Projects
A platform that analyzes your entire codebase (via .ZIP upload) and tells you whether your projects are strong enough to get hired, adapted by target role and seniority level.
- •Designed a multi-agent system that evaluates architecture, testing, maintainability and code quality, generating a hexagonal chart with actionable improvement suggestions.
- •Implemented a Go backend that reconstructs repository structure, identifies languages and patterns, and feeds a role-aware scoring pipeline.
AI-driven data pipeline automation platform built with a multi-agent architecture using LangGraph. PRD2Blog transforms Product Requirements Documents (PRDs) into publication-ready technical blog posts using a coordinated multi-agent workflow.
- •Multi-agent system with LangGraph that researches external sources, verifies factual consistency, and iterates until quality thresholds are met.
- •Delivers polished articles while tracking token usage, execution costs, and pipeline metrics.
Current Projects
Building a Transformer-based LLM inference engine from scratch in modern C++, implementing the complete mathematical and systems pipeline required for autoregressive text generation. The project includes a custom tensor library, optimized matrix multiplication, softmax, RMSNorm, SiLU activations, positional embeddings (RoPE), and a full Multi-Head Attention implementation with KV Cache for efficient token-by-token inference. The engine is designed with a modular architecture featuring custom memory management, multithreaded request scheduling, tokenizer integration, and an HTTP server, closely resembling the architecture of production inference systems such as llama.cpp and vLLM.
- •Implementing the complete Transformer forward pass from scratch, including Multi-Head Attention, RMSNorm, RoPE, SiLU, residual connections, and autoregressive decoding.
- •Developing a custom tensor library and core mathematical operations such as matrix multiplication, softmax, normalization, and activation functions.
- •Designing an inference engine with KV Cache, multithreaded request scheduling, and modular components inspired by modern LLM serving architectures.
- •Building the surrounding infrastructure, including tokenizer integration, custom memory management, worker threads, and an HTTP API for model serving.
- •Structuring the project to closely mirror real-world inference engines, enabling future support for pretrained Transformer checkpoints and optimized execution.
AI assistant for accounting firms that answers tax questions always citing the official ARCA/AFIP regulations. RAG over a vectorized corpus of official regulations, with persistent per-client context, full history, multi-accountant, and automatic regulation sync.
- •RAG over official ARCA/AFIP regulations with verifiable citations (document + URL) in every answer.
- •RAG chatbot (retrieval & generation) with Gemini 2.5 that tailors answers to each client's real context.
- •Ingestion bot running as a nightly cron job: detects changes by hash and re-indexes only what changed.
- •Per-client workspace with full history and multi-accountant login with row-level security.