Blog
Technical writing on full-stack development, AI/ML engineering, and lessons learned building production systems from scratch.
How I built a Retrieval-Augmented Generation pipeline that lets users ask natural-language questions over any PDF — covering chunking strategy, vector similarity search, and prompt engineering.
A deep dive into fine-tuning a pre-trained DistilBERT model on Google's GoEmotions dataset (58K Reddit samples) to achieve 88% accuracy across 27 emotion classes — including lessons on class imbalance, tokenization, and evaluation.
The architecture decisions behind FlowCRM — org-scoped real-time rooms, Bull queue offloading for 3x response-time improvement, RBAC at the org level, and how I handled multi-tenancy in PostgreSQL with Prisma.