AI engineering · Go Fig take-home · online demo
Inbox Triage
Automatically sorts a shared inbox, selects actions and drafts replies. Uncertain cases go to a person; sending and external writes require confirmation.
System, design and demo →My interest
in AI.
I want to understand how AI applications work, especially the architecture around the model: how it uses documents, how agents coordinate and how answers are verified. In my projects I experiment with these choices and document their results and limits.
Computer Engineering student · Politecnico di Milano
01Projects
Problem, contribution and result.
RAG · Agent Research · WikiLLM
RAG — design decisions
I measured retrieval, reranking and generation choices on financial questions. A decision lab compares quality, time and cost; the page also introduces agentic search and advance synthesis.
Decisions and results →80.5%
Agents · experimental evaluation
τ²-bench
Six prompt versions for a support agent, with the model held fixed: an 80.5% average across four runs, compared with 68% for the default agent in one run. I examined errors in Langfuse traces and reported a benchmark defect.
Method and results →
Java · team · 30/30
Mesos online
In a four-person team, I built the Controller and Socket/RMI server paths and contributed to the model UML. Different types of clients can join the same match.
The product and my role →
C · individual project · 30/30 with honours
Shortest paths
Movhex: minimum costs on a changing hexagonal map. Computed land neighbours save memory; Dijkstra with a min-heap and stopping at the destination limit search work. Individual C project, 30/30 with honours.
Problem and solution →
Agents · human review · online demo
Quote review interface connected to Langfuse
An agent prepares quotes and messages, an automated judge evaluates them, and a nontechnical reviewer checks them before sending. The human and automated judgments appear together in Langfuse.
How it works and demo →02Skills
AI applications
Document retrieval, model decisions and human approval, with explicit validation rules.
See the workflow ↗Testing and observability
Repeatable experiments, automated checks and traces to investigate errors.
See the evaluation ↗Software engineering
Server design, modular interfaces and algorithms mindful of time and memory.
See the team project ↗