I build applied AI systems — for production.
Agents, pipelines, and AI products. From architecture to deployment.
Sydney, Australia•5+ years in tech — from industrial systems to applied AI•Industrial engineer → AI & software•Open to full-time AI/ML roles
Applied AI systems · Production-ready agents, workflows, and intelligent software•
Full-stack delivery · Product thinking, architecture, frontend, backend, and integration•
Technical depth · Systems thinking, engineering rigor, and real implementation detail•
Open to roles · Applied AI, AI systems, and product-minded technical teams•
Applied AI systems · Production-ready agents, workflows, and intelligent software•
Full-stack delivery · Product thinking, architecture, frontend, backend, and integration•
Technical depth · Systems thinking, engineering rigor, and real implementation detail•
Open to roles · Applied AI, AI systems, and product-minded technical teams•
AI Engineering
Two ways to explore the portfolio. Both live.
Grounded Q&A from the knowledge base and a guided agent that matches visitor intent. What follows is the flagship work behind these systems.
Q&A · ClaudeKnowledge-grounded Q&A
Q&A Assistant
Visitors ask questions and get answers anchored in structured portfolio context—so the conversation stays on real work and credibility, not generic chat.
Claude Sonnetpgvector HNSWHybrid searchSupabase
Retrieval pipeline
Query
Embed
Hybrid
Score
Claude
vector HNSW + full-text search · multi-signal ranking
Try it
Agentic · GPT-4oGuided by intent
Portfolio Agent
Classifies intent and visitor type, then routes to the right projects and angles—recruiters, technical peers, and hiring managers don’t all need the same tour.
GPT-4oOpenAI Agents SDK7 custom toolsClaude Haiku router
Detected
recruiter
0.87✓
Router → recruiter (0.87)✓
Route: role_fit↻
Tool: recruiter_evidence7 custom tools · file_search · input/output guardrails
Try it
Featured work
Work that runs in the real world.
From production systems to AI agents — engineered for impact, not just demos.
Software · AI Agent
agent = Agent(model, tools=[...])
result = await agent.execute(task)
# → reasoning → action → done
Q&A Assistant — RAG Grounded
Hybrid retrieval over an autobiographical knowledge base. Every answer cites its source. Handles EN + ES natively.
Claude SonnetSupabasepgvectorNext.js
AI Agent · GPT-4o
guide = Agent(gpt4o, portfolio_tools)
response = guide.chat(visitor_query)
# → navigating projects & skills
Portfolio Agent — Agentic Explorer
GPT-4o with 8 custom tools and a Vector Store. Ranks projects against your role and answers in cards, not paragraphs.
GPT-4oAgents SDKNext.js
Coming Soon
pipeline = ProductionPipeline(ml=True)
pipeline.automate(decisions)
# → 80% automation, -25% lead time
Automated Production System
Strategic automation of production processes with ML-enabled decisions at CEM S.A.S. 80% automation, -25% lead time.
PythonTensorFlowML Automation
Coming soon
Coming Soon
forecast = model.predict(catalog)
alloc = optimize(forecast, supply)
# → +18% forecast accuracy
Sales Forecasting Dashboard
Forecasting + stock allocation for the full catalog in a constrained-supply region at AB InBev. +18% forecast accuracy.
PythonRegressionPower BI
Coming soon
Coming Soon
model = train(dataset, epochs=50)
results = model.evaluate(test_set)
# → research in progress
AI Thesis — Industrial Master's
Final thesis for the Master's in Industrial Engineering (Pontificia Universidad Javeriana). Pivoted to an AI topic after the Attention Is All You Need awakening.
AI / MLResearchPython
Coming soon
Coming Soon
cnn = build_baseline(input_shape)
vgg = transfer_learn(VGG16, unfrozen=4)
# → gender + age multi-output
UTKFace Multi-Output Model
Gender classification + age prediction. Baseline CNN vs VGG16 transfer learning, with and without augmentation. CQU Master of IT deep-learning coursework.
CNNVGG16TensorFlow
Coming soon
Software · AI Agent
agent = Agent(model, tools=[...])
result = await agent.execute(task)
# → reasoning → action → done
Q&A Assistant — RAG Grounded
Hybrid retrieval over an autobiographical knowledge base. Every answer cites its source. Handles EN + ES natively.
Claude SonnetSupabasepgvectorNext.js
AI Agent · GPT-4o
guide = Agent(gpt4o, portfolio_tools)
response = guide.chat(visitor_query)
# → navigating projects & skills
Portfolio Agent — Agentic Explorer
GPT-4o with 8 custom tools and a Vector Store. Ranks projects against your role and answers in cards, not paragraphs.
GPT-4oAgents SDKNext.js
Coming Soon
pipeline = ProductionPipeline(ml=True)
pipeline.automate(decisions)
# → 80% automation, -25% lead time
Automated Production System
Strategic automation of production processes with ML-enabled decisions at CEM S.A.S. 80% automation, -25% lead time.
PythonTensorFlowML Automation
Coming soon
Coming Soon
forecast = model.predict(catalog)
alloc = optimize(forecast, supply)
# → +18% forecast accuracy
Sales Forecasting Dashboard
Forecasting + stock allocation for the full catalog in a constrained-supply region at AB InBev. +18% forecast accuracy.
PythonRegressionPower BI
Coming soon
Coming Soon
model = train(dataset, epochs=50)
results = model.evaluate(test_set)
# → research in progress
AI Thesis — Industrial Master's
Final thesis for the Master's in Industrial Engineering (Pontificia Universidad Javeriana). Pivoted to an AI topic after the Attention Is All You Need awakening.
AI / MLResearchPython
Coming soon
Coming Soon
cnn = build_baseline(input_shape)
vgg = transfer_learn(VGG16, unfrozen=4)
# → gender + age multi-output
UTKFace Multi-Output Model
Gender classification + age prediction. Baseline CNN vs VGG16 transfer learning, with and without augmentation. CQU Master of IT deep-learning coursework.
CNNVGG16TensorFlow
Coming soon
Wheel here to move cards left/right. Move cursor out to continue page scroll.
Software · AI Agent
agent = Agent(model, tools=[...])
result = await agent.execute(task)
Q&A Assistant — RAG Grounded
Hybrid retrieval over an autobiographical knowledge base. Every answer cites its source. Handles EN + ES natively.
Claude SonnetSupabase
AI Agent · GPT-4o
guide = Agent(gpt4o, portfolio_tools)
response = guide.chat(visitor_query)
Portfolio Agent — Agentic Explorer
GPT-4o with 8 custom tools and a Vector Store. Ranks projects against your role and answers in cards, not paragraphs.
GPT-4oAgents SDK
Coming Soon
pipeline = ProductionPipeline(ml=True)
pipeline.automate(decisions)
Automated Production System
Strategic automation of production processes with ML-enabled decisions at CEM S.A.S. 80% automation, -25% lead time.
PythonTensorFlow
Coming soon
Coming Soon
forecast = model.predict(catalog)
alloc = optimize(forecast, supply)
Sales Forecasting Dashboard
Forecasting + stock allocation for the full catalog in a constrained-supply region at AB InBev. +18% forecast accuracy.
PythonRegression
Coming soon
Coming Soon
model = train(dataset, epochs=50)
results = model.evaluate(test_set)
AI Thesis — Industrial Master's
Final thesis for the Master's in Industrial Engineering (Pontificia Universidad Javeriana). Pivoted to an AI topic after the Attention Is All You Need awakening.
AI / MLResearch
Coming soon
Coming Soon
cnn = build_baseline(input_shape)
vgg = transfer_learn(VGG16, unfrozen=4)
UTKFace Multi-Output Model
Gender classification + age prediction. Baseline CNN vs VGG16 transfer learning, with and without augmentation. CQU Master of IT deep-learning coursework.
CNNVGG16
Coming soon
01 — About
Engineer
first.
Builder
always.
Fabian Roberto
Software & Applied AI Systems Builder · Sydney, Australia
Most AI builders start from code. I started from real operations — and that changes how I design, build, and ship.
Engineering Rigor
5+ years of systems & process thinking from industrial engineering.
Systems DesignOptimizationLean/Six SigmaOperations ResearchData Modeling
Software Execution
Full-stack development with product design sense and modern tooling.
TypeScriptPythonJavaReact/Next.jsPostgreSQLDocker/AWS
Applied AI Systems
Since 2021 — from ML models to production agents & LLMs.
TensorFlow/PyTorchLangChainClaude/OpenAI APIRAGAgentsNLP
years experience
industries
countries
tools in stack
OpenAI
Anthropic
Google Gemini
Meta Llama
Mistral AI
Cohere
LangChain
LangGraph
LlamaIndex
PyTorch
TensorFlow
scikit-learn
OpenAI
Anthropic
Google Gemini
Meta Llama
Mistral AI
Cohere
LangChain
LangGraph
LlamaIndex
PyTorch
TensorFlow
scikit-learn
Python
PostgreSQL
Redis
SQLite
MongoDB
Supabase
AWS
Docker
Kubernetes
Linux
Git
GitHub
Node.js
Firebase
Vercel
Python
PostgreSQL
Redis
SQLite
MongoDB
Supabase
AWS
Docker
Kubernetes
Linux
Git
GitHub
Node.js
Firebase
Vercel
02 — Process
How
I
engineer
01
Systems first
Industrial engineering trained me to see the full system — constraints, data flows, failure modes — before writing a line of code. The architecture always decides the outcome.
Arena SimulationProduction Control SystemFreshMart OS
02
Research-driven
I don't follow hype. I read papers, benchmark alternatives, and validate with experiments. From RAG architectures to agent frameworks — every decision has evidence behind it.
Hybrid retrievalVector + FTS scoringEval-driven iteration
Lessons shape the next system
04
Ship, measure, iterate
Production is where the real feedback lives. I ship fast, instrument every decision path, and iterate based on real user behavior — not assumptions about how agents should work.
Q&A AssistantSQL PlaygroundFreshMart OS
03
AI-augmented building
I don't just build AI — I build with AI. Cursor, Claude, and custom agents are part of my daily workflow, from rapid prototyping to production-grade code.
Claude CodeCursorCustom agent pipelines
01
Systems first
Industrial engineering trained me to see the full system — constraints, data flows, failure modes — before writing a line of code. The architecture always decides the outcome.
Arena SimulationProduction Control SystemFreshMart OS
02
Research-driven
I don't follow hype. I read papers, benchmark alternatives, and validate with experiments. From RAG architectures to agent frameworks — every decision has evidence behind it.
Hybrid retrievalVector + FTS scoringEval-driven iteration
03
AI-augmented building
I don't just build AI — I build with AI. Cursor, Claude, and custom agents are part of my daily workflow, from rapid prototyping to production-grade code.
Claude CodeCursorCustom agent pipelines
04
Ship, measure, iterate
Production is where the real feedback lives. I ship fast, instrument every decision path, and iterate based on real user behavior — not assumptions about how agents should work.
Q&A AssistantSQL PlaygroundFreshMart OS
↑ 01Lessons shape the next cycle
03 — Story
My
Journey
2018 – 2020💼 Work+7
Production Analyst → Junior Manager
E.C. Ingeniería S.A.S · Colombia
Hired through university applied project to implement a production control system. Led process improvement, workflow simulation, and automation proposals. Promoted to Junior Manager — operations planning, import logistics, and production scheduling.
15% throughput improvement via digital twin simulation~8% unit cost reduction
2020 – 2022💼 Work+9
Director of Strategic Projects
CEM S.A.S · Bogotá, Colombia
Led ~20 people, reported directly to the board of shareholders. Applied ML professionally — regression models and neural networks with Python and TensorFlow for demand forecasting, production scheduling, and import planning. Deployed Power BI dashboards and integrated ERP systems. Achieved 80% production automation and 25% lead time reduction, measured via Power BI dashboards reporting to the board over 24 months.
80% production automation (Python/TensorFlow + ERP + rule-based control)25% lead time reduction vs 2020 baseline
Jun – Sep 2022💼 Work+10
Sales Supervisor
KOPPS COMMERCIAL / AB InBev · Cartagena, Colombia
Managed a sales team of ~7 people across the Norte/Cartagena region — KPI tracking, metrics compliance, and NPS. Identified opportunities to integrate technology into sales operations, building forecasting and automation tools that drove measurable results.
Nov 2023 – Present💼 Work+8
Specialized Tech → Lead Tech
Procom / Ingenico · Sydney, Australia
Payment terminal support for major Australian banks — NAB, CBA, Westpac, Suncorp. Hardware, software, networking across Windows and Linux. Promoted to Lead Tech (Dec 2025) — coordinating technicians, managing inventory, ensuring compliance.
Clients: NAB, CBA, Westpac, Suncorp + othersPromoted to Lead Tech — Dec 2025
04 — Contact
Let's
build
something.
I moved from Colombia to Sydney to be closer to where AI is being built. If you're working on something that matters — I want to hear about it.
Open to opportunities
AI/software engineering roles, research collaboration, and interesting side bets. If you think there's a fit — I'd love to have a conversation.
or chat on WhatsApp