Porfolio page

Jeanne Shih

Here are eighteen hours of movie soundtracks

I’m a formally trained computer & electrical engineer based in Montreal,
Quebec, Canada

My research interests
(unfinished & exploratory):

Compression techniques.
Hybrid realm quantum classical computing algorithms.
Classical and quantum based formalisms.
Steganography and Subliminal learning.
Uncertainty literature.
Eastern ways and philosophy

1

Frequency based linguistic message sending experiments
2025

2

Separability with PCA visualizations of periodicity patterns in text
2025

3
Pasqal quantum cloud emulator experiments
2025

4
Finetuning my own voice to voice model
2023

5
Perspective
A screenshot RUST app
2025

6

Contextualize
GPT-3 research app
2020

7

Novelty generator
A new research app
2025

8

ETH global Waterloo
2023

9
Quaintance AI
Image recognition app
2018

10

Bumpin
Quantum AI dating
2017

11

Vlyy
Video app
2014

12

Ungraduate university TCP peer to peer server and client project
2005

13

Mintmtl
2022

14

Energy
Grants collective on Zora
2023

15
The S.W.A.P. Team
2014

Steganography project 001

2025

We imagine that three representative agents (ALICE, BOB, and CHARLIE) should be able to simultaneously transmit hidden messages through a single shared channel using distinct carrier frequencies (1.0, 2.0, and 3.0 Hz respectively).

The 3D FFT visualization clearly shows three distinct spectral peaks at their respective carrier frequencies, demonstrating successful frequency separation. Each agent then would occupy a distinct frequency band (0.8 Hz bandwidth) with minimal spectral overlap, allowing for independent recovery via bandpass filtering.

This work is published as a technical preprint paper.

001

Top ↑

PCA project 002

2023

Synthetic experiment on PCA projections for viewing separablility of features of periodicity in text.

This work investigates whether language models can hide secret messages in their reasoning chains in a set of periodic patterns of transmission and whether they are separable in PCA analysis.

The study developed a custom BERT-based classifier system that simultaneously performs overt tasks such as truthfulness classification while encoding and detecting hidden messages. The periodic patterns of hidden messaging were achieved on a linguistic level, encoding periodicity using reasoning verbosity to encode bit string messages like "HELLO" and "SECRET" across a sequence of chained reasoning statements. The periodic patterns are separable using PCA visualizations (pastel).

This synthetic experiment is a proof of concept that periodic patterns can exist and be detected in reasoning like text.

002

Top ↑

Pasqal emulator experiments

2023

These are experiments conducted on Pasqal Cloud, a neutral atom quantum computer, using their cloud emulator to factor the numbers 21, 35 with 8-15 qubits and the breakage of 5-bit ECC using the KAIST methodology.

003

Top ↑

Persepective

2025

This is a RUST implementation of an iOS Screenshot AI Server that you can host locally to automatically track your Macbook ios desktop screenshots as you take them.

Timestamps and summarizes all screenshots to your Telegram bot timeline. Automatically track them all. If you’re a visual note taker and take a lot of screenshots this may be useful. See more.

003

Top ↑

Voice modeling project 004

2025

I trained my own voice model to sound like Uma Thurman, Quentin Tarantino, Kazuo Ishiguro ways of expressing themselves, including a character from a clip of the Richard Linklater movie, Waking Life.

005

Top ↑

UMA

QUENTIN

KAZUO

LINKLATER

Novelty generator

2025

This is a research tool integrating my custom model based on a truthfulness/untruthfunless BERT based classifier to generate novel Ideas that are fleshed out with Sakana AI Scientist. These are implemented to be mintable as NFT’s on Base (L2) and they may be registered as Story Protocol Intellectual Property. Licenses with terms are also assignable through Story. The researcher may also take down notes and mint these on Mode network (L2).

In this way a market of ideas is produced on two different defi blockchains. Eliza agent from ai16z is integrated. This won two hackathon bounty prizes in separate competitions here and here.

006

Top ↑

Contextualize

2025

A context-based recommendation engine built on GPT-3, OPEN AI’s then most advanced natural language model.

I built this to call the GPT-3 API in 3 different ways. The aim was to give context to a topic by training GPT-3 to think and speak like three different persons.

This was accepted for demonstration as one of the first GPT-3 demos presented to Pioneer.app in 2020 as part of “The First Open GPT-3 Demo Day” (Watch at approximately 41:50). The prototype was submitted as a “contextual reference recommendation engine using gpt-3” meant to serve as a research tool.

007

Top ↑

ETH GLOBAL WATERLOO 007

2025

My submitted project was an implementation of LLM’s and ERC-6551.

The project ingests a prompt that is presumed to be relevant to two pieces of embedded text or more. By way of example, I embedded a book by Nietzche and Jane Austen’s Pride & Prejudice as well as the prompt. Nearest neighbour search between the prompt embedding and each of the two embedded pieces of text was performed to generate a new vector embedding representing the three way relationship to be used to fine tune other LLM’s or the original model.

The new embedding is then stored as metadata on an NFT that by virtue of the being an ERC-6551 NFT has its own wallet. Having its own wallet means the NFT, which holds the embedding vectors representing the relationship between three pieces of text, is now able to autonomously transact itself in a hypothetical marketplace of AI text embeddings.

As they are stored as metadata on NFT’s, the embeddings may also be transacted by humans as valuable items for the refinement of personal LLM’s, thus combining a decentralized blockchain protocol, the ERC-6551 and LLM’s, within the context of an autonomous or voluntary marketplace.

In other words, it’s designed purpose is to improve LLM’s as cognitive tools for both humans and AI agents through an NFT marketplace of relationship bearing vector embeddings stored as metadata on token bound NFTs.

008

Top ↑

Quaintance AI

2025

Quaintance was an AI engine using Clarifai’s API, for IKEA showroom discovery to purchase furniture according to a person’s overall style & tastes.

These are the office and bedroom results from the prototype AI trained on IKEA room layouts, populating the most “similar” IKEA room layouts from an input image

009

Top ↑

Bumpin

2018

Bumpin was a AI-based dating app, meant to scrape users' social media metadata and images to generate representations of people using higher dimension vectors, calculating metrics and matching them according to the resolution of various quantum problems as optimization problems using quantum compute.

This project was proposed as part of the Creative Destruction Lab’s first Quantum Machine Learning stream.

010

Top ↑

It is for Jeanne

Top ↑

I
Research

II
Projects

III
Community