BEG

AI portfolio · Mexico City

Mexican corporate lawyer who builds with AI.

Energy, hydrocarbons and mining · about 8 years of practice
Spanish (native), English (professional)

  • Claude Code plugins and agents
  • retrieval over legal text
  • n8n workflows
  • Python data pipelines
  • websites
  • video
  • AI agents for marketing

00 · In one sentence

I write the legal analysis, and I build the systems around it and beyond it. I use AI every day. This portfolio is a sample, not a complete list: some projects are confidential and others haven't been released yet.

Benjamín Esponda García seen from behind at his desk, wearing headphones, with an n8n workflow on one monitor and code on the other

This is me

This is where all of this gets built.

My desk: an n8n workflow on one screen and code on the other.

Learning and community

At the forefront: new tools get tested on real work.

I use AI every day and I stay at the forefront of new AI tools and features: I test them on real work as soon as they come out.

Benjamín Esponda García in a suit and tie, smiling at his laptop, with an abstract painting behind him
  • Anthropic's free online courses (Anthropic Academy).
  • Harvey's free online courses on AI for legal work.
  • Live and pre-recorded courses from the community of tech influencers such as Juan Lombana and Enrique Rocha, and from digital schools such as Coderhouse and OrangutanAI.
  • Claude talks and workshops in Mexico, such as Claude for AI Engineering (Claude Community CDMX, October 2026), Claude for Legal, Claude for Finance, Claude Community House, Claude for Revenue, Claude for Games, Claude at the Chamber of Deputies and Claude at the Mexican Senate, among others. At Claude for AI Engineering I summarized the five talks (tool use, MCP, context engineering) and turned the notes into an agent that answers as each speaker would.
  • Grok Bot talks and workshops.
  • Networking dinners with CTOs of large and mid-size companies.
  • n8n automation course (in progress): the daily Official Gazette digest was the first deliverable.

Academic background

  1. 2023 to 2025

    Master's in Business Law (Maestría en Derecho de la Empresa)

    Universidad Anáhuac

    Field: corporate law

  2. 2023 to 2024

    Master of Business Administration (MBA)

    EAE Business School (Madrid, Spain)

    Field: business administration

  3. 2015 to 2020

    Law degree (Licenciatura en Derecho)

    Universidad Iberoamericana (Mexico City)

    Field: law

My personal brand and consultancy · pivotek.mx

Pivot

Strategic pivot: a fundamental, deliberate change in a business's strategy (its product, target market or revenue model) when it becomes clear the current direction isn't working, in order to survive and grow.

Tek

From tech: technology.

I named it this way because I know that in a few years the practice of law won't be what it is today, or what it was before AI arrived. You have to stay at the forefront of technology and good practices.

01 · At a glance

Six numbers, one judgment.

  • 25+

    More than 25 projects in 2026, from cross-border regulatory audits to a legal-tech product that is on sale.

  • 374

    energy-law instruments indexed for natural-language search.

  • 11,112

    chunks, split along the hierarchy of Mexican statutes.

  • 560,000+

    public procurement contracts in a pipeline that shows oil and gas suppliers where the money goes.

  • 10,000+

    browser actions automated from Claude Code.

  • 1

    product live: Con Criterio, on sale at concriterio.mx since November 5, 2026.

Automation with judgment: every workflow keeps a person in the loop where signatures, money or legal judgment are involved.

  • Stripe
  • Gmail
  • Google Calendar
  • Google Drive
  • Vercel
  • Higgsfield
  • Apify
  • Canva
  • browser automation
  • all driven from Claude Code

02 · AI skills map

Twelve skills, each with something built.

  • 01

    AI agents and agentic workflows

    Claude Code plugins, skills and subagents for seven areas of Mexican law; a nine-role multi-agent system with typed handoffs; social media run by Grok Bot and Meta's Muse.

  • 02

    Retrieval over documents (RAG)

    Legal RAG with law-specific embeddings, pgvector and hybrid ranking over 374 energy-law instruments.

  • 03

    Data engineering

    A pipeline over more than 560,000 public procurement contracts with weekly updates and audits; financial models in Python and Excel.

  • 04

    Workflow automation

    n8n designs for a 20-step real-estate closing and a daily digest of the Official Gazette (DOF); scheduled tasks for the monthly regulatory review and the weekly second-brain review.

  • 05

    AI for legal work

    Cross-border regulatory audits, contract and compliance reviews, opinions and letters, every data point with source and cut-off date.

  • 06

    Product building

    Con Criterio, from content to payments, automated delivery and QA with simulated buyers.

  • 07

    Web development

    concriterio.mx, pivotek.mx, casalegado.mx and carocomunicando.com.

  • 08

    AI video and content

    A stage-entrance video with AI voice-over and motion graphics, and an automated Reels pipeline with local transcription and computer vision.

  • 09

    AI marketing

    More than a dozen marketing and creative skills applied to real campaigns, SEO audits and ad copy.

  • 10

    Integrations (MCP connectors)

    Stripe, Gmail, Google Calendar and Drive, Vercel, Higgsfield, Apify, Canva and more than 10,000 browser actions.

  • 11

    Judgment and quality

    Human approval gates, anti-fraud rules, labeled uncertainty, golden test sets and reproducible builds.

  • 12

    Continuous learning and community

    Daily use of AI and of new tools as soon as they launch; free courses from Anthropic and Harvey; live and pre-recorded courses from tech influencers and digital schools; Claude and Grok Bot talks and workshops; networking dinners with CTOs.

03 · Featured · Live product

Con Criterio: an AI toolkit for Mexican lawyers.

A guide (116 pages in the Claude edition, 134 in the ChatGPT edition), 7 plugins by practice area and 7 drafting skills. On sale at concriterio.mx since November 5, 2026. Since November 11 it also offers a beginners' guide to everyday AI for the general public.

7 plugins by practice area

Each plugin is a set of Claude Code skills grounded on a curated corpus of statutes and verified court criteria, with one citation format, data labels and a validation block that runs before any answer is delivered.

7 drafting skills

  • residential lease (Mexico City)
  • mutual NDA
  • commercial agency
  • professional services
  • letter of intent
  • simple power of attorney
  • severance calculator
  1. 1/7

    LFPDPPP

    Data protection

    Privacy notices, processor agreements, transfers, infringement risk, defense before the authority.

  2. 2/7

    LGSM

    Corporate

    Shareholder minutes, bylaws audits, powers of attorney, corporate due diligence, step-by-step restructurings.

  3. 3/7

    LFPPI

    Trademarks

    Registrability, oppositions, nullity and lapse, infringement, deadlines before the trademark office.

  4. 4/7

    LFDA

    Copyright

    Ownership opinions, transfer clauses, exceptions, notice and takedown, works made with AI.

  5. 5/7

    Commercial Code and LGTOC

    Commercial contracts

    Drafting and review, promissory notes, credit operations and trusts, interest and usury, collection routes.

  6. 6/7

    Mexico's 2025 energy laws

    Energy

    Permit roadmaps, power and hydrocarbons projects, sanctions, a 2014 versus 2025 comparison.

  7. 7/7

    Notarial and registry law

    Notarial practice (Mexico City)

    Deed requirements, wills and advance directives, probate before a notary.

How it was built, in numbers

  • 115

    verified claims and 46 screenshots in the Claude edition guide. The ChatGPT edition was ported with install steps for Windows.

  • 101

    automated tests passing on the Cloudflare Worker that delivers the files and runs the newsletter. Sales run on its own page, Stripe payment links and a Stripe webhook.

  • 118

    distinct findings and zero blockers in a simulated round with 20 buyers; 114 of the 115 open findings were closed the same day.

  • Day 1

    of every month, a scheduled regulatory review runs. A Python monitor reads the Official Gazette (DOF) through the government's open-data web service and flags anything that touches the statutes the product covers.

  • Its own legal layer: terms and conditions, privacy notice and license of use.
  • Marketing: content produced with Claude Code and published by AI agents, SEO audit, two Google Ads campaigns prepared and the Meta Pixel connected to the payment flow.

04 · Featured · RAG

A retrieval engine over Mexican energy law.

Problem. Mexico's energy rules are scattered across the Official Gazette (DOF) and the websites of eight energy authorities, and the 2025 reform replaced much of the framework. Finding the instrument that actually applies, and knowing whether it still applies, takes hours.

What I built. A pipeline that collects, structures and indexes Mexican energy legislation so it can be queried in natural language. No LangChain or LlamaIndex: plain Python and SQL.

Scale (April 2026)

374 documents

11,112 chunks, all embedded

352 in force 22 repealed

  • CRE
  • SENER
  • CENACE
  • CNH
  • CENAGAS
  • CNE
  • ASEA
  • Congress
  1. Stage 1

    Collection

    Async scrapers over the Official Gazette and regulators' sites, with polite delays.

  2. Stage 2

    Extraction

    Text from PDFs with an OCR fallback for scanned documents.

  3. Stage 3

    Chunking

    Splits text along the hierarchy of Mexican legal instruments (title, chapter, article) instead of fixed sizes.

  4. Stage 4

    Storage

    PostgreSQL with pgvector: documents and chunks, validity flags and a validity note for rules that are formally in force but suspended in practice.

  5. Stage 5

    Embeddings

    Voyage AI voyage-law-2, built for legal text.

  6. Stage 6

    Search

    Hybrid ranking (vector similarity plus keyword, merged with reciprocal rank fusion), synonyms, statute detection and boosts.

  7. Stage 7

    Answers

    Claude answers natural-language questions over the retrieved chunks.

  8. Stage 8

    Quality

    A "golden" test set used as a regression guardrail (9 of 9 passing) and database audits for duplicates, validity and publication dates.

05 · Featured · Data at scale

SupplySense: where the money goes in oil and gas.

Problem. Suppliers to the oil and gas sector in Mexico bid blind: they don't know who buys what they sell, at what price, or when. What I built: a price-intelligence tool built on public procurement data.

More than 560,000 contracts from CompraNet and 150,000 from ComprasMX.

More than 560,000 contracts from CompraNet (2023 to 2025) across 26 categories

150,000 recent contracts from ComprasMX collected with Playwright

Each dot: 1,000 contracts

  • 01

    Load

    More than 560,000 contracts from CompraNet (2023 to 2025) across 26 categories, plus 150,000 recent contracts from ComprasMX.

  • 02

    Updates

    Weekly automatic download with deduplication; data enriched with geography and level of government; per-subsidiary links for the national oil company.

  • 03

    Dashboard

    KPIs, watchlists, line-of-business filters, global search, and CSV and PDF export; basic multi-tenancy.

  • 04

    Audits

    Caught and fixed misclassified categories and outliers, plus a market sizing (TAM) analysis.

06 · Featured · n8n

A 20-step real-estate closing, automated with a hand on the brake.

Problem. A real-estate law firm in Baja California closes property purchases by foreign buyers through Mexican bank trusts. Each closing has 20 steps and depends on a notary, a trust bank, an appraiser, an escrow agent and the public registry. Most of the lawyers' time goes to chasing documents and answering "what's the status of my closing?".

  • 11can run almost on their own
  • 7are assisted
  • 2stay human: the legal review of the deed draft and the signing before the notary

One block per step, in closing order. Hover to see which step it is.

Trigger

Escrow opens

  • Foreign-affairs permit
  • Deed draft
  • Appraisal
  • Certificates

Steps 4 to 7 start in parallel the moment escrow opens, which could cut 7 to 10 days per closing [ESTIMATE].

The 9 components of the design

  • 01

    Single deal record

    One row per deal with buyers, unit, bank, notary, escrow, dates per step, validity periods and status.

    Google Sheets first; own database if volume grows

  • 02

    Folder per deal

    Fixed structure created automatically: 01 Client to 08 Post-closing.

    Google Drive

  • 03

    Outgoing email

    Bilingual templates for client, notary, bank, appraiser, HOA and escrow, filled from the record.

    Gmail

  • 04

    Incoming email reader

    Classifies third-party emails, links them to the deal, files attachments and moves the status.

    Gmail and AI

  • 05

    WhatsApp

    Notices, reminders and a bot that answers "what's the status of my closing?" from the record.

    WhatsApp Business API

  • 06

    Document AI

    Reads passports and certificates, compares the deed draft, summarizes video calls and bank comments.

    Claude API, with human approval

  • 07

    Timers and gates

    Deadlines per step, certificate validity and the signing gate.

    n8n

  • 08

    Partner dashboard

    Deals by stage, days per step, who is blocking, alerts.

    Google Sheets or Looker Studio

  • 09

    Logs and errors

    Record of every run and an instant alert if a flow fails.

    n8n

Signing gate

No one books the closing until six conditions are met.

  1. Foreign-affairs permit granted
  2. Appraisal received
  3. Certificates still valid on the proposed date, not just today
  4. Deed draft approved by the trust bank
  5. Funds confirmed in escrow
  6. Final balance paid

6 of 6: the closing can now be booked

Anti-fraud · red alert

No workflow ever sends bank details, and any email about new wiring instructions raises a red alert and requires a verification call.

  • Document AI with human approval. It reads passports and certificates and compares the deed draft against the record.
  • Client data never lives on my infrastructure. The demo runs on fictitious data and production goes on the firm's own server.

Status · design stage (November 2026)

Proposal delivered; decisions kept in a decision register; rollout planned in 4 phases over 10 to 13 weeks [ESTIMATE], starting with a pilot on one development. Expected savings: 50% to 70% of busywork hours per closing [ESTIMATE], to be measured in the diagnosis.

07 · n8n · Daily Official Gazette digest

The energy news in the Official Gazette (DOF), in my inbox at 8:30.

Problem. Every morning I reviewed the Official Gazette by hand for anything affecting my energy clients: about 30 minutes a day [ESTIMATE].

  1. Schedule

    Mon to Fri, 8:30

  2. HTTP Request

    today's index

  3. HTML

    title, agency, link

  4. Filter

    energy keywords and agencies

  5. Aggregate

    all in one

  6. LLM

    one line + priority

  7. Gmail

    daily email

About 100 hours a year saved [ESTIMATE]

Result. One email a day with the energy publications sorted by priority. The workflow tells me what to read; reading it is still my job.

Built as the first deliverable of an n8n automation course, with a documented prompt and an error-handling plan.

09 · Websites, video and digital marketing with AI

Four sites live, video and agents that publish.

  • concriterio.mx

    Sales site of Con Criterio: a page for lawyers and /para-todos for beginners, with a build-your-purchase panel.

    Pages and legal texts generated with Python scripts; Stripe payment links and a Cloudflare Worker for checkout and delivery; Cloudflare Pages, DNS and Web Analytics; Meta Pixel and two Google Ads search campaigns set up.

  • pivotek.mx

    Home of this portfolio. It replaced the previous Next.js site of PIVOTEK, my personal brand and consultancy, which I also built (TypeScript, contact form with Resend, technical SEO).

    Designed and built by Fable 5.2 (Claude) from the plain text of my README, under my direction: 20 sections with their own scroll animations, self-hosted GSAP, ScrollTrigger and Lenis under a strict content security policy; deployed on Vercel.

  • casalegado.mx

    My biographical documentary brand: five-page site plus brandbook, sales dossier and sales presentation.

    Static HTML and CSS under a strict brand system (two typefaces, a warm palette and voice rules), illustrative images generated with Higgsfield, privacy notice, sitemap and a content security policy; deployed on Vercel.

  • carocomunicando.com

    Site of a public-speaking and communication trainer: home with an animated cover, her method, her TEDx talk, cases and brand pages.

    Static site with self-hosted GSAP, ScrollTrigger and Lenis; a downloadable PDF guide with optional email capture; GA4, Microsoft Clarity and Vercel Analytics declared in the privacy notice and cookie banner; SEO audit and blog articles.

Video for a speaker

  • 102s

    Stage-entrance intro

    Projected before she walks on stage. Five versions; the current one combines motion graphics built in Remotion, an AI voice-over (ElevenLabs v4 through Higgsfield), original music and sound effects, and her photos. It comes with a waiting loop for the screen and three short Instagram Reels. Every version can be rebuilt from scripts (Python, Node and ffmpeg).

  • 40s

    Event highlight

    Her role as host and moderator of a university entrepreneurship summit: 9:16 for Reels, TikTok and LinkedIn and 16:9 for YouTube, with original music timed to the moment the stage lights come on.

  • 10s

    Book launch loop

    An animated screen with quotes that rotate every 10 seconds and verified QR codes to her Instagram and website.

Automated Reels pipeline

Drop a raw video in Downloads and get a finished Reel.

  1. 1

    Transcription

    Word by word, locally, with Whisper.

  2. 2

    Style rules

    Learned from a reference creator: four Reels analyzed frame by frame (a hard cut to B-roll every 5 to 9 seconds, title pills, a closing question).

  3. 3

    Supporting visuals

    Licensed real photos from Wikimedia Commons and Openverse, with credits, or B-roll generated with Higgsfield.

  4. 4

    Subject cut-out and face tracking

    With Apple's Vision framework (Swift).

  5. 5

    Automatic edit

    In Python: format, subtitles, pills, cards, photo credits and the speaker's frame; a history file avoids repeating formats.

Six Reels ready to publish in November 2026, a monthly set of scripts, and a technical guide to publish through the Instagram Graph API with n8n.

Social media run by AI agents (Con Criterio)

  • Claude Code

    Produces

    The content and the publishing calendar, the prompts for the agents and the checklists to verify their work.

  • Grok Bot (an agent that operates my Mac)

    2 Reels a day

    On TikTok, plus Instagram prospecting and direct messages; if the Mac is asleep or asks for a code, it stops and warns.

  • Muse (Meta's AI marketing agent)

    3 posts a day

    On Instagram and Facebook; it runs the campaigns and continues the conversation when a prospect answers.

The labeling of AI-generated content was reviewed with Muse, and verification prompts audit what each agent actually did.

  • humanizalo-es
  • social-content
  • content-strategy
  • copywriting
  • marketing-psychology
  • pricing-strategy
  • ad-creative
  • campaign-plan
  • seo-audit
  • Higgsfield skills
  • frontend-design
  • emil-design-eng
  • apify-lead-gen
  • abogado-el-diablo

Marketing skills and where I applied them

  • humanizalo-esCopy for carocomunicando.com and its Reels, Con Criterio pages and posts, course deliverables.
  • social-contentReels scripts and calendar for the trainer; Con Criterio's social strategy.
  • content-strategyCon Criterio's content strategy for Instagram, Facebook and TikTok.
  • copywritingBusiness plans and the trainer's site.
  • marketing-psychologyThe trainer's offer and pages; business plans.
  • pricing-strategyPricing in business plans and the trainer's proposals.
  • ad-creativeAd variations for the trainer's campaigns.
  • campaign-planTraffic and contacts plan for the trainer (November 2026).
  • seo-auditAudits of carocomunicando.com and concriterio.mx.
  • Higgsfield skillsSpeaker intro, B-roll and book loop; Casa Legado images; Con Criterio explainer pieces.
  • frontend-design and emil-design-engCon Criterio's sales page.
  • apify-lead-genProspect base for a website service in Los Cabos.
  • abogado-el-diabloCon Criterio's launch strategy.

10 · Connectors (MCP)

Claude Code reaches my tools: I automate, query and generate.

Counts come from my own session logs.

  • Claude in Chrome and the built-in browser

    more than 10,000 actions

    Operate web dashboards, test my sites end to end, upload files and check what is live.

  • Higgsfield

    more than 1,000 calls

    Images, video, voice-over and music for sites, Reels and the speaker intro.

  • Computer use

    more than 900 actions

    Drive native Mac apps when there is no API.

  • Stripe

    about 800 calls

    Products, prices, payment links and webhooks for Con Criterio.

Automate

  • Gmail, Google Calendar and Google DriveWeekly review of calendar and inbox, drafts and follow-ups, file search.
  • Scheduled tasksWeekly second-brain review, monthly regulatory review, campaign follow-ups.
  • VercelProjects, domains and web analytics of my sites.

Query

  • ApifyWeb scraping, lead lists and web browsing for research.
  • Wispr FlowSearch across my meeting notes.

Generate

  • CanvaDesigns from brand kits.
  • Claude artifactsShareable pages.

Plus two external agents (Grok Bot and Muse) and local command-line tools: Obsidian, GitHub, Whisper and ffmpeg.

11 · Desktop companion

BeCa: my cat knows what Claude is doing.

BeCa is my cat turned into a sticker-style companion that lives in a small floating window on my Mac. It listens to Claude Code's lifecycle hooks, so I can tell at a glance what Claude is doing without looking at the terminal. Why BeCa? To bring a sense of humor to my day-to-day work with AI.

clickClick it once and it waves; click again and it plays with Clawd for about four seconds.

It runs locally, uses no network and stores no content. The hooks are global, so BeCa works in every project.

  1. Claude Code starts a session

    SessionStart

    Waves twice, then rests

  2. Works on a request or runs a tool

    UserPromptSubmit PreToolUse PostToolUse

    Types in the air

  3. Edits or writes a file

    PostToolUse of Edit or Write

    Inspects the work, paw on chin

  4. Finishes its turn

    Stop

    Jumps four times and waves

  5. Waits for my permission or answer

    PermissionRequest Notification

    Asks with its paws together

  6. Hits an API error

    StopFailure

    Deflates once

  7. Closes the session

    SessionEnd

    Rests

How it was built: a native macOS app in Swift, made with Claude Code; the art was generated with Higgsfield from three photos of my cat and cut into animation strips with Python scripts; the sounds were extracted from AI-generated video clips.

12 · Tools and systems

Systems that work while I work.

Second brain on Obsidian and Claude Code

One note per project, person and official document; every working session leaves a log entry; a scheduled task reads my calendar and email every Sunday (read-only) and updates the vault; hourly backup with launchd. This portfolio was assembled from it.

Personal multi-agent system

Nine specialist roles read one shared record, pass typed signals instead of prose and follow a declared precedence rule when they disagree. Personal use; no data shown.

My own internal CRMs

Built on the architecture of auto-crm, an open-source local CRM with Next.js, SQLite and Claude Code that adapts to each business.

one shared record
9 roles · typed signals · precedence rule

13 · Feasibility studies

Every study, with a verdict and traffic lights by phase.

Own ventures and services of PIVOTEK, my consultancy.

Thanks to artificial intelligence tools and our own custom skills, we have produced many business plans, all with conclusions I have validated personally.

14 · How I work with AI

Seven rules, none of them negotiable.

  1. Sources first.

    Official texts are downloaded into the project, and each data point carries its source and cut-off date.

  2. Uncertainty is labeled.

    Every deliverable marks [ESTIMATE], [ASSUMPTION] and [DATA GAP].

  3. People approve what matters.

    Signatures, money and legal judgment always go through a person.

  4. Reproducible deliverables.

    Markdown first, then PDF with pandoc and Typst; generator scripts rebuild the final files.

  5. Decisions are logged.

    Decision registers and GREEN, YELLOW and RED status lights.

  6. Test like a user.

    Simulated buyers, golden test sets, automated tests and checks against the live site.

  7. Validation is explicit.

    Studies state what must be confirmed with the competent authority and are not legal advice.

15 · Stack

AI
  • Claude Code (skills, plugins, subagents, MCP connectors, hooks, scheduled tasks)
  • Claude API
  • Voyage AI embeddings
  • ChatGPT and Codex
  • Higgsfield (image, video, voice and music)
  • Grok Bot
  • Meta's Muse
Data
  • Python
  • PostgreSQL with pgvector
  • Playwright
  • pdfplumber and OCR
  • Excel
Automation
  • n8n
  • Docker
  • launchd
  • scheduled tasks
Video
  • Remotion
  • ffmpeg
  • Whisper
  • Apple Vision (Swift)
Web
  • Next.js
  • TypeScript
  • static HTML and CSS
  • GSAP
  • Cloudflare Workers
  • Vercel
  • Stripe
  • Resend
  • Web3Forms
Marketing and analytics
  • GA4
  • Microsoft Clarity
  • Vercel Analytics
  • Google Ads
  • Meta Pixel
  • Google Search Console
Documents
  • Markdown
  • pandoc
  • Typst
  • Word with tracked changes
Knowledge
  • Obsidian