Cristian Desivo
AbstractSince 2016 I have taken projects from a set of requirements to a working system: modeling the problem, building the data and models, writing the code and delivering it. My recent work focus on making language models useful in places where a wrong answer is not acceptable.
§ 1Selected cases
Constrained question answering
An assistant whose every reply is drawn from a fixed set of pre-written answers. It cannot invent facts, and nothing can be injected into the prompt to change this, by design.
LLM features that can't make things up
DirectRetrieval, my Python library for AI assistants that can't make things up, running live on this site. Plus LLM features built for clients, from document drafting in Zendesk to chatbots …
People counting with computer vision on a Jetson Nano
A camera system that counts people entering and leaving an area and checks whether workstations are occupied, running in real time on a Jetson Nano.
Optimization: container loading and factory scheduling
Demos on sample data: a shipment packed into 2 containers instead of 3, and a factory schedule that cuts late orders from 10 to 3. Client work includes container loading, …
Games: Pixel P.I. and earlier prototypes
Pixel P.I., the detective game I designed, programmed and released on Steam on my own, from the Unity client to the Python server and my own writing tools. Plus earlier …
Report generation from Excel, Word, CSV and text files
A desktop tool that pulls data out of Excel, Word, CSV and text files and builds reports to the client's specification, with a simple window to pick the inputs and …
Reinforcement learning in a driving simulator (CARLA)
An agent trained with reinforcement learning to complete a task inside the CARLA driving simulator, deciding from what its camera sees.
Circuits that work as neural networks
Electronic circuits that compute like a neural network (weighted sums, activations, signal propagation), designed and tested in SPICE simulation before hardware.
Flight instruments on a Raspberry Pi
Simulated aircraft instruments on a Raspberry Pi, driven by physical controls, with a web page to configure them.
§ 2Games of my own
Pixel P.I.
A detective game I make as a solo developer. Players type their own questions; a statement unlocks when the question matches what the police asked.
View on Steam ↗
Murder in the Aethertree Forest
A supernatural murder-mystery visual novel, made in one month for the Spooktober Visual Novel Jam.
Play on itch.io ↗§ 3Methods
Reliable AI systems
Retrieval, agents and structured outputs, bounded to approved content.
Mathematical optimization
Operations research, non-linear optimization, metaheuristics.
Deep learning
Datasets, training and deployment, computer vision.
Local inference
llama.cpp, vLLM, Ollama; choosing and quantizing models.
Games
Unity, Godot and Ren'Py.
Python, thoroughly
Django, Flask, FastAPI, NumPy, PyTorch, PyQt and more.