Caio is available for hire
Hire CaioCaio Theodoro
Verified Expert in Engineering
Software Engineer and Developer
Curitiba - State of Paraná, Brazil
Toptal member since August 19, 2026
Bio
For the past seven years, Caio has built agent infrastructure and production LLM systems. He has worked extensively across fintech, education, and supply-chain industries. His work centers on Python, TypeScript, MLops, and AWS for clients including Adopt AI, Avenza, and Acorns.
Portfolio
Adopt AI
Temporal, Agent Orchestration, AI Engineering, AI Architecture, AI Guardrails...
Avenza
Agent Orchestration, AI Engineering, AI Architecture, Amazon Web Services (AWS)...
Acorns
Rails, Node.js, React, Next.js, GraphQL, Software Engineering, Microservices...
Experience
- Microservices - 7 years
- TypeScript - 7 years
- Software Engineering - 7 years
- AI Architecture - 3 years
- Retrieval-augmented Generation (RAG) - 3 years
- LLM Evaluation - 3 years
- AI Guardrails - 2 years
- AI Engineering - 2 years
Work Experience
Senior AI Engineer
2026 - PRESENT
Adopt AI
- Architected the core agent execution harness, implementing turn and message lifecycles, idempotent claim semantics, and replay-safe Temporal orchestration to support all agent capabilities on the platform.
- Built the live agent evaluation dashboard that scores conversation quality in production, surfacing degradation early enough for the team to intervene before customers feel it.
- Shipped the human-in-the-loop intervention system end to end, including a signal-driven resolve state machine with optimistic locking and tenant isolation, a 3-action resolution model, and a Temporal signal contract for escalations.
- Designed the generative UI architecture for tax and accounting workflows using a registry of approved components, reducing dashboard prompts and saving 90% on multi-turn follow-up edits with a JSON-Patch delta.
- Built a policy engine for agent guardrails with policy-version pinning and full audit logging, and integrated AWS Bedrock for multi-model inference behind one selection interface, cutting latency and cost through provider routing and prompt caching.
Technologies: Temporal, Agent Orchestration, AI Engineering, AI Architecture, AI Guardrails, Amazon S3 (AWS S3), Software Engineering, Machine Learning, Retrieval-augmented Generation (RAG), Microservices, TypeScript, LLM Evaluation, LLM Workflows, Multi-agent Systems, Tool Calling, Model Context Protocol (MCP), LangGraph, Prompt Engineering, Context Engineering, Observability, CI/CD Pipelines, REST, Temporal Workflows, NumPy, Pandas, TensorFlow, PyTorch, Machine Learning Operations (MLOps), Inference Optimization, Reinforcement Learning from Human Feedback (RLHF), LLM-as-Judge, Embeddings & Semantic Search, Vector Search, LangChain, Responsible AI, Explainability, Calibration, Quantization, OpenAI, Google Gemini, Hugging Face, Natural Language Processing (NLP)
Founder and Principal Engineer
2025 - PRESENT
Avenza
- Directed a 7-engineer team across client discovery, solution architecture, delivery, and engineering standards, delivering systems tied to roughly $2.8 million in client revenue over eight months.
- Architected production AI agents across logistics, customer support, operational workflows, and computer vision using structured outputs, retrieval, model-backed classification, and human review loops.
- Delivered end-to-end software and automation for a foreign-exchange house covering operational workflows, back-office processing, and third-party system integrations.
Technologies: Agent Orchestration, AI Engineering, AI Architecture, Amazon Web Services (AWS), AI Guardrails, Python, TypeScript, Software Engineering, Machine Learning, Retrieval-augmented Generation (RAG), Microservices, LLM Evaluation, LLM Workflows, Multi-agent Systems, Tool Calling, Model Context Protocol (MCP), LangGraph, Prompt Engineering, Context Engineering, Observability, CI/CD Pipelines, REST, Temporal Workflows, Go, NumPy, Pandas, TensorFlow, PyTorch, Machine Learning Operations (MLOps), Inference Optimization, Reinforcement Learning from Human Feedback (RLHF), LLM-as-Judge, Embeddings & Semantic Search, Vector Search, LangChain, Responsible AI, Calibration, Natural Language Processing (NLP)
Senior Software Engineer
2025 - 2026
Acorns
- Shipped production changes across the full stack (Rails, Node services, React, Next.js, GraphQL APIs) throughout a phased monolith-to-microservices migration.
- Built GraphQL-driven admin and support tooling that gave internal operations direct visibility into servicing workflows and resolved schema and data-consistency failures in the gateway introduced by splitting Rails-owned data across newly indep.
- Shipped ephemeral per-PR environments to production, closing the staging/production parity gap that had been a recurring source of regressions.
Technologies: Rails, Node.js, React, Next.js, GraphQL, Software Engineering, Microservices, TypeScript, Tool Calling, Model Context Protocol (MCP), LangGraph, Prompt Engineering, Observability, CI/CD Pipelines, REST, Kafka, Go, NumPy, Pandas, TensorFlow, PyTorch, Vector Search
Software Engineer
2022 - 2025
MB Labs
- Built production systems on AWS (ECS, RDS, ElastiCache, S3, CloudFront, ALB) with Cloudflare and Redis for low-latency fintech workloads serving millions of users.
- Designed microservices and APIs for high-traffic fintech clients; connection pooling and regional caching cut latency while right-sizing compute brought infrastructure spend down.
- Architected event-driven data pipelines (SQS, async workers) for compliance-grade fintech operations.
Technologies: Amazon Web Services (AWS), Cloudflare, Redis, Amazon Simple Queue Service (SQS), Software Engineering, Microservices, TypeScript, Prompt Engineering, Observability, CI/CD Pipelines, REST, Kafka, Rust, Go, NumPy, Pandas, TensorFlow, Machine Learning Operations (MLOps), Recommendation Systems
Software Engineer
2021 - 2022
Mactus Informática LTDA
- Modernized legacy business applications and improved stability for critical operational workflows.
- Designed cloud-ready services with clearer performance, maintenance, and cost boundaries.
- Led new products discovery and implementation inside the company.
Technologies: JavaScript, TypeScript, Python, Software Engineering, Observability, CI/CD Pipelines, REST, Recommendation Systems
Software Engineer Intern
2020 - 2021
Atla
- Built Python scrapers processing 5,000+ records per day and REST APIs for internal education data workflows.
- Built backoffice softwares that empowered the core main product.
- Implemented Agile methods inside the squad to improve efficiency.
Technologies: Python, Software Engineering, TypeScript, Scraping, REST, Recommendation Systems
Software Engineer
2019 - 2021
Haken
- Delivered web and mobile software for early clients while owning project scope, delivery plans, and client communication.
- Established effective communication channels between technical teams and stakeholders.
- Implemented RESTful APIs to streamline data processing workflows.
Technologies: JavaScript, TypeScript, Python, Software Engineering, REST
Experience
ReconForge
https://github.com/caiotheodoro/reconforgeI fine-tuned a 1.7-billion-parameter model with LoRA on a laptop to detect financial reconciliation exceptions. I optimized for severity-weighted recall, the metric tied to dollar risk, reaching 0.913 versus 0.872 for a frontier model, with 100% recall on high-severity exceptions and zero API cost. I built the event pipeline and reconciliation state on Kafka, Temporal, Postgres, and Neo4j, so exceptions are caught, orchestrated, and traced end-to-end without relying on a hosted LLM API.
LossBench
https://github.com/caiotheodoro/lossbenchI built an evaluation and control plane that scores AI agents on expected operational loss instead of raw accuracy: severity-weighted, calibrated, and replayable. Every decision runs through a record-calibrate-decide-escalate-replay pipeline backed by a hash-chained decision ledger, so outcomes can be audited and reproduced after the fact. Ships with LangGraph and DeepSeek Harness adapters, making it drop-in for agent stacks that already use those frameworks and need loss-aware evaluation rather than accuracy alone.
Calibrated Evaluation Foundry
https://github.com/caiotheodoro/substrateI reproduced ARC-AGI-3's benchmark methodology as a repeatable pipeline, then used it to stress-test its own parameters against real solvers and a real LLM judge rather than trusting the published methodology at face value. The goal was to find where the benchmark's assumptions break down: which parameter choices meaningfully change solver rankings, and where an LLM judge disagrees with or diverges from solver-based scoring. It was built as a foundry so the calibration pipeline can be rerun against new solvers or judges as they change.
Education
2018 - 2025
Bachelor's Degree in Computer Science
Technological Federal University of Paraná - Brazil
Certifications
SEPTEMBER 2024 - PRESENT
Cloud Technical Essentials
Amazon Web Services (AWS)
JANUARY 2024 - PRESENT
DevOps and Software Engineering
IBM
JANUARY 2024 - PRESENT
Machine Learning
IBM
JANUARY 2024 - PRESENT
AWS Fundamentals
Amazon Web Services (AWS)
JANUARY 2024 - PRESENT
Cloud Solutions Architect
Amazon Web Services (AWS)
AUGUST 2023 - PRESENT
Full-stack Software Developer
IBM
AUGUST 2023 - PRESENT
Data Analytics
Skills
Libraries/APIs
PyTorch, TensorFlow, Pandas, NumPy, Node.js, React, OpenCV
Tools
GraphRAG, Amazon Simple Queue Service (SQS), Terraform
Languages
Python, TypeScript, JavaScript, SQL, GraphQL, Go, Rust
Paradigms
Microservices, REST, Automation, Model Context Protocol (MCP)
Platforms
Harness, Docker, Apache Kafka, Cloud Native, Amazon Web Services (AWS), Azure, Kubernetes
Frameworks
LangGraph, Next.js
Storage
PostgreSQL, Neo4j, Amazon S3 (AWS S3), Redis
Other
AI Engineering, AI Architecture, Machine Learning, LLM Workflows, Agent Orchestration, Structured Outputs, AI Guardrails, Retrieval-augmented Generation (RAG), Embeddings & Semantic Search, Vector Search, Inference Optimization, Temporal Workflows, Multi-agent Systems, Tool Calling, LangChain, Prompt Engineering, Machine Learning Operations (MLOps), OpenAI, Computer Vision, Natural Language Processing (NLP), Kafka, CI/CD Pipelines, Software Engineering, Scraping, LoRa, LLM Evaluation, Data Cleansing, Cloud Computing, Temporal, Rails, Cloudflare, GCP, Context Engineering, Responsible AI, Explainability, LLM-as-Judge, Calibration, Reinforcement Learning from Human Feedback (RLHF), Quantization, Google Gemini, Hugging Face, Predictive Analytics, Recommendation Systems, Observability, Infrastructure as a Service (IaaS)
Collaboration That Works
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
1
Share your needs
Discuss your requirements and refine your scope in a call with a Toptal domain expert.
2
Choose your talent
Get a short list of expertly matched talent within 24 hours to review, interview, and choose from.
3
Start your risk-free talent trial
Work with your chosen talent on a trial basis for up to two weeks. Pay only if you decide to hire them.
Top talent is in high demand.
Start hiring