Portfolio

Selected Work

A few things I have designed, built, and shipped.

jorgecastro.ai
Project 01AI Agents & SaaS
Agentic Content & SEO Automation Platform
An agentic platform that writes and publishes SEO content, answers reviews across platforms, and tracks how a brand shows up in Google, ChatGPT, Gemini and Perplexity. Built for a Swedish SEO consultancy and run by its clients every day.
Role
AI & Full-Stack Engineer
Client
Marketlytics (Jorge Castro)
Timeline
Ongoing
- Blog generation, reputation replies and rank tracking each run as a tool an agent can call
- Background work moved onto Cloud Tasks with idempotent retries and an inspectable queue
- One gateway across OpenAI, Anthropic and Gemini, so a model can be swapped per task
- Tracks brand visibility in ChatGPT, Gemini, Perplexity and Google AI Overviews
View live
Node.jsExpressMySQLNext.jsReduxGoogle Cloud RunCloud TasksOpenAI+4 more
skimming.ai
Project 02AI SaaS
Skimming, Bulk AI Workspace for Video and Documents
A multi-tenant RAG workspace that ingests YouTube videos, podcasts and documents in bulk and answers questions across all of them, grounded in the sources rather than the model's memory.
Role
AI & Full-Stack Engineer
Client
Marketlytics
Timeline
Ongoing
- Whisper transcription for video and audio, Docling for PDFs, slides and spreadsheets
- LanceDB vectors through LlamaIndex, four language models behind LiteLLM
- Teams, roles and per-tenant isolation on an Express and MySQL layer
- Deterministic language detection replaced an LLM call that slowed every request
View live
PythonFastAPILlamaIndexLanceDBOpenAILiteLLMWhisperDocling+5 more
responsly.ai
Project 03AI Automation & SaaS
Responsly, Autonomous Google Reviews AI
An autonomous Google Business reviews responder for multi-location businesses. It fetches new reviews, replies in the business's own voice and language, manages posts and events, and bills per location.
Role
Lead Engineer, team of three
Client
Marketlytics
Timeline
Ongoing
- Led the build with two junior developers; owned architecture, code review and releases
- Per-location Stripe billing: one customer per user, one subscription per location
- Google accounts keyed on the stable sub claim rather than a mutable email address
- Every degraded AI path now refuses rather than publishes, after an outage posted the wrong language
View live
Node.jsExpressMySQLRedisSocket.ioGoogle Business Profile APIGoogle OAuthOpenAI+5 more
www.findurlawyer.com
Project 04Web App & SaaS
FindUrLawyer, Legal Marketplace for the Philippines
A two-sided marketplace where people in the Philippines describe a legal problem in plain words and verified lawyers reply to them directly. Security hardening, production monitoring, moderation and a headless blog.
Role
Full-Stack Engineer
Client
FindUrLawyer
Timeline
Ongoing
- Hardened the whole authentication surface: reset flow, OTP rate limiting, admin routes
- A 16-probe monitor with incident dedup, running every five minutes against live traffic
- User reporting wired into the moderation console, with an audited action trail
- Headless WordPress blog with webhook revalidation, and replay analytics on the funnel
View live
Next.js 16TypeScriptSupabasePostgreSQLRow Level SecurityWordPress REST APIContentsquareApache
play.google.com
Project 05Mobile App
Soor Exchange, Shariah-Compliant Crypto Trading App
A Shariah-compliant crypto exchange app for iOS and Android with live trading and fiat payments, built from the trading-API integration through to both store listings.
Role
Senior React Native Developer
Client
ProgrammX
Timeline
6 months
- Live on the App Store and Google Play
- Trading-API and payment-gateway integration through to store submission
- Cold-start load time cut 35 percent, 99.9 percent uptime since launch
- Admin panel in Next.js for listings and compliance
View live
React NativeExpoTypeScriptNode.jsNext.jsREST APIsBlockchainPayment gateway integration+1 more
nvizy-decentralized-ai-health-app
Project 06AI & Blockchain Health
Nvizy, AI Doctor and Patient App on Chain
A decentralised healthcare app where an AI intake chat gathers the patient's history, flags the key findings, and hands the doctor a one-page summary, with records secured on chain.
Role
React Native & AI Developer
Client
ProgrammX
Timeline
4 months
- AI intake chat collects history and symptoms before the appointment
- Key findings surfaced for the doctor as a one-page summary
- Patient records written to chain, readable only by the people the patient allows
- Biometric and liveness checks before a record is unlocked
Private project
React NativeNode.jsOpenAIBlockchainSmart contractsExpressMongoDB
videofunker.ai
Project 07AI SaaS
Video Funker, AI Sales Video SaaS
Done-for-you LinkedIn video for B2B founders: an eight-section AI campaign generator, HeyGen avatar and voice-clone rendering, and Stripe plans where every advertised limit is enforced in code.
Role
Full-Stack & AI Engineer
Client
Video Funker
Timeline
3 months
- Eight-section campaign generator: ICP, research, article, LinkedIn posts, outbound scripts
- HeyGen avatar creation and voice-clone rendering, queued and resumable
- Every plan limit on the pricing page has a matching enforcement check on the generation path
- Free tier gives the full written campaign plus a watermarked 30-second preview
View live
Node.jsExpressReactOpenAIHeyGen v3StripeGA4Docker
outreach-engine
Project 08AI Automation & SaaS
Outreach Engine, AI Cold Email at Scale
A cold-email engine that sources leads from Apollo, writes one grounded opening line per lead, sends through a 30-mailbox Smartlead fleet at its 900-a-day ceiling, and classifies every reply into an action.
Role
AI & Full-Stack Engineer
Client
Marketlytics (Jorge Castro)
Timeline
Ongoing
- 30 mailboxes at their real ceiling of 900 sends a day, across 19 campaigns
- Reply classifier and action engine: urgent, draft, domain-block, referral
- Autonomous multi-turn reply agent, off by default, one full gate stack before every send
- Measured that Smartlead's cross-campaign dedup does not work, and enforced it at the database instead
Private project
Node.jsExpressMySQLRedisNext.jsReduxApollo APISmartlead API+5 more
okami-museum-ai-guide
Project 09RAG & AI Chatbot
Okami Museum Multilingual AI Guide
A multilingual RAG guide chatbot deployed on a Japanese museum's walking robots and its visitor app, answering questions from the museum's own archive rather than the open web.
Role
AI & Full-Stack Engineer
Client
Okami Museum, Japan
Timeline
3 months
- Answers only from the museum's own content, never from the model's memory
- Runs on the museum's walking robots and in the visitor app
- Handles visitor questions in several languages from one index
- Admin panel for training sources, so the museum updates the guide itself
Private project
Node.jsExpressRAGVector searchOpenAIReactMySQL
jenkins-real-estate-flip-saas
Project 10Web App & SaaS
Jenkins, US Real Estate Flip-Finder SaaS
A SaaS that finds undervalued US homes, estimates after-repair value, models each market from historical flips, and automates outreach to sellers.
Role
Full-Stack & AI Engineer
Client
Marketlytics
Timeline
Ongoing
- Scores undervalued homes and estimates after-repair value
- Market models built from historical flip data, per metro
- Automated seller outreach from the scored list
- Dashboards for flip activity, confidence mix and market coverage
Private project
Next.jsNode.jsExpressMySQLPythonDashScopeDockerGoogle Cloud Run
saleslights.com
Project 11AI SaaS
Saleslights, AI Sales Video Platform
An AI sales-video platform with OpenAI content generation, HeyGen avatars and Stripe billing, plus a headless-WordPress blog and a marketing site with rollback deploys.
Role
Full-Stack Engineer
Client
Saleslights
Timeline
4 months
- OpenAI content generation and HeyGen avatar video from one brief
- Stripe subscriptions and JWT auth on an Express and MongoDB API
- Headless-WordPress blog server-rendered inside the app, with a sanitised HTML pipeline
- Marketing site with a deploy script that rolls back on failure
View live
Node.jsExpressMongoDBViteReactOpenAIHeyGenStripe+3 more
toolbox-tim
Project 12Enterprise SaaS
Toolbox Tim, Websites and Phone Lines for UK Tradespeople
In development. An enterprise SaaS for UK tradespeople: generated multi-tenant websites on one SSR renderer, a virtual landline with missed-call text-back, and call-to-quote with live transcription.
Role
Architecture & Build Lead
Client
Toolbox Tim (UK)
Timeline
In development, 2026
- One SSR renderer serves every tenant site by Host header, with edge cache purged on publish
- Twilio virtual landline, missed-call text-back, and UK area-code consent logging
- Deepgram streaming transcription for call-to-quote
- Prisma 7 on Postgres, Cloud Tasks workers, Better Auth, all in a Turborepo monorepo
Private project
Next.js 16React 19Express 5Prisma 7PostgreSQLRedisCloud TasksBetter Auth+4 more
aust.edu.pk
Project 13Website
Abbottabad University Website & Admin Portal
The public website and a complete admin portal for a Pakistani university, built so non-technical staff manage admissions notices, programmes and news themselves.
Role
Full-Stack Developer & Project Manager
Client
Abbottabad University of Science & Technology
Timeline
3 months
- Live at aust.edu.pk
- Admin portal that non-technical staff run without a developer
- Admissions, programmes, faculties and news managed from one place
- Delivered on time as project manager and lead developer
View live
Next.jsNode.jsExpressMySQLTailwind CSSDockerAWS
ai-portfolio-assistant
Project 14RAG & AI Chatbot
AI Portfolio Assistant
The assistant on this site: a RAG chatbot that answers a recruiter's questions about my work from my real project data, and knows when to say it does not know.
Role
Full-Stack & AI Engineer
Client
Personal
Timeline
2 weeks
- Grounded in the same project records that render this portfolio
- Answers what I built, which stack, and whether I am available, from data rather than a script
- Declines questions it has no source for instead of inventing an answer
- Conversations are logged so I can see what people actually ask
Private project
Next.jsExpressMySQLOpenAIGeminiRAGSocket.io
ai-posture-coach-app
Project 15AI & Mobile
AI Posture Coach, Fitness Tracking App
A workout app with a trained pose-estimation model for live posture tracking, form correction and rep counting, running on the phone.
Role
React Native & ML Developer
Client
Freelance
Timeline
2 months
- Live posture tracking from the phone camera, on device
- Form correction cues during the exercise, not after
- Automatic rep counting per exercise
- Trained pose model tuned for gym movements
Private project
React NativeTensorFlow LitePose estimationNode.jsFirebase
ai-coding-buddy-platform
Project 16EdTech & Developer Tools
AI Coding Buddy & VS Code Extension
A real-time collaborative coding-education platform with a companion VS Code extension, where students learn and build together with an AI helper alongside.
Role
Full-Stack Developer
Client
Sarte Solution
Timeline
3 months
- Real-time shared editing between students and instructor
- Companion VS Code extension so students work in their own editor
- AI helper that explains rather than solves
- Built and shipped at Sarte Solution
Private project
Next.jsNode.jsSocket.ioVS Code Extension APIOpenAIMySQL
chess-forum
Project 17Web App
Chess Forum
A community platform for chess players with live game analysis, interactive lessons, an active forum and global tournaments.
Role
Full-Stack Developer
Client
Personal
Timeline
6 weeks
- Live game analysis with move-by-move evaluation
- Interactive lessons with playable positions
- Forum with threads, replies and moderation
- Tournament brackets with global standings
Source code
DjangoPythonPostgreSQLJavaScriptWebSockets

Research

Research & writing

Papers, articles, talks and open source where I share what I learn.

Paper2026

Context-Efficient Orchestration of Production LLM Agents

A working paper proposing token-efficient patterns for coordinating multi-agent LLM systems with the Model Context Protocol, drawn from production agent deployments. Manuscript in preparation.

AI AgentsMCPLLM Systems

Paper2026

Source-Grounded RAG: Reducing Hallucination with Attributed Retrieval

A working paper on retrieval-augmented generation that attaches verifiable source citations to each claim and abstains under low retrieval confidence. Manuscript in preparation.

RAGHallucinationRetrieval

Paper2025

Green-LLM: KV-Cache Compression via Reinforcement Learning

A reinforcement-learning framework for adaptive KV-cache compression during Llama-3 inference, reducing energy use by 28 percent with under 1 percent perplexity loss, aligned with green-AI goals.

LLM EfficiencyReinforcement LearningInference

Paper2024

Lightweight Image Matching with XFeat

A study replicating and extending the XFeat accelerated features approach for efficient local feature detection and matching on resource constrained devices.

Computer VisionCNNFeature Matching

Link

Paper2024

FedArabic: Federated LoRA Instruction Tuning of Llama-3 on Arabic Social-Media Data

A LoRA-based federated fine-tuning framework that collaboratively trains across clients on Emirati-dialect tweets, reaching 96 percent of centralized BLEU performance with minimal communication overhead.

Federated LearningLoRANLP

Recognition

Awards & recognition

Honors and recognition for the work and the impact behind it.

2024

Employee of the Year

Sarte Solution

Recognized for leading the team and delivering all projects on time.

Professional

2023

Vice President, Cyber Security Society

AUST

Led initiatives and events promoting cybersecurity awareness.

Community

2023

Merit Scholarship

AUST

Awarded for outstanding academic performance throughout the program.

Professional

2023

Foster Flagship Training Program

Foster

Completed a flagship program focused on business management fundamentals.

Professional

2022

HULT Prize Regional Finalist

HULT Prize Foundation

Recognized among the top teams at the regional round for a social impact startup concept, pitched to a panel of judges.

Lahore, PakistanTop 6 of 120+ teams

ProfessionalView

Have something in mind?

If you are building a product or need an extra pair of hands, I would love to hear about it.

Start a conversation

Agentic Content & SEO Automation Platform

Skimming, Bulk AI Workspace for Video and Documents

Responsly, Autonomous Google Reviews AI

FindUrLawyer, Legal Marketplace for the Philippines

Soor Exchange, Shariah-Compliant Crypto Trading App

Nvizy, AI Doctor and Patient App on Chain

Video Funker, AI Sales Video SaaS

Outreach Engine, AI Cold Email at Scale

Okami Museum Multilingual AI Guide

Jenkins, US Real Estate Flip-Finder SaaS

Saleslights, AI Sales Video Platform

Toolbox Tim, Websites and Phone Lines for UK Tradespeople

Abbottabad University Website & Admin Portal

AI Portfolio Assistant

AI Posture Coach, Fitness Tracking App

AI Coding Buddy & VS Code Extension

Chess Forum

Context-Efficient Orchestration of Production LLM Agents

Source-Grounded RAG: Reducing Hallucination with Attributed Retrieval

Green-LLM: KV-Cache Compression via Reinforcement Learning

Lightweight Image Matching with XFeat

FedArabic: Federated LoRA Instruction Tuning of Llama-3 on Arabic Social-Media Data

Employee of the Year

Vice President, Cyber Security Society

Merit Scholarship

Foster Flagship Training Program

HULT Prize Regional FinalistVerified

Have something in mind?

HULT Prize Regional Finalist