Work

The challenge. A legal consulting firm manages document-heavy case files for financial institutions. Each case begins with PDFs and unstructured documents that must be analyzed, validated, and turned into calculations and formal reports. The process depended on many manual tasks, was difficult to measure, consumed a great deal of time, and required human review at every stage.

My work. I took responsibility for product definition, technical architecture, and construction of the platform. I designed a system of specialized agents that extract information, prepare calculations, and generate documents. Tasks run in the background and stop at review points where a person validates the result before the case moves forward.

The platform also manages documents, the state of each case, different client workflows, and decision traceability. The system combines automation with a complete application through which professionals can review and control the work performed by the AI.

Result. The platform is in production with real client workloads. The process has moved from a sequence of manual tasks to a traceable workflow in which every case has a visible state, AI outputs can be reviewed, and decisions are recorded. This has reduced manual work from weeks to hours without delegating the final decision to the system.

Technologies. Google ADK, Gemini, React, FastAPI, Firestore, Cloud Storage, Cloud Run, and asynchronous queue-based processing.

The challenge. A poker education business provides tools for learning the game and has thousands of courses, videos, transcripts, documents, images, and notes containing highly specialized knowledge. The content was spread across different formats, and the business needed a way to support students using that material.

My work. I designed the knowledge processing and retrieval architecture. The system transforms PDFs, images, audio, and video through several AI stages that normalize the content and extract domain-specific metadata such as game type, concepts, format, and difficulty.

On top of that foundation, I built an agentic RAG system with chunking, vector search, hybrid search, keyword search, query expansion, and result reranking. The product includes a chat experience that answers with citations to the source material, along with tools for managing documents and exploring the knowledge base.

Result. The system is in production and makes it possible to query a multimodal corpus through answers linked to curated content. The business can add and organize new material without requiring manual application changes. Retrieval and responses include observability for evaluating quality, detecting failures, and analyzing costs.

Technologies. Gemini, Cohere, PostgreSQL with pgvector, Langfuse, Python, Next.js, Hono, AI SDK, and Cloud Run.

The challenge. A pharmaceutical distributor receives thousands of operational documents across more than 40 different types, including orders, quality documentation, and transport documents. The information had to be reviewed and manually re-entered into SAP from Freshdesk, causing errors, delays, and repetitive work across support, operations, and back office.

My work. I developed a system that covers the full document workflow: intake from the Freshdesk support channel, classification, structured AI extraction for each document type, validation, and submission of the resulting records to SAP.

The solution is divided into document modules with configurable rules and mappings that adapt to different suppliers and formats. It also includes validation, error handling, and integration with SAP responses that do not always behave as expected.

Result. The system is in production and continues to evolve. It turns unstructured documents into usable records inside SAP, reduces manual data entry, and allows the extraction layer to adapt as new formats or suppliers appear. It also opens a new path for accessing data: information available in SAP can be queried through AI agents.

Technologies. Vertex AI with Gemini, Google ADK, Cloud Run, Datastore, Cloud Storage, SAP, and CI/CD.

The challenge. An organization managing the electronic catalog of millions of products verified the classification and keys of pharmaceutical products manually. The process depended on people, produced delays and transcription errors, and did not scale with the volume of medical data.

My work. I worked on the design and development of a serverless AI agent system that separates seeing from understanding: one agent transcribes product images with high-fidelity OCR, and a second agent combines that transcription with visual context to extract data into a strict, structured schema. The system is exposed as an API on Cloud Run.

Result. Product classification and validation went from a manual process taking hours to seconds, with a correction rate above 95%. Team members moved from entering data to validating its quality.

Technologies. Google GenAI (Gemini), Cloud Run, Python and FastAPI, and strict data validation with Pydantic.

The challenge. Coding assistants work well with source files, but their integration with real Google Docs and Drive documents is more fragile. Large documents, Markdown conversion, and format preservation introduce errors into workflows that should be able to run autonomously.

My work. I maintain an extended version of google-docs-skill. I added Markdown support and reliable insertion of large documents into Google Docs. The tool can be used from Claude Code, Gemini CLI, Codex, and Cursor.

Result. It is a public tool used in real agent workflows. It allows AI assistants to read, transform, and create Google Workspace documents without requiring manual steps for every operation.

Technologies. Google Docs and Drive APIs, TypeScript, Markdown, and skills for AI assistants.

Final-year thesis on machine learning applied to portfolio allocation. I implemented and compared LSTM, CNN-LSTM, LightGBM, XGBoost, random forest, stacking, and ARIMA models. I built the hyperparameter search, evaluation, and backtesting workflow with vectorbt and qlib, along with Dash and Plotly dashboards for analyzing the results.

I implemented several computational linear algebra methods from scratch: Householder QR factorization, power iteration, LU factorization comparisons, and analysis of bases under perturbation. Each implementation is accompanied by a LaTeX report examining its numerical behavior.

I built a complete workflow for extracting, cleaning, and analyzing web analytics data. The project combines exploratory notebooks with a Streamlit application for examining the results through an interactive dashboard.