2025-01
RAGPostgreSQLpgvector
Hybrid Search with pgvector & BM25
How I combined dense vector search with BM25 keyword scoring to get retrieval quality that neither approach achieves alone. Reciprocal Rank Fusion merges both result lists.
> ls notes/
Notes
Short-form technical writing. Lessons from building things that actually work.
2025-01
LLMArchitecture
LLM Provider Fallback — Gemini to Claude
Building a provider-agnostic adapter that falls back from Gemini to Claude transparently, with streaming SSE preserved across the switch.
2025-01
SecurityLLMGuardrails
Two-Layer Prompt Injection Defense
Regex patterns catch obvious attacks fast. A lightweight LLM classifier handles the subtle ones. Three strikes → shadow-limit. How I keep the AI twin honest.
2025-01
SecurityGDPRCryptography
Encrypting Chat Messages at Rest
Using Fernet symmetric encryption with PBKDF2 key derivation to encrypt visitor chat messages in PostgreSQL. GDPR Article 32 compliance without sacrificing query performance.
2024-12
CeleryArchitectureRAG
Async Knowledge Ingestion with Celery
Chunking, embedding, and indexing documents without blocking the API. Celery workers process uploads in the background while the user gets immediate feedback.
2024-12
ObservabilityFastAPIPrometheus
OpenTelemetry in FastAPI — Practical Setup
Instrumenting a FastAPI app with OpenTelemetry, exporting traces to Grafana Tempo and metrics to Prometheus. What actually matters vs. what's noise.
Want to discuss any of these topics?
Ask the AI twin →