Projects and Publications
ClaimValidator: Browser extension to increase trust while using the internet
Using open-source LLMs to check whether sources truly validate the claims made users on Reddit. In general, making the internet more trustworthy with AI assistants.
Release Page:
Chrome Webstore Link
Implementation Page:
https://github.com/veezbo/claim_validator
Llourney: Stable Diffusion with a fine-tuned LLM instead of CLIP
Exploring the feasibility of replacing CLIP with an LLM fine-tuned using a frozen Stable Diffusion model and diffusion training loop.
https://github.com/veezbo/llourney-dev
AkkadianOracle: Understanding the Ancients through Conversation
Pre-training and Retrieval-Augmentation Generation with LLMs and English-translated Akkadian language corpuses for the ultimate goal of better understanding the “umwelt” of these ancient peoples. Talking with Penn Museum to integrate into the visitor experience and direction for future research.
Live Demo:
https://poe.com/AkkadianOracle (free)
https://poe.com/AkkadianArchon (req. Poe subscription)
Implementation:
https://github.com/veezbo/akkadian_oracle
Paper:
Arxiv link coming soon
Long-term Financial Asset Optimization
Novel AI to optimize the highly unconstrained problem of spending and saving money to maximize portfolio and minimize lifetime taxes paid. Individualized, optimized advice is provided with new market data every year, optimized up to a 40-year timeframe.
Evolutionarily-Curated Curriculum Learning for Deep Reinforcement Learning Agents [@ Imbellus]
https://arxiv.org/abs/1901.05431
Paper representing the effort I led at Imbellus to develop a system to procedurally generate games (tasks) in our assessment while maintaining similar difficulty using novel deep RL agents trained to play the games.
Metadata Feedback Loop for Dynamic Spam Cohort Detection [@ Tinder]
From concept to implementation: a novel dynamic feedback loop that automatically generates suspicious combinations of metadata and results in large cohorts being caught at once. Critically, the fully dynamic system automatically adapts to the shifting suspicious behavior of bad actors. Implemented as a novel streaming infrastructure in Flink for real-time detection, before it impacts real users.
Latent Transaction Feature Discovery with Deep Generative Models [@ Visa]
Led the effort to train deep generative models (VAE, GAN, and hybrids) on transaction data with continuous and categorical features for synthetic transaction generation and the learning of latent features to improve merchant recommendation and fraud detection through dataset augmentation.
Prevalence and recoverability of syntactic parameters in sparse distributed memories [@ Caltech]
https://arxiv.org/abs/1510.06342
We stored the syntactic parameters of various world languages in Kanerva Networks with corruption, and used the recoverability as a proxy to understand the universal prevalence and relationships of those parameters.