeductiveLabs IT Troubleshooting Engine

A chat that walks IT support through a diagnosis one step at a time, behind a dashboard that tracks sessions, confidence, knowledge base coverage and skill gaps. Summary emails go out automatically.

React · Gemini · n8n · Supabase

n8n canvas of the DeductiveLabs agent flow: a webhook and a chat trigger, a product manager agent, an incident summary agent, and a switch that routes to four specialist agents
n8n canvas

Multi-Agent Orchestration System

Part of the DeductiveLabs engine. A product manager agent reads the issue, a summary agent condenses it, and a switch hands it to one of four specialist agents, each with its own OpenAI model.

n8n · OpenAI

n8n canvas of the Problem Identification Agent: a webhook from the app, branches that read images and files, a knowledge base search, an optional web search, and a reply to the app
n8n canvas

Problem Identification Agent

Part of the DeductiveLabs engine. It takes a question and any files from the app, reads images, PDFs and text files, searches the knowledge base, runs a web search only when needed, and sends one answer back to the app.

n8n · Gemini

Chapter Three2 builds

Instagram and ElevenLabs

Voice & chat agents

An Instagram agent built to book consults in the DMs, and voice work measured to the second.

  • Instagram AI chat agent
  • ElevenLabs voice project

Instagram AI chat agent

Built for a client to answer Instagram DMs for med spas and book consults inside the thread. The engagement ended in July 2026. These are its real test conversations.

  1. 01

    Someone asks to book. It asks which service, then when, and offers two open times from the spa’s calendar.

    An Instagram DM thread: the person asks to book a consult, the agent asks which service (Botox) and when (weekday morning), offers two open times on Monday, and asks how to send the intake form.
  2. 02

    Booked, inside the DM. It takes an email for the intake form and confirms. A calendar event, the GoHighLevel pipeline and a team alert follow.

    The end of a second test booking: the person picks email, gives an address, and the agent replies that the Botox appointment is confirmed and the intake form is on its way. The spa's location is covered.
  3. 03

    It remembers people, and knows when to hand over. A returning client is welcomed back by name. Someone unsure about a date gets a callback, flagged for the team.

    The agent greets a returning client by name, mentions their last Botox visit and asks what they would like to book.The person asks for a callback instead of picking a date; the agent asks for a time of day and a way to reach them, then says the team has been flagged.
  4. 04

    Behind it: 7 n8n workflows. This one is the brain, shown in two rows. It reads the message, checks the knowledge base, runs the AI agent, decides the next step and replies in under 2 seconds.

    The n8n workflow behind the agent, in full color: a webhook, the returning-contact lookup, the knowledge base cache, the AI agent with its OpenAI model and memory, the next-step decision, the calendar lookup and the reply.

Screenshots from testing in the client’s sandbox, so no real customer is shown. Cropped to remove account names and photos; the grey bar covers the spa’s town. The test person is me.

Built so every text DM becomes a GoHighLevel contact within 60 seconds.

For
An agency that sets up Instagram booking agents for med spas.
What it does
Answers DMs from each spa’s own knowledge base (a Google Doc their team edits) and books consults in five steps: service, preferred time, two open slots, how to send the intake form, contact details.
Follow-ups
If someone goes quiet, it follows up after 1 hour and after 20 hours, then stops at Meta’s 24-hour limit.
People
Medical, refund, legal or “talk to a person” messages are never answered by the AI: it takes a callback time and contact, then alerts the team.
Built with
ManyChat for Instagram, 7 n8n workflows, an OpenAI model (gpt-4o-mini), GoHighLevel, Google Calendar and Gmail.
Checks
219 automated checks across 7 test suites, all passing on 27 April 2026. The rule: no workflow change went back into n8n until they passed.
Status
Client build, delivered. The engagement ended in July 2026.
Read the case study

ElevenLabs voice project

A client project, October 2026: a hands-on review of ElevenLabs for an AI tools review site, with every take measured.

One click turned a 189-word script into 71.3 seconds of speech, ready in about 18 seconds.

Listen: the first 34 seconds of the narration

0:34

AI voice (ElevenLabs, Eleven v4): the same script in another library voice, made on the paid Starter plan. The company in the script is made up.

One click gave two versions. The second cost 0 credits.
Dubbing v1 into Filipino, quoted at 3,566 credits for the 71-second file (both ends of the same bar).
Accuracy
The take was read back by speech-to-text and compared with the script. The only differences were number formats (“6” as “six”) and the made-up name Kestrelwick, heard as “Kestrel Wick”.
Long takes
A long script of 1,521 words became 8 min 40 s of audio in about 128 seconds. Its last 2 minutes (342 words) read back with no wrong words; the only differences were number formats and spacing.
Pauses
Pause tags drift. Pauses ran 1.0 to 1.2 seconds up to 4:50, then 1.6 to 1.9 seconds after 5:13, while the average speaking speed held at 186 to 187 words a minute.
Dubbing
For the same 71-second file, dubbing v2 quoted 16,046 credits and v1 quoted 3,566. Filipino and Spanish dubs took about 2 minutes each on v1.
Billing
Speech to text was charged exactly ElevenLabs’ stated 330 credits a minute (660 credits for a 2-minute file, 392 for a 71-second one).
Tools
ElevenLabs text to speech (Eleven v4), speech to text, dubbing, voice cloning and the API (read only).

Chapter Four6 builds

Outreach & marketing

Cold email that stops itself before it does harm, and engines that write the content.

  • Cold-email safety checks
  • Sourced lead list
  • Lead scraper and icebreakers
  • Hook bank and UGC briefs
  • Instagram hook engine
  • Static ad generator

Cold-email safety checks

Cold email can hurt a business fast: emails that bounce, a mailbox in poor health, a reply nobody reads. So the campaigns run behind five checks.

  1. 1

    The campaign can be read.

    Without this, a count of 0 could mean nothing happened, or that the check could not look.

  2. 2

    Every sending mailbox is active, warmed up and healthy.

    A mailbox that is missing, inactive, has warm-up off or scores under 90 (Instantly’s own bar) stops it.

  3. 3

    Bounces stay at 3% or under, per person emailed.

    Only judged once 50 or more people have been emailed, because 1 bounce in 20 says nothing. Above 2% is a warning.

  4. 4

    The list of sent emails can be read.

    Otherwise the bounce rate was never really checked.

  5. 5

    No AI reply agent is switched on.

    It could answer a prospect in the client’s name before a person has read the reply.

  6. !

    If any one fails, the send stops. The check runs at every check-in.

For
An Australian plumbing business: 4 campaigns, about 940 contacts.
Runs on
Node.js and the Instantly API. The daily check only reads; it never sends anything itself.
Bounces
Counted per person emailed, not with Instantly’s own “contacted” figure, which matched neither the people nor the emails.
It warns
When anyone replies (a person reads every reply) or unsubscribes, and when a send setting differs from the plan. It also prints what it cannot see: complaints inside replies, and replies that go to another inbox.
Also
A preflight of 50+ checks before a campaign launches. Every opt-out copied to a record that outlives Instantly, for the Spam Act. Mailboxes stepped from 10 to 20 to 30 emails a day, with a rule to step back.
Read the case study
Instantly’s analytics said 0 unsubscribes. 3 people had used the unsubscribe link.
Found on 31 October 2026, from the lead records. An earlier weekly report had repeated the 0. The check now counts unsubscribes from the lead records and labels that analytics line as missing them.

A lead list where every emailed address has a source

The list behind the property-manager campaigns: every address was found on a real page, and that page was checked for a no-marketing notice before the address was loaded.

For
The same plumbing business: commercial property and strata agencies to email.
Sources
Collected from public agency websites and a property-listing site. Every address that gets emailed keeps the page it was read from. 97% of the list was read off a page; the rest, guessed from a name pattern, are never emailed.
The gate
Before an address could be emailed, a real browser re-read its source page for opt-out notices: 1,044 pages in 2 hours 32 minutes.
Customers
Checked against the client’s own job records, so his customers came off, except 2 firms he asked to keep in.
Built with
Claude Code and Node.js, with a Spam Act 2003 note that cites every rule by section and clause.
Status
Delivered in September 2026, and in use.

Engines that find leads and write the content.

Tap a row to see the real build.

  • Part of the Apollo lead scraper’s n8n canvas: the groups that fill the sheet and write the icebreakers

    A form builds an Apollo search, and an Apify actor pulls the matching leads into a Google Sheet. For each new lead, the flow reads their website and OpenAI drafts a personal icebreaker into the same sheet.

Chapter Five3 builds

Amazon

E-commerce

Amazon systems that tell a seller what to fix first.

  • Seller dashboard
  • SP-API sync engine
  • Book sourcing engine

Seller Central knows everything. It doesn’t tell you what to fix first.

A copy of the seller dashboard I built, shown with sample data. It opens on today’s problems, worst first: out of stock and still selling, needs shipping to Amazon, and stock Amazon is not selling.

What needs doing today, worst first.

  1. Out of stock, still selling. Products with orders and no stock. Always at the top, because they lose sales every day.
  2. Needs shipping to Amazon. What to send and how many days are left to cover the season ahead.
  3. Stock Amazon is not selling. Units Amazon holds but won’t sell, flagged so someone checks Seller Central.

A seller dashboard I built for a US outdoor brand on Amazon. Amazon SP-API, GitHub Actions, Supabase and React. Client build, in delivery. The first pre-season purchase order has already gone out through it. Shown here with sample data.

Read the case study

Two more Amazon systems I’ve built.

Amazon SP-APIorders · returns · financial eventssettlements · inventoryFBA shipments · feesn8n, one workspaceabout 45 workflowsdaily · hourly · gap-fillshared Reports API looprequestpolldownloadparseuntil readyClient inventory systemUS · Canadaemail alerton failure
How the sync engine runs, simplified.

Amazon SP-API Sync Engine

For
An Amazon FBA business selling in the US and Canada.
What it does
Brings orders, returns, financial events, settlements, inventory, FBA shipments and fees into their inventory system, daily and hourly. If a run is missed, it catches up.
Runs on
About 45 n8n workflows in one workspace.
Status
Client build, live. Walmart built and staged on the same pattern.

Every sync emails the moment it fails, so a quiet morning means it ran.

Read the case study
AbeBooksBiblioAlibrislistings scannedabout 2,000each morningbuying rulesabout 700 passpriced under its six-month averagelive check, each survivorprice · condition · seller locationreal browser for bot-protected shopsmargin per bookafter postage and Amazon feesdaily buy list6 to 8 booksinside the budget
How one morning’s run narrows, simplified.

Used-Book FBA Sourcing Engine

For
A used-book Amazon FBA seller.
What it does
Scans about 2,000 listings on AbeBooks, Biblio and Alibris each morning, opens every survivor live for price, condition and seller location, counts postage and Amazon fees, and writes a buy list inside the daily budget.
Runs on
Node.js, and a real browser for the shops behind bot protection.
Status
Client build, in delivery. 600 automated checks guard every change, and a second check re-reads Amazon and the shops before a list goes out.
Read the case study

In my clients’ words.

5.0

Every review here is 5 stars, verified on Upwork

Read every review on Upwork (opens in a new tab)

n8n Fix

5.0Aug to Oct 2026

Basti has been incredible to work with. He goes above and beyond the task given and brings solutions proactively. He is also an excellent communicator and keeps you updated…

Basti is short for Sebastian, my middle name.

Endorsed by client

  • n8n
  • Google Sheets Automation
  • Automation
  • Solution Oriented
  • Clear Communicator
  • Accountable for Outcomes

  • AI Automation, Image Generation

    5.0Jan 2026

    Eleazar brought the vision into reality. I wasn’t sure how the automation is going to work but he showed us that everything was possible. He works really quick and very knowledgeable

    Endorsed by client

    • Reliable
    • Committed to Quality

  • n8n Multi-Agent Framework Design

    5.0Dec 2025 to Feb 2026

    He has deep knowledge in workflow automation. His technical background helped the project to be more than I envisioned.

    Endorsed by client

    • Committed to Quality
    • Solution Oriented
    • Detail Oriented

  • Workflow Automation (Asana, Zapier, HubSpot)

    5.0Oct to Nov 2025

    Great experience working with this developer. They were responsive, detail-oriented, and delivered high-quality work on time. Excellent communication throughout the project. Highly recommended.

    Endorsed by client

    • Reliable
    • Collaborative
    • Clear Communicator
    • Accountable for Outcomes

Where I’ve worked.

Windows and Azure systems in enterprise IT from 2012 to 2026. Now building AI automation full time, from Calamba, Laguna, Philippines.

  1. Aug 2026 to presentBook-Sourcing EngineFreelance · Used-book Amazon FBA seller (Upwork)Details
    • Building a sourcing engine that scans about 2,000 used-book listings each morning across AbeBooks, Biblio and Alibris, and ends with a daily buy list that fits the client’s budget.
    • Every surviving listing is opened live for its current price, condition and seller location. Postage and real Amazon fees are inside the margin before a book can be recommended.
    • 600 automated checks (as of November 2026) pin behavior before any change ships; a deliberate gate keeps the scraping engine from being edited casually.
    • A second check goes back to Amazon and the shops and reproduces every figure from the source, so no list reaches the client unverified.
    • Phased, milestone-based engagement. Sourcing failure modes the client caught by hand are now guarded by automated checks.

    Skills: Node.js, Web Scraping, Chrome DevTools Protocol, Amazon FBA

  2. Apr to Jul 2026AI DM Agent (Instagram to GoHighLevel)Freelance · Agency client (Instagram bookings for med spas)Details
    • Built the full agent to read the intent of each Instagram DM, answer from the client’s knowledge base and book consults inside the DM thread, with an OpenAI model (gpt-4o-mini).
    • Built the first version on Instagram’s own webhooks (the verification handshake and inbound message events), with replies sent through ManyChat.
    • Set it up so every DM created or updated one GoHighLevel contact (no duplicates), and a confirmed booking added a calendar event, moved the pipeline opportunity to Booked, sent the intake form and alerted the team.
    • Medical, refund, legal or “talk to a person” messages were never answered by the AI: they went to a person, through a callback request and a team alert.

    Skills: n8n, OpenAI, ManyChat, Instagram Graph API, GoHighLevel, AI Agents

  3. Nov 2025 to presentAutomation Specialist, Amazon SP-API and SourcingFreelanceDetails
    • Built automations against Amazon’s Selling Partner API: LWA authentication, request throttling, pagination and the asynchronous Reports API.
    • Live sales and inventory dashboards with re-order alerts; inbound-shipment and order-status workflows synced to the inventory system.
    • Two-stage LLM product-sourcing bot: strict brand / UPC / MPN / SKU matching plus SOP rules, with Apify supplier scraping and Keepa pricing. Buy / no-buy verdicts are delivered to Slack with sourced reasoning.
    • Added logging and monitoring so failures are easy to track and debug in production.

    Skills: n8n, Amazon SP-API, Python, OpenAI, Apify, Keepa

  4. Sep to Nov 2025Creative IntelligenceSanlava (e-commerce and social brand)Details
    • Designed a “creative intelligence” workflow that scrapes ads, UGC and reviews from TikTok, Instagram, Facebook and Google.
    • Structured data into Notion databases (Organic Posts, Ads, Reviews, Trends, Hook Bank), a single source of truth for creative ideas.
    • Used OpenAI and Gemini to generate hooks, UGC briefs and static ad prompts based on scraped data and custom scoring rules.
    • Automated weekly runs, scoring and QA so the marketing team gets fresh, ready-to-use creative insights without manual research.

    Skills: n8n, Apify, Notion, OpenAI, Gemini, UGC

  5. 2012 to Aug 2026System Administrator, Windows and Azure CloudEnterprise InfrastructureDetails
    • Managed large fleets of Windows Server environments with strong focus on uptime, security updates and compliance.
    • Automated patching and reporting via PowerShell and Azure Update Manager, keeping systems consistently up to date at scale.
    • Troubleshot complex L3 issues across VMware, networking, storage and backups in production environments.
    • Worked with global teams to plan changes, communicate risks and document runbooks for stable, repeatable operations.

    Skills: Windows Server, Azure, PowerShell, VMware, Patch Management

Tools I use. Automation: n8n, Make, Zapier. AI and agents: Claude, Claude Code, OpenAI, Gemini. Amazon: SP-API, Keepa API. Full stack: Supabase, React, Node.js, Python, Netlify. Data and CRM: Airtable, Notion, HubSpot, GoHighLevel. Scraping and APIs: Apify, custom APIs.

Fixed price, per milestone.

You get a written scope and a fixed price for each milestone. You sign off before I build anything, so there are no surprise invoices.

  • AI Agent and Workflow Build

    $600 to $1,800 per build

    Covers
    • AI customer agents that know your knowledge base
    • n8n and Make workflows
    • Internal AI assistants that answer with citations
    • Data pipelines
    • API and CRM integrations

  • Amazon System Build

    $900 to $2,400 per build

    Covers
    • SP-API sync workflows for orders, inventory, financials and returns
    • Live seller dashboards with re-order alerts
    • Sourcing and buy-decision tools
    • Runs on your accounts, credentials stay yours
    • Handoff doc and recorded walkthrough

  • Maintenance and Growth

    $200 a month

    Covers
    • Up to 5 hours of updates a month
    • Monitoring and small fixes
    • Priority support
    • A monthly summary

Each build is split into fixed-price milestones. Smaller job? Hourly consults and fixes from $15 an hour.

How a build goes.

  1. Day 1

    Call and audit

    A 30-minute call. I look at your current setup, ask where the hours go, and point out what is worth automating first.

  2. Days 2 to 3

    Written scope

    Deliverables, timeline and a fixed price per milestone, in writing. You sign off before I start.

  3. 1 to 3 weeks

    Build

    I work in your tools, so the handoff is clean. You see working pieces every few days, with a weekly check-in.

  4. Final week

    Handoff

    It runs on your accounts, with full documentation and a recorded walkthrough. Yours to keep.

Questions.

Most builds take 1 to 3 weeks from signed scope to handoff. Full audits and small fixes take 3 to 5 days. AI agent and knowledge base projects take 2 to 4 weeks, depending on their size.

Yes. I’m happy to sign a mutual NDA before we start, and I already work with several clients under NDA.

Yes. I work in your stack: Slack, Notion, HubSpot, Airtable, Sheets or whatever you already use. I’m not tied to one platform, so I pick the right tool for the job.

Your hosting (AWS or GCP), domain costs and third-party API spend such as OpenAI, Apify or Pinecone. They are listed in the scope, and I don’t mark them up.

Yes. The $200 a month Maintenance and Growth plan covers up to 5 hours of updates, monitoring and small fixes. Bigger ongoing work is priced per scope.

Yes. Every build comes with documentation, and I can run paid training sessions so your team can extend and fix the workflows without me.

Not sure which fits? We’ll work it out on the call.

Book a call (opens in a new tab)

Tell me what you check every morning.

In 30 minutes, I’ll tell you plainly if I can automate it and what it would take. If it’s outside my lane, I’ll say so.

Book a call (opens in a new tab)

I’m taking new projects, and I usually reply within 24 hours.

Fixed price per milestone. Runs on your accounts. Code, credentials and docs stay yours.

© 2026 Eleazar Sebastian Martinez. Built in React, with entirely too much n8n.