Gennaro Francesco Landi / Selected work
I work on the boundary between AI applications and the infrastructure beneath them. These projects focus on measurable behavior, inspectable decisions and tools that make experiments easier to reproduce.
Can a short request prefix choose a cache policy? A held-out evaluation over seven public traces, with synthetic overlays and explicit fixed-policy baselines.
59/63 family predictions correct. +0.141 pp mean hit-rate gain over the best fixed baseline.
Results, methods & technical note →Request-level browser workload traces and a reproducible cache replay. Measure the requests behind an agent run and inspect what the cache can reuse.
On the scripted trace, GDSF improves request hits but reduces byte hits at 5 MiB. Both metrics are reported.
Explore the cache tradeoff →Typed, composable AI workflows with loops, branches, maps and Zod validation. A deterministic mock runtime makes orchestration testable without inference calls.
Local benchmark samples and schema-validation regression tests accompany the implementation.
Read the engineering case study →First place in the Chain IQ case challenge at START Hack 2026. A procurement prototype built around deterministic decisions and an auditable execution path.
Public hackathon code and an architecture write-up.
Architecture & project story →V2X co-pilot: traffic/weather adapters, spatio-temporal filtering, TTL caching and fuzzy risk scoring. TORCS: behavioral cloning with nearest-neighbor search. Broletter: personalized science delivery from arXiv to Telegram.
The research pages distinguish measured findings from open questions. They link to the exact reproduction code and describe where the released data limits the conclusion.