ALEX
JERPELEA

sophomore at columbia university studying cs & math

aij2115 [at] columbia [dot] edu

// about

I’m Alex Jerpelea, a rising sophomore at Columbia University studying CS and math. My interests are all over the place. On the AI side I love deep learning, mechanistic interpretability, multi-agent systems, latent communication, and machine translation. Beyond that I’m into efficient algorithms, systems design, UNIX, decentralized networks, and building fun consumer apps. I grew up in Bucharest, Romania, where I spent most of high school doing competitive programming (see the awards below). From there I fell into NLP, then machine learning more broadly, then startups, and the list keeps growing. Mostly, I’m just having fun learning as much as I can.

currently

Most of my time goes to a model-diffing project at Columbia’s NLP lab: DINOv3 and CLIP have the same architecture and see the same data, so whatever separates them comes from the training objective, and I want to know what language supervision costs a vision model in fine-grained structure. Most of the effort there is method work — sparse dictionaries whose atoms are subspaces rather than directions, and a resolver I designed to stop them from splitting a single concept across many features. Alongside that I do research at Columbia’s DAPLab on a shared belief memory to mitigate inter-agent misalignment in LLM multi-agent systems, drawing on Theory-of-Mind research, and I run my own experiments on how Transformers could be more efficient and deeper learners.

// portfolio

research projects

Feature-pair trajectories over training: partners vs bystanders vs random pairs

Model Diffing: CLIP vs. DINOv3

My main research right now. Diffing vision models CLIP and DINOv3 to understand how CLIP’s joint language objective costs it fine-grained visual features. Designed a novel tournament-style top-K algorithm for block-sparse SAEs, significantly reducing feature splitting. Experimented with block-sparse cross-coders, alongside Matryoshka and Archetypal SAEs, in order to find the best diffing recipe. With Amith Ananthram (Anthropic Fellow), in Prof. Kathleen McKeown’s lab; in preparation for ICLR.

model diffinginterpretabilityvisionwip
AroTranslate

AroTranslate

Built the first neural MT system for Aromanian, an endangered language related to Romanian; ran the largest bilingual Aromanian–Romanian data collection and deployed the system online, reaching 5,000+ users. Primary author of the COLING 2025 paper; represented Aromanian preservation at the Forum of European Minority Regions (San Sebastián) and UNESCO’s “Language Technology for All” (Paris).

nlptranslationpaper paper → site →
Latent Belief States

Latent Belief States

A shared belief memory to mitigate inter-agent misalignment in LLM multi-agent systems, drawing on Theory-of-Mind research. I’m exploring methods from belief probes to KV-cache sharing to RLVR training, benchmarked against both memory and ToM baselines. At Columbia’s DAPLab with Prof. Eugene Wu.

diagram from an NVIDIA blog

multi-agenttheory of mindongoing read →
2D-Transformers

2D-Transformers

A testbed for my hypotheses on making Transformers more efficient and deeper learners. Very much a work in progress.

transformersefficiencywip read →
Affective State Identification

Affective State Identification

Affective states are a more open-ended notion than emotion. Grown from a Romanian pipeline into a multilingual affective-states benchmark, with interpretability experiments that locate emotion manifolds and diff broader affective-state representations against them. At Columbia NLP with Prof. Kathleen McKeown; in preparation for EACL.

nlpemotioninterpretabilitywip read →

hackathon projects

DopaMAXX

Top 3 · NY Tech Week

DopaMAXX

A Chrome extension that optimizes your feed for dopamine using EEG signals.

chrome extensioneeg / bci open →
neuro-trading

Best Use of Hermes · Nous Research

neuro-trading

fMRI-enhanced swarm simulations for prediction markets.

yhack 2026pytorchmulti-agent open →
Syscall

Best Use of Nemotron · NVIDIA

Syscall

Agents coordinating agents to solve software & kernel engineering.

sw26 buildathonclaude agent sdkcuda open →
scaled

Best Use of Solana · MLH

scaled

Crowdsourced diffusion data for world models, on Solana.

hack@brown 2026diffusionsolana open →
Open Disaster

Best Social Impact · Best Use of Featherless

Open Disaster

AI agents in real-time disaster scenarios.

devfest 2026three.jsagents open →

other projects

RoMemes: a Romanian meme and its annotation mask

RoMemes

Compiled and annotated the first digital corpus of Romanian-language memes, with specialized text-extraction and political-sentiment-detection algorithms. Co-authored “RoMemes: A Multimodal Meme Corpus for the Romanian Language,” awarded Best Paper at ConsILR 2024.

nlpmultimodalpaper read →
open-deadlocks

open-deadlocks

A benchmark pipeline that injects realistic Java deadlocks and tests whether AI agents can detect and fix them.

pythonjavaagents open →
slopache

slopache

A full open-source, end-to-end pipeline that turns one script into a finished character-grounded vertical reel (ComfyUI · FLUX · Wan · ElevenLabs).

comfyuidiffusionffmpeg open →
hive

hive

Hive is an open-source platform where AI agents collaboratively improve shared artifacts. My contribution: a system that lets an agent self-verify it has actually completed a round of training.

multi-agentopen sourcepython open →

// experiences

  • Model Diffing: CLIP vs. DINOv3

    Jan 2026 – present

    Columbia NLP · New York, NY · w/ Amith Ananthram (Anthropic Fellow)

    Diffing vision models CLIP and DINOv3 to understand how CLIP’s joint language objective costs it fine-grained visual features. Designed a novel tournament-style top-K algorithm for block-sparse SAEs that significantly reduces feature splitting, and experimented with block-sparse cross-coders alongside Matryoshka and Archetypal SAEs to find the best diffing recipe. In preparation for ICLR.

  • Latent Belief States for Multi-Agent Systems

    Jun 2026 – present

    Columbia DAPLab · New York, NY · w/ Prof. Eugene Wu & Yusen Zhang

    Designing a shared belief memory to mitigate inter-agent misalignment in LLM multi-agent systems, drawing on Theory-of-Mind (ToM) research. Exploring methods ranging from belief probes and KV-cache sharing to RLVR training, benchmarked against both memory and ToM baselines.

  • Multilingual Affective State Identification

    Jan 2026 – May 2026

    Columbia NLP · New York, NY · w/ Prof. Kathleen McKeown

    An affective state is a more open-ended type of emotion. Built a multilingual affective-states benchmark and ran interpretability experiments locating emotion manifolds and diffing broader affective-state representations against them. In preparation for EACL submission.

  • Researcher · AroTranslate

    Dec 2023 – Mar 2025

    Human Language Technologies Research Center · Bucharest, Romania

    Built the first machine translation system for Aromanian, an endangered language related to Romanian. Ran the largest bilingual Aromanian–Romanian data collection to date and deployed the system at arotranslate.com, reaching 5,000+ users. Primary author of “Dialectal and Low Resource Machine Translation for Aromanian” (COLING 2025, Abu Dhabi). Represented Aromanian preservation at the Forum of European Minority Regions (San Sebastián) and UNESCO’s “Language Technology for All” (Paris), as well as in mainstream media.

  • ML Engineer Intern · Veridion

    Jun 2024 – Aug 2024

    Veridion · Bucharest, Romania

    Developed and deployed large language models in production to gain firmographic insight from internet data.

  • Research Assistant · RACAI

    May 2024 – Aug 2024

    Research Institute for Artificial Intelligence, Romanian Academy (RACAI) · Bucharest

    Compiled and annotated the first digital corpus of Romanian-language memes; developed specialized text-extraction and political-sentiment-detection algorithms. Co-authored “RoMemes: A Multimodal Meme Corpus for the Romanian Language,” awarded the Best Paper Award at ConsILR 2024.

  • Research Intern · SpeeD Lab, UPB

    May 2023 – Jan 2024

    SpeeD Lab, University POLITEHNICA of Bucharest · Romania

    Machine learning to classify exercises from sensor data; built custom movement-tracking bracelets and ran a 23-participant data-collection campaign.

// awards

…thousands of Codeforces problems later

  • 2025

    I.I. Rabi Scholar, Columbia University

    One of 12 students selected for the program at admission.

  • 2022–25

    National Olympiad in Informatics, Romania

    Gold & 7th nationally (2025) · Silver (2024) · Bronze (2023, 2022) · Extended National Team, IOI (2025)

  • 2024

    International Autumn Tournament in Informatics (IATI)

    Bronze medal at this international, olympiad-style competitive-programming contest in Shumen, Bulgaria.

  • 2023

    Romanian Masters of Informatics

    Bronze medal at this international olympiad-style contest.

  • 2024

    Best Paper Award, ConsILR 2024

    “RoMemes: A Multimodal Meme Corpus for Romanian”

  • 2024

    National Olympiad in AI, Romania

    Extended National Team (international competition)

  • 2024

    National Olympiad in Mathematics, Romania

    Qualified to the national phase

// press

selected coverage of AroTranslate, the Aromanian translator