Daniel Scheidemantel Camargo

Capellaris AI (SaaS, agentic AI platform for SMBs)

November 2024 — September 2025

• Designed and built an LLM-agnostic multi-tenant agentic SaaS platform end-to-end as sole engineer (business partner owned commercial, financial and business processes), enabling SMBs to deploy specialised AI agents across business domains; also owned post-sales and customer training.
• Built a full RAG pipeline: document ingestion, chunking, embedding generation, pgvector storage, hybrid search, reranking and source-grounded answers, reused across all agent domains.
• Implemented a hierarchical multi-agent orchestration layer (manager agent + specialists; CrewAI + MCP) with tool-calling and memory management.
• Designed a two-tier agent memory: long-term RAG over business rules, the business plan and previously well-rated answer formats (appended to the vector DB), plus dynamic short-term memory where an automatic cyclic summary of the chat and prior context is written at 60% of the agent’s context window, prioritising newer context.
• Tracked a 1–5 star user rating for every project delivered by the manager agent, from the first single-agent deliveries through hierarchical multi-agent runs: later projects scored higher and needed fewer refinement rounds.
• Piloted with three small businesses (an independent consultancy, a restaurant and an art atelier) plus internal use.
• Delivered financial control agents that gave a pilot customer a unified view of financial planning across departments, replacing fragmented manual processes.
• Implemented the MLOps lifecycle: model integration, prompt versioning, evaluation, observability and CI/CD with GitHub Actions.
• Full-stack delivery: React front-end with a Node.js backend-for-frontend (BFF).
• Infrastructure: GCP (Cloud Run, Firebase), Supabase PostgreSQL + pgvector, Cloudflare; OAuth 2.0, IAM, multi-tenant secrets management.

Mercado Livre (logistics network & facility simulators, ~5k internal users, LATAM)

August 2025 — August 2026

• Integrated OpenAI and Claude APIs as decision-support agents in a production simulation platform used by ~5k supply-chain specialists across LATAM: agents validate user-entered parameters and suggest corrections (~25% fewer input errors vs. before the agents) and interpret computed KPIs by comparing a summary of the full dataset against the desired behaviour.
• Evaluated the agents on p95 latency under 30s, user-rated quality (4.6/5 average) and token consumption per user; owned the lifecycle from prompt design and versioning through deployment and iteration.
• Redesigned the simulation data-input flow, splitting one large spreadsheet into smaller sequential steps, reducing input errors by 17%.
• Built data pipelines over terabytes in BigQuery: chained queries and transformations that break down real historical demand and apply leadership growth forecasts by category, region and facility type, producing demand projections over an adjusted logistics network.
• Designed the async FastAPI layer between frontend and backend: receives user parameters, handles database operations and job status without blocking other requests, so the application keeps scaling.
• Moved data processing from Pandas to Polars, absorbing a 2x increase in input rows from user-submitted Excel files.
• Used Dataiku for transformations holding business rules owned by domain specialists, so rules could be edited without a code deploy.
• Mentored 2 mid-level engineers in an 8-person team; one was promoted to senior.
• Maintained CI/CD pipelines (GitHub Actions) and observability (Datadog, Grafana); added new metrics through new endpoints.
• Stack: Python, FastAPI, BigQuery, MySQL, Dataiku, Docker, React (front-end); limited exposure to Go (edited one microservice).
• Placed 4th of 22 teams in an internal AI hackathon (02/2026, team of 6), using n8n to evaluate operational metrics.

Buscaônibus / DeÔnibus / Travelier (travel meta-search and BI platform)

April 2021 — January 2025

• Sole technical owner of a high-traffic Python backend processing millions of daily searches across hundreds of bus operators (on-call, incident response, full production lifecycle); platform ranked 4th-largest in Brazil by web traffic (SEMrush).
• Led a complete platform migration (Python 2.7 to 3.10, Django 1.9 to 5, App Engine to Cloud Run) in well-defined stages, using feature flags backed by cached keys for immediate rollback; new complex APIs on Cloud Run were built in FastAPI and specific microservices as Flask Cloud Functions. Average API latency −25% and cloud cost −40% (mean before vs. after the migration).
• Redesigned the Redis caching layer: all database and partner data cached with per-source TTLs, structured key management, and a Redis manager to inspect and manipulate keys. Reached a 92% cache hit ratio (GCP Redis metric, 30-day average).
• Built on-demand scrapers (Python, Playwright, Selenium, Beautiful Soup on Cloud Run/Docker) for 5 small bus companies without digital integration, and built a monitoring dashboard that surfaced when a company stopped returning data.
• Built a Gemini RAG agent with engineered prompting to automate triage and responses for customer emails from the website contact forms, with responses rated 89% accurate on average by the support specialist who handled these emails.
• Designed a BigQuery analytics layer delivered through Plotly Dash dashboards (Pandas/NumPy) and Looker Studio views: market, competitor and route-coverage intelligence for partner bus companies’ strategic decisions, plus an internal sales BI for the commercial team.
• Mentored a junior engineer (a content copywriter turned frontend developer) toward full-stack on seniority, project management, ticket lifecycle and pull-request practices.

Portobello Group (ceramics manufacturing & retail, ~4,000 employees)

December 2019 — April 2021

• Built and maintained a regional fiscal discount simulator (Python/Django on Heroku with PostgreSQL) that calculates the full tax burden of lists of complex, diverse products, integrated into Salesforce (Apex API layer and UI button) and rendered there as HTML served by the service; worked with fiscal stakeholders to raise fiscal-rule coverage from 70% to 96%.
• Part of the team that delivered the company’s 2021 digital transformation, including the CRM/ERP migration to Salesforce across 124 retail stores (mostly franchises), with no integration-related incident tickets (only usage questions); designed the product data model and the product-to-order connection.
• Administered the Microsoft SQL Server database that held company-wide data and integrated multiple ERPs and CRMs (ETLs, transformations) ahead of the Salesforce migration, hosted on an AWS EC2 virtual machine.
• Worked across Microsoft Azure, where every new service was a microservice (serverless functions and a data pipeline), and AWS (SQL Server on EC2).
• Operated Kubernetes (Rancher) and Jenkins in an API-management integration epic with attached microservices.

Stormtech (consumer loans, regulated)

August 2018 — November 2019

• Developed RPA bots (Python, Scrapy, Selenium, Beautiful Soup) that scrape Brazilian loan-bank portals to update loan status automatically in an internal PHP loan-management system; wrote about half of the bots and maintained the rest, updating them whenever a bank changed its contract or portal.
• Implemented a TypeScript dashboard with real-time visibility of bot health (running, failed, unused), cutting the time to identify a broken pipeline from about 2 hours to under 1 hour. The dashboard also showed stability per bank and alerts.
• Automated roughly half of the operators’ manual workload by adding bots for new banks.
• Ran the bots mostly on AWS Lambda (EC2 for those requiring Windows), with Step Functions-orchestrated Lambda ETL feeding DynamoDB (bank update data) and MySQL (client data and security layer), across 18 integrated banks.

ReciclaComigo (startup, cloud-connected reverse vending machine, beta stage)

July 2016 — February 2018

• Developed the edge software of a beta reverse vending machine for PET bottles and aluminum cans: a conveyor brings each item to a camera where Python + OpenCV reads the barcode, shape and label; after validating product and user against the database, the machine credits points to the user.
• Wrote the Arduino (C++) firmware for motors and presence sensing and the Raspberry Pi Python and Shell scripts (GPIO automation) that tie the microcontroller, an RFID reader/keypad and the cloud together.
• Designed and configured the company’s AWS environment (EC2 on Linux, RDS MySQL, S3, security groups, IAM roles) accessed over Wi-Fi/HTTP by the machine; managed the developer who designed the database schema and built the HTML/JavaScript/CSS front-end.
• Took part in startup and product definition applying Lean Startup practices (MVP, Design Thinking, business model canvas, competitor analysis, business plan led by the CEO). Team of five; single unit tested in a controlled environment; the company closed before public launch.
• Trello for task management and Bitbucket for version control.

UDESC: Colmeia & LabP2D (research)

March 2012 — March 2017

• Developed energy monitoring solutions for data centers (GreenHop project).
• Created algorithms for virtual machine consolidation (CHAVE project).
• Published 13 peer-reviewed research papers (journals and conferences).

25 teams, 36h, live demos to investors (prize: cash + 1-week residency at the Swiss Data Science Center)

May 2026

• Co-built a simulated 256-rack datacenter where three specialist LLM agents (thermal, workload, energy; Pydantic AI + OpenAI) propose actions in sequence and a deterministic resolver enforces SLA and safety rules.
• Owned datacenter requirements, PUE/cost metrics and simulator realism, plus the manager-facing view (Phaser, TypeScript); 39% lower energy vs. a no-agent baseline in a simulated 10-min failure scenario.

UDESC — Colmeia & LabP2D

2013 – 2016

• Taught introductory and advanced Arduino programming in C++ to undergraduate students.
• Covered embedded systems concepts, sensor integration, and hardware automation.
• Applied in research projects such as GreenHop (data center environmental monitoring).

UDESC — Colmeia & LabP2D

2013 – 2016

• Taught Git fundamentals and collaborative workflows to research lab members.
• Topics included branching strategies, pull requests, and open source contribution practices.

UDESC — Colmeia & LabP2D

2013 – 2016

• Instructed students on LaTeX typesetting for academic papers, theses, and conference submissions.
• Covered BibTeX references, mathematical notation, and IEEE/ACM formatting templates.