Personal Reflections & Tech Insights

Blogs

Building Safe LLM Agents in Production

LLM agents offer powerful automation, but their inherent unpredictability makes production deployments tricky. Architecting for reliability and safety is crucial to avoid surprises.

AI Models Decay. Here's How to Keep Them Useful

AI models degrade in production as data changes. Continual learning architectures offer strategies to keep models fresh and effective, but they add complexity.

Causality in Production AI Systems

Observing correlations in AI system behavior is easy. Understanding the underlying causes is much harder but essential for robust production systems and effective debugging.

Making Large AI Models Fit Small Hardware

Deploying powerful AI models on edge devices or resource-constrained hardware often means they're too big or too slow. Quantization and pruning are key techniques to shrink models while keeping them useful.

Beyond Single Agents: Building Robust Multi-Agent AI

Building complex AI often means moving past single agents. Multi-agent systems can tackle harder problems, but they come with significant design challenges.

Vector Search Speed: Advanced Indexing

Vector search is powerful for similarity, but brute-force comparisons don't scale. Efficient indexing is the key to making it practical for large datasets.

ML Supply Chain Security: Beyond the Model

Securing an ML system goes beyond the model itself. Protecting the entire supply chain, from data to deployment, is crucial for integrity and reliability.

Multi-Modal AI: It's Not Just About More Models

Building multi-modal AI systems means combining different data types or models. It's more complex than just chaining components; true integration is the real challenge.

Serving Fresh AI Features: The Real-time Store Connection

Real-time feature stores bridge streaming data with AI models, solving staleness and training-serving skew for low-latency predictions.