Day 31: Lambda + Bedrock: Sync, Stream, and Batch
Choose Lambda sync, response streaming, or batch Bedrock invokes by client type — and surface throttles instead of hiding them behind retries.
Posted inAI System Design
Day 4: Vector Indexes Without Magical Thinking
HNSW vs IVF vs pgvector: pick indexes by latency, recall, and filter patterns — a SQL predicate often beats a larger ANN index.
Posted inAI AWS Machine Learning
Day 30: Project: Multi-Tenant Code RAG on Bedrock Knowledge Bases
Ship multi-tenant code RAG on Bedrock Knowledge Bases with hard isolation filters, per-tenant eval suites, and citation fail-closed gates.
Posted inAI Machine Learning
Day 3: Embeddings That Survive Domain Jargon
Generic Wikipedia embeddings fail on code and ops jargon — choose, version, and evaluate embedding models against your domain queries and corpus.
Posted inAI System Design
Day 29: Failure Taxonomy: Wrong File vs Wrong Advice
Triage RAG failures into wrong-file vs wrong-advice classes so retries, alerts, and product UX stop treating every miss the same.
Posted inAI Machine Learning System Design
Day 2: Transformer Attention as a Systems Diagram
Treat transformer attention, KV cache, and batching as a production data path — latency, memory, and throughput come from systems choices, not paper diagrams.
Posted inAI Machine Learning
Day 28: Multimodal RAG: Diagrams, PDFs, and Screenshots
Decide what to OCR, caption, or keep as images in multimodal RAG so diagrams and screenshots retrieve without drowning the index in garbage text.
Posted inAI AWS System Design
Day 75: Autoscaling Retrieval and Embed Jobs
Autoscale embed/retrieve workers from queue depth with concurrency caps so ingest cannot melt the vector DB.
Posted inAI Machine Learning
Day 1: Tokens, Context Windows, and Why Models Forget Mid-Task
Tokenization, context windows, and recency bias silently break chatbots — measure token budgets and attention to the middle before you pick a model or RAG design.
Posted inAI AWS Machine Learning
Day 74: Batch Inference Overnight
Bedrock batch inference for overnight repo-wide refactors — daytime humans apply validated intents.