Notes, experiments, and technical deep-dives into models and architectures.
Why we ripped out our Vision Transformers for State Space Models in production.
A deep dive into rank deficiency and catastrophic forgetting when fine-tuning Llama-3 70B.
Notes on reward clipping and advantage estimation in noisy financial environments.
Why standard semantic search is dead, and how agents are taking over retrieval.
Achieving 5k tokens/sec on an A100 cluster.
A retrospective on 10 Kaggle competitions and why XGBoost still reigns supreme.
Feeding screenshots and DOM trees into an autonomous web-testing agent.
Using frontier models to generate high-quality Kaggle datasets for small model tuning.
Implementing an online learning pipeline for our recommendation engine.
Why giving your agent more time to "think" beats training a larger model.
Replacing autoregressive models with score-based generative models for market prediction.
Debugging NCCL timeouts with Hugging Face Accelerate and FSDP.
Using Vision-Language models to detect manufacturing defects without labeled data.
Benchmarking Parameter-Efficient Fine-Tuning (PEFT) methods.
How to prevent your Mixture of Experts model from only using two experts.
Why adding more features usually destroys your trading model.
Client-side ML is finally viable for real workloads.
Moving beyond simple sliding windows for LLM context.
Modeling transaction networks to catch bad actors.
When standard PyTorch operations aren't fast enough.
Why policies trained in simulation fail catastrophically in the real world.
Why human-in-the-loop and LLM-as-a-Judge are mandatory.
Extracting microstructural alpha from Level 2 data.
Combining visual inputs with semantic reasoning for path planning.
How we sandboxed our code-executing agents.
Training deep neural networks without centralizing sensitive patient data.
Advanced knowledge distillation techniques for Large Language Models.
Extracting sentiment alpha from 10-K and 10-Q reports.
Balancing recall and latency in billion-scale vector databases.
Choosing the right orchestration framework for multi-agent systems.
Learning visual representations without a single labeled image.
Why Direct Preference Optimization is taking over model alignment.
Making models aware of changing market volatility states.
Compiling Transformers for Nvidia hardware.
Using Playwright and VLMs to navigate dynamic single-page applications.
Solving the noise problem in financial transformer models.
Closing the reality gap with domain randomization.
Don't pay OpenAI twice for the same question.
You can't teach a model to run before it knows how to walk.
Integrating Markowitz Portfolio Theory directly into neural networks.
Text-extraction ruins PDF structure. Vision models fix it.
How to build systems that get smarter every time a user interacts with them.