Renan – Python, LLM, RAG

It's an enterprise AI platform that automates the creation, review, and approval of complex data-governance and compliance plans. The workflow evaluates how projects use data across privacy, jurisdictional, legal, ethical, and operational risk dimensions.

• Implemented automated deterministic risk detection that evaluates a curated risk library against all assessment responses.
• Designed the risk-suggestion flow so that automatically identified risks remained explainable and traceable to explicit conditions.
• Added a human-review stage allowing accountable reviewers to approve or reject automatically suggested risks before finalization.
• Built document-upload and processing for PDF, DOCX, and PPTX so the AI copilot could use project-specific materials as context.
• Corrected aggregate risk-rating logic so category-level results reflected the highest applicable factor instead of masking severe risks.
• Evaluated a newer LLM generation through a structured migration spike and smoke-tested existing AI-assisted features.
• Identified output-quality regressions and supported a safe rollback rather than promoting a model that did not meet product expectations.

It's an end-to-end NLP and semantic search solution that transforms unstructured vehicle warranty feedback into actionable product intelligence.

• Built an end-to-end NLP pipeline that transformed unstructured warranty claims into structured sentiment, feature, ownership, and issue-analysis signals.
• Consumed curated warranty and product-reference data from BigQuery and prepared the text fields required for downstream language processing.
• Implemented sentiment analysis to distinguish positive feedback from complaints and prioritize the cases most relevant to product-quality investigation.
• Designed semantic-search logic that compared customer-described symptoms with functional feature descriptions rather than depending on exact feature names or keywords.
• Ranked semantically similar feature candidates and mapped the selected result to the appropriate internal feature owner for routing and investigation.
• Handled the vocabulary gap between non-technical customer language and the engineering terminology used in the vehicle-feature catalog.
• Aggregated classified claims to identify features with recurring complaints and connect issue frequency with warranty-related cost analysis.
• Deployed and operated the solution on Google Cloud Platform, taking ownership of the workflow beyond model development.
• Defined repeatable cloud infrastructure with Terraform and maintained the monitoring required to observe the deployed pipeline.
• Protected confidential product and customer information by keeping internal feature names, ownership structures, claim content, volumes, and cost values outside public documentation.

It's a document-intelligence solution for a healthcare reimbursement workflow.

• Developed the first computer-vision binary classifier to distinguish medical requests from unrelated image submissions and analyzed why visual appearance alone did not generalize across handwritten, printed, photographed, and scanned documents.
• Investigated an accuracy ceiling of approximately 40% and traced the main business risk to false negatives, where valid medical requests could be incorrectly rejected.
• Built an OCR layer for both images and PDF documents and transformed extracted text into signals suitable for downstream classification.
• Designed a domain-aware keyword strategy using terms and expressions characteristic of medical examination and procedure requests.
• Integrated an LLM-based classification stage that evaluated extracted content and document context, increasing final accuracy to 92% and materially reducing false negatives.
• Used Amazon EC2 for the training workloads in the first version, including parallel model experiments and optimization with Adam.
• Used Amazon SageMaker and Bedrock for managed machine-learning and generative-AI capabilities within the revised AWS solution.
• Implemented event-driven processing with Amazon S3, API Gateway, and Lambda, and containerized the application with Docker.
• Orchestrated application containers with Amazon ECS and supported repeatable delivery through CI/CD practices.

It's a fleet analytics solution that transforms high-volume vehicle telemetry into an explainable score to drive efficiency and fuel-conscious behavior.

• Built Python and SQL data pipelines that extracted raw telemetry from MySQL, standardized inconsistent signals, filtered unusable records, and stored analysis-ready data in a cleaner downstream table.
• Investigated missing, delayed, and anomalous telemetry caused by tunnels, connectivity loss, signal blockers, and jammers, and added validation rules so low-quality data would not distort driver scores.
• Performed exploratory and comparative analysis across routes, vehicle models, weight classes, and operating conditions to establish realistic behavioral baselines.
• Mapped manufacturer specifications and vehicle manuals into model-specific operating ranges, including economical RPM, torque, power, and engine-braking bands.
• Designed an interpretable scoring strategy that rewarded coasting with the vehicle engaged and progressive braking, and penalized prolonged idling, harsh braking, extended speeding, excessive RPM, and inefficient engine operation.
• Normalized driving events by distance and compared vehicles against relevant peer groups, reducing bias introduced by different routes, vehicle categories, and usage profiles.
• Created analytical dashboards and reporting artifacts with Plotly, Dash, Power BI, and automated PDF outputs to make technical findings understandable to non-technical stakeholders.
• Presented data narratives that demonstrated at least 10% potential monthly savings in fuel and maintenance through improved driving habits.