Daniel Mora

AI Systems & Research Engineer

I study how models represent information and build AI systems that survive contact with the real world.

About

Who I am

Hey, I'm Daniel. I'm an AI systems and research engineer focused on interpretability, automation, and applied machine learning. I like building systems that are useful in the real world: trained models, agentic pipelines, research tooling, client software, and the occasional game when the creative itch wins.

IE University

AI & ML Builder

Spanish & English

Selected systems & research

Flagship work

Three projects that best represent how I think: begin with a hard question, make the system legible, and test the part most likely to fail.

01Model · weights released

JEPA Transformer

A two-stage equity forecaster. ArgusJEPA learns point-in-time asset-day embeddings; frozen, they become the sequence input to CassaFormer, a causal RoPE + Mixture-of-Experts Transformer with custom Triton kernels.

2: training stages
MoE: expert routing
2.4×: vs torch.compile

Architecture case study ArgusJEPA CassaFormer

Pretrain, freeze, forecast

Stage I · ArgusJEPAself-supervised · one day

Asset-day featuresprices, filings, macro, news

Context encoderhalf the features masked

EMA teacherfull view, latent target

Freeze · one embedding per day t

Stage II · CassaFormersupervised · sequence of days

Causal attentionRoPE, QK‑RMSNorm

Mixture of expertsTriton kernel

Student-t headmulti-horizon location, scale, tails

02Thesis · paper

Brain–LLM Feature Correspondence

A cross-description test of whether sparse language-model features and human brain responses depend on the same representational structure.

59: participants
200: brain parcels
6: SAE systems

Case study Paper PDF

Cross-description transfer

Select, then freezesource description only

Description Aspoken story, 59 listeners

SAE feature bank256 feature IDs, locked

Transfer unchanged to description B: same events, new words

Predict both targetsheld-out description

Brain · fMRI200 parcels, r = .07–.09

Model · LM hidden stateprediction transfers

Same features, same order? ρ = −.06 inconclusive: the test rarely detects planted agreement

03Autonomous media system

TC_ACC Documentary Studio

A multi-stage production system that turns a case reference into a researched, narrated, rendered documentary episode.

28: stages
5: departments
2: review gates

Architecture case study

A montage of documentary imagery produced for a TC_ACC episode

Research5 stages
Story4 stages
Assets7 stages
Gate
Animation2 stages
Gate
Publish8 stages

Additional work

More systems, with the work left visible

These are supporting projects, not footnotes. Each one shows what I owned, what I built, and where the work can be inspected.

Summaries are shown first.

Model research · release

Triplet SAE Project

Compared LoRA fine-tuning with SAE feature steering to determine which technique was better for subject–predicate–object extraction used in Graph-RAG navigation.

Role, engineering details & stackHide project details

Research question Which approach performs better for Graph-RAG triplet extraction: LoRA fine-tuning or targeted steering through sparse autoencoder features?

Built comparable SAE-steering and LoRA fine-tuning workflows, then evaluated their subject–predicate–object extraction quality for downstream Graph-RAG traversal.
Fine-tuned Gemma 3 4B IT with LoRA and packaged deployment-oriented GPTQ 4-bit and 8-bit variants.
Ran the GPU-heavy analysis and training workflow on Google Cloud compute.

PythonPyTorchGemma 3Graph-RAGSAEPEFT / LoRAGoogle CloudHugging Face

Interpretability infrastructure

OLMo SAE Builder

A readable, single-GPU pipeline for turning large text corpora into released sparse autoencoders for OLMo-3-7B-Think.

Role, engineering details & stackHide project details

Scope Data collection, activation infrastructure, SAE training, validation, and public model packaging.

Streamed Dolma and Dolci-Instruct-SFT data, captured residual activations with forward hooks, and wrote disk-backed chunks instead of loading the corpus into memory.
Trained TopK SAEs across four model layers on roughly 20 million token activations per layer and released seven SAELens-compatible artifacts.
Built and operated the Google Cloud GPU workflow needed to move data, run long jobs, preserve checkpoints, and publish the outputs.

PythonPyTorchOLMo 3 7BTopK SAECUDAGoogle CloudSAELens

Agent experiment

Dual-Agent AI Chatbot

A controlled persuasion experiment that placed a helpful electric-car guide against a deliberately relentless Dacia Duster salesperson.

Role, engineering details & stackHide project details

My role Architect and project manager; owned the experimental design and research write-up while two teammates implemented the agents.

Designed opposing agent goals and the conversation structure.
Turned a playful chatbot into an explicit behavioral experiment with documented findings.

PythonCrewAILLM agentsExperiment design

Learning product · team project

LearnMate

An agentic study partner that combines interactive help with personalized learning-resource recommendations.

Role, engineering details & stackHide project details

My role Built the agent flows, backend, and first Streamlit prototypes in a seven-person team.

Connected conversational guidance to resource discovery and recommendation flows.
Carried the work from early prototype into an integrated team application.

PythonStreamlitLLM agentsRecommendation systemsAPIs

Cloud application · team system

Cloud Banking Platform

A mock banking platform built as a six-role software team, with application code, infrastructure, delivery automation, and secret handling treated as one system.

Role, engineering details & stackHide project details

My role One of two full-stack developers, working across the Vue interface, Flask API, and deployment infrastructure.

Implemented product work across frontend, backend, and PostgreSQL-backed data flows.
Provisioned Azure infrastructure with Bicep and supported CI/CD, secrets, sprints, and ticketed team delivery.

VueFlaskPostgreSQLAzureBicepGitHub Actions

Solo automation system · sunset

Automated Content Poster (ACC)

A source-to-publish media pipeline that rewrote content, created voiceovers, assembled posts, and published them automatically.

Role, engineering details & stackHide project details

Ownership Solo build; operated it in production, tracked the output, and chose when to sunset it.

Connected GPT rewriting, ElevenLabs narration, media processing, scheduling, and Instagram delivery.
Deployed the long-running automation on Google Cloud Run and handled its production lifecycle.

PythonOpenAIElevenLabsCloud RunAutomationMedia APIs

C++ systems · game

Alien Invasion

A Space Invaders-style SFML game used to practice real-time state, rendering, testing, and native build tooling.

Team project details & stackHide project details

Collaboration Team C++ project.

The project includes four enemy types, difficulty scaling, collision and scoring systems, plus a persistent leaderboard.
Its engineering surface includes unit tests, CMake build tooling, and Sphinx/Doxygen documentation.

C++SFMLCMakeUnit testingSphinx / Doxygen

Skills & expertise

The toolkit behind the work

Languages

Python (advanced)
C++
SQL
JavaScript
TypeScript

Modeling & Interpretability

PyTorch — SAE training, forward hooks, mixed precision
Sparse autoencoders — TopK training, ablation, feature steering
Transformers — Gemma, OLMo, Llama, Qwen, DeepSeek, Kimi, PEFT / LoRA
Local models — local inference, evaluation, and model-serving workflows
Evaluation — ridge models, RSA, structured extraction metrics
Hugging Face — model cards, quantized exports, SAELens releases

Research Compute & ML Systems

Google Cloud and Vast.ai — GPU training workloads and remote Linux environments
Large activation datasets — streaming, chunking, storage, transfer
Long-running jobs — checkpoints, recovery, validation, artifact publishing
CUDA-aware workflows, bfloat16 / float16, memory-conscious data loading
Triton — GPU kernel work for ML workloads
Distributed systems — coordination, checkpointing, failure handling, and recovery

Agentic Systems & Backend

Agentic workflow design, structured outputs, and provider integration
OpenAI, Claude, Gemini, DeepSeek, Kimi, CrewAI, LangGraph, provider integrations
Flask, FastAPI, SQLAlchemy, PostgreSQL, Redis
Automated media workflows with ffmpeg and Remotion