Giacomo Cappelletto

GiacomoCappelletto

I build software, ML systems, and robotics around messy real-world data.

View workResume

Software is most useful when it survives contact with the real world.

  1. AI engineering in Milan
  2. Biomechanics research at Boston University
  3. Seeking Summer 2027 software and AI roles

Experience

Four environments, from enterprise AI systems to physical sensing and product delivery.

Aug 2026 - Present

Boston, MA

Boston University · Prof. Brian Kulis

Undergraduate Machine Learning Researcher

Researching robust contextual metric learning for human-motion retrieval, learning fixed-length embeddings from temporal 2D/3D pose sequences for nearest-neighbor search across labeled movement classes.

Python / PyTorch / Metric Learning / Pose Embeddings / MotionBERT / MMPose

Jun 2026 - Aug 2026

Milan, Italy

Banca Mediolanum

Software & AI Engineering Intern

Lead developer, as an intern, for internal AI agent and prompt-governance platforms; own backend, frontend, data workflows, evaluation, deployment, and stakeholder requirements.

Python / SQL / Azure Databricks / Unity Catalog / Delta Lake / Databricks AI Search

Jan 2026 - May 2026

Boston, MA

Boston University College of Engineering

Undergraduate Researcher

Built a single-camera rowing biomechanics pipeline spanning stabilization, MMPose 2D keypoints, and MotionBERT 3D lifting.

Python / MMPose / MotionBERT / OpenCV / NumPy / Pandas

May 2025 - Oct 2025

Treviso, Italy

Società Cappelletto S.R.L.

Full-Stack / Systems Developer

Rebuilt an Electron inventory application with Tauri, React, and Rust, reducing app size by 70%, memory use by 80%, and startup time by 60%.

Rust / Tauri / React / TypeScript / Firebase / Shopify GraphQL

Nov 2024 - Feb 2025

Remote

TickIT

Full-Stack Engineer

Worked across Next.js and TypeScript interfaces plus Rails and PostgreSQL backend features as part of a four-developer remote event-ticketing team.

Next.js / TypeScript / React / Rails / PostgreSQL / PostGIS

Skip featured work

Production software, applied ML, and physical systems built around real operational constraints.

MOVE selected

Exploded MOVE system showing an authenticated web request entering the governed agent runtime and returning through a versioned typed UI contract.
Enterprise AI / Platform2026

MOVE

A governed enterprise AI copilot that turns natural-language questions into scoped, traceable answers across approved CRM and document sources.

View project
Exploded Deskinator software stack showing sensing, signal stabilization, controller states, coverage planning, actuation, evidence, and the separate stepper pulse timing boundary.
Robotics / Autonomy2025

Deskinator

An end-to-end tabletop-cleaning robot that turns touchless edge sensing into a fitted workspace, a safe coverage plan, and measurable motor commands.

View project
Exploded inventory execution stack connecting staff SKU input, React and Tauri, Rust orchestration, Shopify reads and adjustments, Firestore logging, and refresh.
Systems / Product2025

Inventory System Rebuild

A Tauri desktop workflow that resolves exact SKUs across two locations, decrements one available unit through Shopify, records an app-owned audit log, and exposes bounded recovery paths.

View project
Exploded TickIT current platform showing public and organizer interfaces, authenticated organization-scoped Rails services, and PostgreSQL plus PostGIS event operations.
Full-Stack Product Engineering2024-2025

TickIT

A full-stack event platform spanning public discovery, ticket commerce, QR access control, organizer operations, and analytics.

View project
Exploded NoteWorthy system showing image preparation and server request guards, streamed Gemini LaTeX generation, template composition, configurable HTTP compilation, browser output, optional persistence, and separate historical rails.
Applied AI / Product2026

NoteWorthy

A multimodal notes-to-LaTeX pipeline that streams Gemini-generated structure into a styled template, compiles it through configured HTTP backends, and returns PDF plus source.

View project
Skip research showcase

Research

Studies on robust learning, from noisy motion embeddings to biologically inspired adaptability.

Contextual Similarity selected

System view / research question

Pose → context → retrieval
Corrupted query → clean-gallery retrievalExplanatory animation / not measured output
Metric Learning / Pose Retrieval2026

Contextual Similarity

A retrieval-first study of whether contextual metric learning can preserve motion similarity when joints jitter, disappear, or frames drop.

Read proposal
Side-view rowing frame with full-body pose landmarks, face and hand tracking, ergometer detection, joint angles, and a 3D pose inset.
Research / Applied ML2026

Rowing Biomechanics Pipeline

An inspectable research system that learns a stroke-level force representation from synchronized video and RP3 telemetry, then packages the same kinematic contract for video-only inference.

View case study
Three independent evidence lanes relate MLC study and query episodes, PFC-inspired context mechanisms, and plasticity-loss markers with a separate continual-backpropagation intervention.
Research / Technical Writing2025

Forging Adaptability

A literature-based examination of systematic compositionality, continual learning, adaptive plasticity, and the limits of biomimetic AI.

Read paper

Notes

All notes

Summer 2027 software and AI opportunities

Let's build something that has to work.

Email