Joan Escamilla — AI Agent Developer

Background

Experience

Built and ship data-entry screens, reports, and functional modules; kept the Medication Adherence app stable across multiple framework migrations.
Volunteered onto the company's internal AI initiative — most developers remain skeptical, so I've been the one piloting new AI workflows and proposing where else AI could help, beyond my own tickets.
Built an agent that encodes the codebase's unwritten SQL conventions — cuts routine script time in half, complex schema-design tasks by a similar margin.
Built a second agent that audits code against WCAG rules and proposes phased, human-approved fixes, extended to generate per-file SVN patch files, then rebuilt to scan an entire codebase on its own — projected to cut the 50-to-60-hour-per-system estimate to about 5 hours including diff review, across 50+ systems needing the work; full-system validation is still pending.
Proposed and built the company's first MCP server after the CEO asked whether AI could cut a 40-to-50-hour manual client-import process, then built two more for internal dev workflows — one that answers schema and convention questions directly against our database, and one that uses SVN to answer codebase and commit questions and group edits into commits, cutting a class of manual SSMS/SVN lookups to seconds.
Led a technical meeting with AWS engineers on scaling our MCP infrastructure via AgentCore, Bedrock, and Lambda — researched and presented the architecture and identified AgentCore as the fix for our missing LLM API endpoint within our existing AWS budget; later demoed the servers directly to a senior dev, who called it a step toward proper system documentation.
Leading an inaugural UX/UI initiative to modernize a 30-year-old interface — architected a reusable, scalable CSS system with tokenized colors and folded ADA compliance into the same pass; after a senior dev demo, rebuilt the internal site's entire frontend from 20-year-old default Bootstrap styling to a modern, modular system in 3 days.
Now leading a new design system initiative born out of downtime while the EMR product awaits contract approval — leading 2 junior developers with a 2-week demo timeline — stripping our most modern product into documented, reusable components and practices for the team.
Selected and mentored a 5-developer university capstone team, hiring one into a junior role — still mentoring him directly; he just hit one year with the company.

ColdFusion SQL Server JavaScript CSS GitHub Copilot MCP SVN AWS Agentic AI ADA / WCAG Design Systems

Tracked progress across 20+ students and delivered structured feedback to parents monthly. Students averaged a 30% improvement in programming skills by internal metrics.

React Python Java Curriculum design

Converted the Tableau JS API into reusable React components to reduce integration overhead. Also built a React app for testing and demoing API endpoints, which accelerated bug identification across the team.

React Tableau Census API JavaScript

Selected work

Projects

Panini WC26 sticker collection view — teams grouped with have/need/extra status per sticker

Trade Builder scoring a proposed trade — outstanding value, emblem/photo bonuses, and completed-team bonuses

Analysis view showing priority targets and best trade bait across teams

Achievements view tracking collection milestones

Panini World Cup Sticker Tracker

Built entirely through Claude artifacts, without ever looking at the underlying code — describe what you want, see it live, ask for changes until it's right. Scan your World Cup sticker collection, see what's missing, and check whether a proposed trade is actually worth making.

No real data persistence in an artifact, so collections import/export as CSV — haves, missing, and duplicates, re-importable next session
Uses distance as a second filter on top of trade quality — whether it's worth the gas or the trip, not just the cards

Claude Artifacts Vibe Coding Trade Evaluation

Harmonic pattern detector drawing an XABCD pattern on a GME chart, with a pattern-history success-rate table

Dynamic Support/Resistance indicator marking live resistance and support zones with percent-to-level labels

Both indicators running together — harmonic pattern, support/resistance zones, and a projected target on the same chart

AI-Assisted Trading Indicators

Built a custom Claude Skill for Pine Script development — it knows the syntax and conventions, flags likely bugs before they ship, and tracks fixes across sessions. Used it to build a suite of indicators, including a harmonic pattern detector and a dynamic support/resistance tool.

Dynamic Support/Resistance detects multi-touch pivot zones and draws them live — 3,000+ member Discord community using it
Harmonics indicator rebuilds the underlying pattern-recognition libraries from scratch — prediction zones, historical outcome tracking

Pine Script TradingView Claude Skills Harmonic Patterns

World Cup Overview tab — full tournament bracket with next-match callout and predicted outcome link

Predictor tab — win/draw/loss probability chart and match verdict for Argentina vs Egypt

Team Comparison tab — Elo ratings, rank, and recent form side by side

Model Health tab — calibration, Brier score, log-loss, and top-pick hit rate

World Cup Predictor

Same shape of project as the UFC predictor, but backed by real APIs this time — the constraint was free-tier data caps, not scraping. Used this one to test how far a mostly hands-off, Claude-directed build could go, guiding it only at the moments that mattered.

Built a caching layer on its own to stay under the free tier's API request limits
Surfaces failed data pulls directly in the UI instead of failing silently

Claude Code Agentic AI Python

Inbox triage view sorting job emails from other mail, with an analyzed job posting showing role, match, and pay details

Analyzed roles table comparing verdict, salary, match score, and quality-of-life score across job postings

AI Job-Search Assistant

A live Claude artifact connected through connectors to email, calendar, and Google Drive — surfaces the job-related messages buried in an inbox and rates them against a resume, instead of combing through email by hand.

Reads email + calendar via connectors to find and rank job-relevant messages against a resume
No real database available here either — Google Drive connector as pseudo-storage, good enough for a personal tool

Claude Connectors Live Artifacts Job Search

Clawdbot — the open-source AI harness used to run this local agent

Personal AI Assistant (Local Agent)

A fully local AI agent run through OpenClaw on a Mac Mini, reachable from my phone over Discord. Tested several Ollama models to balance hardware limits against answer quality, then built it out to handle email, calendar, and trading-setup questions grounded in Bulkowski's chart-pattern rules.

Memory runs on Obsidian's graph links instead of a vector database — markdown notes become a decision tree the agent navigates for context
Full device access meant testing several context-storage approaches before answers were reliable

OpenClaw Ollama Obsidian Agentic AI Python

UFC octagon — cover image for the fight-data scraping and prediction pipeline

UFC Fight Predictor — Data Pipeline

An idea that stalled for years on a data problem — the UFC site has no API, and its fight-record URLs aren't structured in any way you can parse directly. Built with Claude Code, entirely inside a Jupyter notebook, applying data-science methods from college.

Custom scraper built by Claude Code where no public API existed
Pulled and normalized missing data into a clean dataset ready to build predictions on

Claude Code Jupyter Python Web Scraping

ADA Compliance Agent

Senior devs estimated 50 to 60 hours per system for screen-reader compliance, across 50+ systems needing it. Built an agent instead — it now scans an entire codebase on its own, converts to semantic HTML, adds ARIA labels, checks JS for screen-reader UX issues like focus handling after a form submits, and generates a per-file SVN patch so a developer can apply the fix directly.

Every fix ships with an impact/feasibility score and the specific WCAG rule it satisfies — proposes, never applies, so the developer stays in the loop
10 files, 50+ changes applied: ~30 min with the agent vs. an estimated 3 hours by hand
Full-codebase scan: 50–60 hour manual estimate projected to drop to about 5 hours per system — full-system validation still pending, blocked on token limits
Validation plan: benchmark against an existing ADA compliance tool before and after the agent runs, to measure the improvement directly

GitHub Copilot Agentic AI WCAG Semantic HTML

SQL Convention Agent

Templated the codebase's unwritten SQL conventions — audit logging, naming, history tables with trigger generation on new tables — so a GitHub Copilot agent applies them automatically instead of me piecing them together by hand. Ask it to add a column and it doesn't just write the ALTER statement; it adds the logging and naming conventions that go with it.

Routine scripts: ~20 min → under 10
Complex FK backfill across several tables: 30–40 hr estimate cut to about half — first pass was ~80% there, rest went to testing and verification

GitHub Copilot SQL Server Agentic AI Prompt Engineering

MCP Server PDF Import Pipeline cover graphic

MCP Server Infrastructure

The CEO asked whether AI could cut a 40-to-50-hour manual client-import process handled by fiscal managers. Proposed and built the company's first MCP server, then expanded into dev-workflow tools — one that answers schema and convention questions directly against our database, and one that uses SVN to answer codebase and commit questions.

Tested several PDF extraction approaches — text libraries, PDF-to-HTML, PDF-to-Excel, vision-based table detection — none reached clean, production-ready output yet
The actual fix: running a PDF directly through an LLM instead of a coded pipeline sidesteps the extraction ceiling entirely — just needs an LLM API endpoint to run against
DB and codebase MCP servers replaced routine SSMS/SVN lookups — schema relationships, missing convention scripts, and commit history all answered directly, no agent context limits
Led a technical meeting with AWS engineers on hosting these servers via AgentCore, Bedrock, and Lambda — solves the missing LLM API endpoint and moves them from my laptop to something the whole team can use
Demoed to a senior dev in our internal AI meeting — called it a step in the right direction for proper system documentation

MCP SVN AWS Python SQL Server Agentic AI

EMR UI/UX Overhaul

Our EMR's interface hadn't meaningfully changed in about 30 years, and the company has no dedicated frontend or design team. Used AI to architect a reusable, scalable CSS system with tokenized colors and apply it across the product, folding ADA compliance into the same pass instead of treating it as a separate project.

Demoed to the team — senior dev said it should ship immediately if not for the contract-approval process
ADA compliance rules folded directly into the rebuild, using the same standards behind the ADA Compliance Agent
A second senior dev asked for the same UI/UX on our internal site — rebuilt that frontend from 20-year-old default Bootstrap styling to a modern, modular system in 3 days
Now leading a design system initiative born out of downtime while the EMR product awaits contract approval — leading 2 junior developers, demo in 2 weeks

CSS UI/UX WCAG Design Systems

I build AI agents.

About