AIKHEMIST CV

Mauro Orlando

Data & AI Solutions Architect & Developer

28 years turning enterprise data into decisions. Now I design and ship AI systems end to end.

I understand the business problem and the decisions behind it, and turn that into data models, clear visualizations and working systems — without needing a finished technical spec to start.

  • Remote
  • Full-time, contract or part-time
  • Based in SJC - São Paulo, Brazil (BRT, UTC−3)
  • English, Spanish, Portuguese

2nd place in the Dr.DocBench Challenge · EMNLP 2026

Download CV Contact

What I do

AI Agents & GenAI Products

Multi-agent systems and GenAI applications in production: the architecture, the back end and the interface people actually use.

Built on
Software architecture since the start of my career, plus 2 years at IBM.
Recent work
DocBench Agents (2nd place, EMNLP 2026), Kairós, Kasco AI Store, the Cuidar+ AI agent studio.
Models & tools
Claude, Gemini, ChatGPT, LangGraph, HeyGen (avatar), ElevenLabs (voice)

Legacy & Spreadsheet Modernization

COBOL, Visual Basic and spreadsheets turned into web systems, with the original behaviour tested against the new one.

Built on
My own start: COBOL on IBM mainframes with DB2 for Bradesco and Itaú, then 28 years of systems work.
Recent work
Cob.Pilot, a multi-agent pipeline that extracts the business rules from legacy COBOL and validates the translation against the original.

Data for AI, BI & Decision Intelligence

Fragmented data turned into models, pipelines and dashboards that leadership uses — and that an AI system can read.

Built on
8 years in Business Intelligence at Cognos (now IBM), leading project and development teams, plus enterprise ETL/BI projects.
Recent work
Meta Ads performance cockpit, NR-1 psychosocial risk dashboards, Monte Carlo simulation for food companies, investor signal monitoring.
Simulation
Monte Carlo simulation systems: thousands of scenarios generated within company-defined variables, then probabilistic analysis and dashboards showing the most likely outcomes.

AI Evaluation & Guardrails

The layer that makes an AI system safe to ship: anti-hallucination checks, human approval gates, objective metrics and cost limits.

Built on
2 years at ILOG (now IBM) in operations research and simulation, plus the evaluation harnesses I build for my own AI systems.
Recent work
Locked-holdout evaluation in DocBench Agents, per-comment approval queue in Reddit Copilot, human decision screens in Cob.Pilot.
Patterns
Anti-hallucination layers, LLM-provider-agnostic architectures, human-in-the-loop approval channels, cost and rate-limit guardrails.

Selected work

Architecture diagram: client applications call a gateway, which meters usage and credits before routing requests to the computer-vision services.

Kasco AI Store

Backend & frontend architectClient project

Problem
An industrial computer-vision company needed a production backend and store for its AI services.
What I did
Designed the backend and frontend architecture, a usage and credits model, and end-to-end tested flows.
Stack
Node.js, PostgreSQL, Railway, Playwright
Illustration of document parsing: a warped, scanned page with a tilted table, a formula, chemistry notation and a chart goes through a parser and comes out as clean structure - headings, a rebuilt table, the typeset formula and a redrawn chart.

DocBench Agents

ArchitectInternational competition · EMNLP 2026

2nd place overall

Dr.DocBench Challenge, DocInsights workshop at EMNLP 2026. Competed solo, as AIKHEMIST.

Invited oral presentation · HUNGEXPO Budapest · 29 October 2026

  • 85.83Overall score
  • 0.1464Text edit distance ↓
  • 70.68Tables (TEDS)
  • 76.71Formulas (CDM)
  • 86.43Reading order

Large volumes of messy data are where I have worked for 28 years; document parsing is the natural next step: turning page images into structure a system can trust.

Problem
An open international challenge in expert-level document parsing: each page image becomes one Markdown page, with text, tables as HTML, formulas as LaTeX and blocks in reading order, on the pages strong systems still get wrong.
Pipeline
Four stages, no new model trained. Existing document readers read each page, and hard pages are read again (tiny print gets two independent readings, kept only when they agree). Page-type routing picks the best reading per page and element. Rule-based post-processing repairs tables, page furniture, broken words and reading order. Automated checks validate identifiers, structure and the official format.
Agents
A LangGraph improvement loop changes the pipeline and keeps only what helps: an Analyst groups errors into patterns, a Planner ranks hypotheses, an Engineer writes the change and its tests in isolation, a Runner executes it and an Evaluator keeps or discards it page by page. An Auditor and a Packager then check and build the submission.
Measured gains
A second reader rebuilding broken tables lifted table TEDS from 64.31 to 70.00; handling page furniture lifted reading order from 78.40 to 82.91. Only public training data, split into a 677-page development set and a locked 293-page holdout.
Stack
Python, LangGraph, Claude and Gemini models, Azure AI Document Intelligence, MinerU, TeleOCR, Tesseract, PaddleOCR

View leaderboard

The Kairós site once its opening video ends: the headline over a landscape photograph, a short description and the button to start the journey.

Kairós

Architect & developerCuidar+

Problem
Career transition is slow, generic and hard to navigate.
What I did
Architected and built an AI platform grounded in behavioural science, with seven guided phases, real-time generated content, voice and avatar.
Stack
Next.js, TypeScript, Supabase, Claude API, ElevenLabs, HeyGen

Visit site

Cob.Pilot

Architect & developerOwn product, in development

Problem
Critical business rules are locked inside COBOL that nobody on the team can safely change any more.
What I did
Multi-agent pipeline on LangGraph, built against AWS CardDemo: agents read the code, extract the business rules, translate them to a modern language and validate the result against the original behaviour. A human decision screen sits before each critical step, and every agent's work is visible while it runs.
Stack
Python, LangGraph, COBOL, Next.js
The NR-1 diagnostic dashboard: responses, participation, the overall risk index and the eNPS, with a psychosocial risk radar and a bar chart comparing departments.

NR-1 Psychosocial Risk Diagnostic

ArchitectCuidar+

Problem
Companies must map and act on psychosocial risk at work under Brazilian regulation (NR-1).
What I did
Built the diagnostic and analytics dashboards, with risk bands, alerts and eNPS.
Stack
React, Node.js, Railway, analytics dashboards

Visit site

The FoodScan Pro app scoring a food: a risk score, a benefit score, the number of ingredients, and lists of risks and benefits.

FoodScan Pro

Architect & developerOwn product

Problem
Consumers cannot read what is really in their food.
What I did
Built a mobile app that scans food labels and returns risk scores, allergy alerts and healthier alternatives, based on WHO, FDA and EFSA references.
Stack
Mobile, computer vision, LLMs

Visit site

Also building

Smaller systems and internal tools, same architecture-to-production approach.

Thymos

In development — Cuidar+

Mobile-first emotional intelligence app that extends in-person workshops into 90 days of personalized practice.

Next.js, Supabase, Railway, Claude API

Athanor

Internal tool — Kairós

Free career diagnostic engine built as a product-led growth channel, with an admin panel, LGPD compliance and cost guardrails.

Next.js, Supabase, Claude API

Reddit Copilot

Internal tool — Kairós

Human-in-the-loop engagement copilot that monitors Reddit, classifies threads and drafts replies under hard rate-limit guardrails.

Node.js, Railway, Supabase, Claude API

Lapistack

In development — Own product

Connects an Instagram Business account, assembles a simulated 9-expert committee and generates a tailored landing page.

Next.js, Supabase, Instagram Graph API, Claude API

Before AI: enterprise foundation

Two decades of optimization, BI and large-scale data work for major companies across Latin America.

Optimization & Decision Systems

Petrobras · Itaú

High-cost operational decisions with many constraints.

  • Advanced route optimization for armored cash-transport vehicles
  • Oil well drilling optimization based on geographic radar data
  • Multiple logistics optimization systems
  • Supply chain

Operations Research · Optimization · Routing · Simulation

Now: Decision Intelligence

Enterprise Business Intelligence

Unilever · Pemex · Codelco · Johnson & Johnson

Leadership needed one reliable view of the business across many systems.

  • End-to-end architecture and development of BI systems
  • ETL, data cleansing and OLAP models
  • Reporting and executive dashboards

Business Intelligence · ETL · Data Cleansing · OLAP · Reporting · Dashboards

Now: Analytics Engineering

Large-Scale Data for Banking & Insurance

Banks and insurance companies

Very large data volumes with hidden patterns and no structure.

  • Processed and organized high-volume financial data
  • Pattern discovery across large datasets
  • Visualization for business and risk teams

Data Processing · Pattern Discovery · Data Visualization

Now: Advanced Analytics · Data Platforms

Enterprise Systems & Integration

Business systems that did not talk to each other.

  • Designed and built enterprise software
  • Integrated ERP and business systems

Software Architecture · ERP · Systems Integration

Now: Solutions Architecture

Brazil · Mexico · Chile · Latin America

Experience

  1. 1990–19933 years

    Programmer · HECATON and Origin C&P

    COBOL on IBM mainframes with DB2: high-volume systems for large banks (Bradesco, Itaú), plus manufacturing and retail.

    Now: Software / Solutions Architecture

  2. 1999–20078 years

    Cognos (now IBM)

    Services & Pre-sales Manager

    Business Intelligence. Hired and built teams of developers and analysts, led project and development services, and designed solutions with business directors and managers.

    Now: Analytics Engineering · Data Modeling · ELT pipelines · Decision dashboards

  3. 2007–20092 years

    ILOG (now IBM)

    Software Manager

    Operations research and simulation software

    Now: Decision Intelligence · Optimization · Prescriptive analytics

  4. 2009–20112 years

    IBM

    Software Manager

    Optimization software (ILOG) for Latin America

    Now: Enterprise Data & Analytics Architecture

  5. 2012–2020

    Independent BI & data consultant

    Monte Carlo simulations that guided food companies on how to operate and what to sell; BI systems, data warehouses and ERP integration for corporate and local clients.

2020 — present: three parallel tracks

AIKHEMIST

2020–present

AI repositioning since 2021: AI systems architecture and development. Own products and R&D.

Cuidar+

2025–present

Head of Technology

  • Kairós in production since January 2026
  • Thymos in development

Client work

2026

Kasco AI Store — backend & frontend architecture, contract

MBA, FGV

Enterprise systems delivered for companies such as Petrobras · Itaú · Unilever · Pemex · Codelco · Johnson & Johnson · Cargill

Projects across Latin America

Stack

Data & BI

  • SQL
  • ETL/ELT
  • Relational Data Modeling
  • Dimensional Modeling (star schema)
  • IBM Cognos
  • IBM DB2
  • Dashboards

Decision science

  • Operations Research
  • Optimization
  • Monte Carlo Simulation
  • Probabilistic Analysis
  • Scoring
  • Forecasting

AI

  • LLM Applications
  • Claude API
  • Gemini
  • ChatGPT
  • Agents
  • RAG
  • Document AI
  • Document Parsing
  • OCR
  • HeyGen
  • ElevenLabs

Application

  • TypeScript
  • Next.js
  • React
  • Node.js
  • Python

Data platforms & cloud

  • PostgreSQL
  • Supabase
  • Firebase
  • Railway
  • Vercel

Quality

  • Playwright End-to-End Testing

How I work

  • From business problem to working system, without depending on a finished technical spec.
  • Remote-first, async-friendly, overlap with US and Europe hours
  • Full-time, contract or ongoing part-time
  • End-to-end ownership, from architecture to production
  • Clear written communication in English, Spanish and Portuguese

Outside work In 2011 I reached the summit of Aconcagua, 6,961 m, the highest mountain outside the Himalayas. Fifteen days on the mountain. We started as a group of 21; two of us made the top. The last two days were a blizzard: three metres of visibility, −15 °C. I don't quit.

Raw data is lead. Clear decisions are gold.

Have a data or AI challenge? Let's talk.

Download CV