Eric de Oliveira Souza
Sr. Full Stack & AI Engineer
Owning and developing products from architecture to production for 8+ years. I have experience in industries like legal-tech, healthcare, fintech, e-commerce, AI, and others.
I'm fluent in both web and AI. I build everything from complex UIs to integrated REST APIs with React, Node and Python, and AI features from document processing and image generation to prompt optimization, using LLMs and RAG with OpenAI, Anthropic and Gemini.
I also own what surrounds the code: system design, CI/CD and cloud infrastructure on AWS and GCP. I've led teams, set code and delivery standards, and I work directly with clients to turn what they need into features that ship.
Belo Horizonte, Brazil·GMT-3·I work on flexible US business hours
// experience
- Worked at EXIT83 for its client Remedy Product Studio, a product consultancy, building healthcare software for patient and EEG data tracking for their client Wave Neuroscience.
- Built UIs with React, Next, and Material UI for efficient patient and EEG data tracking.
- Developed scalable backend systems with Node, Nest, PostgreSQL, and DynamoDB, deployed on AWS.
- Integrated medical APIs for TMS session validation and EEG data automation, increasing accuracy.
- Optimized EEG and treatment workflows, reducing admin overhead and improving patient management.
- Improved performance with optimized queries and Redis caching, cutting report processing time.
- Developed and scaled the backend of the main financial app using Node, Nest, and PostgreSQL, implementing microservices with Apache Kafka to improve performance, reduce latency, and enhance scalability for high-volume transactions.
- Used Sequelize ORM to simplify database work, cut down on awkward queries, and speed up API responses.
- Built a financial remittance and client management system with React, Next, and Material UI on the frontend and Node, Nest, and PostgreSQL on the backend.
- Automated reporting workflows by generating Jira-based reports in Google Sheets through a Node/Express integration.
- Built a web platform with AI heatmaps so clients could compare locations using real-time data.
- Built the full stack with React, Next, and Material UI on the frontend and Node, Nest, and PostgreSQL on the backend.
- Improved how the AI R&D team's data flowed into the web platform and stored large geospatial files on Azure Blob Storage, with DevOps on Azure.
- Developed and maintained an e-commerce platform for an internet and telephone service provider in South America, improving the online purchase experience with a React, Next, and Material UI frontend and Node, Express, and PostgreSQL backend.
- Built a customer and plan management system for listing users and viewing plan details, with React, Next, and Bootstrap on the frontend and Node, Express, and PostgreSQL on the backend.
// projects
Client work · AI analytics · 2026–present
Enterprise AI Impact Analytics Platform
Helped enterprises see what AI costs, what teams produce with it and which workflows to automate next, with usage data they can pull into their own BI tools without exposing any one employee.
Sr. Full Stack & AI Engineer · customer-facing platform: API, web app, analytics, identity sync and infrastructure
The challenge
An analytics platform for AI-forward enterprises tracks AI spend, adoption, developer productivity and automatable workflows for Heads of AI, CTOs, engineering leaders and finance teams. Customers wanted that data in their own BI tools and their org charts synced from their identity providers, without the data singling out individual employees.
What we achieved
Customers can pull usage data into their own BI tools through new public API endpoints, with customer-facing docs.
Individual-level data stays protected: admins set a minimum group size before individuals are shown, and totals and filters can't single out a person.
Every per-user endpoint follows one shared privacy policy, and every public route was audited against it.
Org charts sync from Okta, Entra ID, JumpCloud and Google Workspace with new employee attributes, and customers can enrich them without breaking the sync.
Worked directly with customers' IT teams to debug their identity provider integrations.
New and extended reporting marts are tested to reconcile with the existing reports.
Client work · Marketing · 2026–present
Amazon Listing Audit & AI Copy Platform
Gave a marketing agency a tool that audits whole Amazon catalogs and drafts better, evidence-backed listing copy for each marketplace and language.
Full Stack & AI Engineer · designed and built the application
The challenge
The agency's teams reviewed Amazon product pages one at a time, by hand, across Amazon marketplaces. They needed a fast way to see which pages underperform across a whole catalog, why, and what better copy would look like, with evidence they could defend to brands.
What we achieved
Audits catalogs of up to 100 products against 34 checks, from Amazon basics (lengths, counts, images, ratings) to effectiveness (keywords, claims, benefits, buyer language).
Pulls recurring strengths, concerns and buyer language out of customer reviews.
Drafts a new title, bullets, Item Highlights and description for each product, with the reason for each change and the evidence behind it.
Adds a page blueprint and a hero image prompt to brief the creative team.
Operators can change the marketplace, the marketplace locale and the app language, and the checks and generated copy follow.
Shows portfolio health across every client, and exports recommendations to CSV and Excel.
Client work · Legal-tech · 2025–present
Essential Estates, 0 to Production
Turned a law firm's manual estate planning service into a production software platform, as its first engineer.
Founding Full Stack & AI Engineer at Summerall Law, through TeamEx · owned product, engineering, hiring, and infrastructure
The challenge
Summerall Law, a Silicon Valley law firm, ran estate planning workflows by hand. There was no product, no engineering team, and no infrastructure, and they needed a platform clients could rely on in production.
What we achieved
Launched Essential Estates to production, built from zero.
Brought AI document processing into the client workflow, replacing manual handling of legal documents.
Hired, onboarded, and led the engineering team, with coding standards, code review, and a clear delivery process.
Releases ship through automated CI/CD instead of manual deploys.
Gave the firm product analytics (funnels and usage metrics) to guide product and marketing decisions.
Side project · Health-tech · Updated Oct 2026
Voice Receptionist Copilot
A voice receptionist for a dental clinic that books, moves and cancels real appointments over a phone-style call, and that you can interrupt mid-sentence like a person.
Solo build · voice pipeline, agent and tools, scheduling, dashboard, and tests
The goal
Most voice demos are turn-based: the bot talks until it finishes, and anything you say meanwhile is lost. A receptionist has to be interruptible, answer fast enough to feel like a call, and never book a slot the caller did not actually hear read back. The goal was a full speech-to-speech loop on free tiers and a CPU, where talking over Sofia stops her within a third of a second, her memory keeps only what you heard, and every booking still matches the database.
What it does
Talk to Sofia in the browser: she checks availability, books, reschedules and cancels appointments, and answers questions about prices, insurance, hours and parking for a fictional clinic in Austin, Texas.
Talking over her for 300 ms stops her and drops the audio already queued; short sounds like "uh-huh" do not, and the history keeps only the sentences the caller heard, marked as cut off.
Every change is proposed, read back, and only applied after the caller says yes on a later turn, so a change the caller cut off is withdrawn instead of booked.
Per-turn latency logging brought the time from the end of speech to the first audio from 4.8 seconds on the first live turn down to 1.7 to 2.0 seconds.
A React dashboard shows the live transcript, each tool call with its arguments and result, a latency bar per turn, and the week's bookings, which jump to and highlight each change Sofia confirms.
Side project · Fintech · Updated Oct 2026
Doc Extraction Copilot
Reads a photo or PDF of a Brazilian boleto, nota fiscal or income statement into typed JSON, checks every field it can in code, and asks a person to review only the fields that failed.
Solo build · extraction, checks, review UI, eval set, and tests
The goal
Boletos, notas fiscais and informes de rendimentos arrive as PDFs and phone photos, and typing them in is slow. A vision model reads them well but can still drop a digit, and asking it to check its own answer proves little. These documents carry their own check digits and totals, so the goal was to let the model read, let code check, show a person only the fields that need a look, and measure with an eval set how many wrong fields would get through unseen.
What it does
Upload a boleto, a DANFE or a broker's informe de rendimentos as a PNG, JPEG or PDF, and get every field as typed JSON for that document type.
Each field is marked checked, review or not checked, so a person only looks at the fields that failed a check or that the model was unsure of.
A boleto whose printed amount does not match its digit line, or a DANFE line total that is off by R$ 10, is flagged on the field that is actually wrong.
The JSON can only be downloaded once nothing is left to review.
On 36 synthetic documents, the default model read 99.9% of fields right, flagged with 100% precision and recall, left 0 silent errors and caught all 6 planted inconsistencies, at a median of 5.4s a document.
Side project · Fintech · Updated Oct 2026
Renda Fixa Copilot
A daily email that finds the best Brazilian CDBs on the web, checks every rate against its source, and shows what R$ 75k, 100k and 150k would earn after tax in each.
Solo build · search, verification, simulation, email, and scheduling
The goal
Bank CDB, LCI and LCA rates have no public API: they live on bank apps, broker pages and daily ranking posts, and change every day. A model asked for them without search makes them up. The goal was a daily email that uses a model only to find and quote offers, then checks each quote and does all the money math in code, so every number in the email can be traced to a source or a formula.
What it does
Every weekday at 08:00 Brasília time, an email ranks the best CDBs by net return for up to 1 year, 1 to 3, 3 to 6 and over 6 years, then compares them with Tesouro Direto, LCIs and LCAs.
Each offer shows the rate, maturity, net return a year against a CDB at 100% of CDI, and the net profit for R$ 75k, 100k and 150k, plus the app or broker to invest through.
Offers from the web are kept only when the page the search returned quotes the rate verbatim; the rest are listed as left out, with the reason.
Warns when an amount would pass the FGC guarantee of R$ 250k per bank at maturity, interest included, and when a Tesouro bond sold early can lose money.
A responsive HTML email that stacks on a phone, with a plain-text version for clients that drop the styles.
// skills
Frontend
Responsive, accessible UIs, from component libraries to full product front ends.
Backend
APIs and services built for clear contracts, reliability and scale.
AI Engineering
LLM integrations (OpenAI, Anthropic, Gemini, Bedrock, Groq), agents, RAG, MCP, evals, vision and voice models, document processing and image generation.
AI Tools
AI in my daily workflow, from coding and review to planning.
Cloud & DevOps
Cloud infrastructure, containers and CI/CD for reliable deployments.
Architecture
Systems designed to scale, from service boundaries to event-driven flows.
Data
Relational and NoSQL data modeling, ORMs, caching and analytics.
Testing & Quality
Unit, integration and end-to-end tests, plus thorough code review.
Tools & Practices
Version control, project tracking, analytics and agile delivery.