# Blog.log
Writing on AI systems, full-stack product work, and engineering that ships.
5 minBuilding RAG Systems That Teams Can Actually Trust
Hybrid retrieval and citation checks turn a flashy demo into something support teams can use with confidence.
RAGAIProductread →
4 minShip AI Features Like Software, Not Magic
Treat models as components with contracts, failure modes, and observability — the same discipline as any production service.
EngineeringAIOpsread →
4 minNotes From Building EventRent End to End
Lessons from turning equipment rental requests into an operable full-stack product with admin workflows.
Full-StackReactProductread →
5 minWrite the Eval Before You Shop for Models
Pick models against a task set, not a benchmark leaderboard — offline evals keep demos honest.
MLEvalsAIread →
5 minPractical Agent Design: Tools Beat Personality
Agents that work feel boring in the best way — clear tools, tight scopes, and exit conditions.
AgentsAISystemsread →
5 minEmbeddings That Match Your Domain, Not Just a Model Card
Chunking, metadata, and query rewriting matter as much as which embedding model you pick.
EmbeddingsRAGSearchread →
5 minFairness Is a Product Metric, Not a Paper Section
Bias-aware training only helps when metrics ship next to accuracy and latency in the same dashboards.
FairnessMLEthicsread →
4 minFastAPI Patterns That Keep ML Services Maintainable
Typed request models, explicit timeouts, and a clean separation between inference and I/O.
PythonFastAPIBackendread →
4 minTypeScript Frontends for AI Products That Stream
Streaming answers need typed events, cancel tokens, and UI states that respect partial data.
TypeScriptReactAIread →
4 minDevOps Habits ML Apps Still Need
Containers, config, and rollback stories remain underrated when demos ship straight from a laptop.
DevOpsCloudMLread →
4 minContext Windows Are Not Product Memory
Long contexts help, but durable knowledge still needs storage, retrieval, and lifecycle policies.
LLMsArchitectureAIread →
4 minStreaming UX That Feels Calm, Not Chaotic
Token streams can feel premium or anxious depending on layout, cursor behavior, and error recovery.
UXAIFrontendread →
5 minObservability for LLM Apps Without Drowning in Logs
Trace prompts, retrieval hits, costs, and user outcomes — not every token forever by default.
ObservabilityAIOpsread →
4 minFeature Flags for Model Rollouts
Ship behind flags so model upgrades are controlled experiments, not all-or-nothing deploys.
ReleaseAIEngineeringread →
5 minFix the Data Before You Fine-Tune
Most “we need a custom model” problems are labeling, coverage, and leakage problems first.
DataMLTrainingread →
More posts soon. Prefer questions? Reach out.