Vector Databases and Semantic Search: What I Learned Building Aria
Building SellTrove's AI assistant required understanding vector embeddings and semantic search. Here is the practical architecture.
DevDick / Field notes
Engineering decisions, research, setbacks and progress. A closer look at what changes—and what it means for people building software.
Building SellTrove's AI assistant required understanding vector embeddings and semantic search. Here is the practical architecture.
Converting trained ML models to ONNX format and running inference via ONNX Runtime simplified deployment and cut latency.
SellTrove's ML pricing engine delivered a 28% conversion uplift in 60 days. Nobody attributed it to the AI.
Building production LLM applications requires the same discipline as any other software engineering.
I built a production LLM agent cutting response time from 8 minutes to 30 seconds. An honest assessment of where LLMs win and fail.
Every ML model overfits during training at some point. The engineers who catch it before production know what to look for.
On the right relationship between ML models and domain experts — from credit scoring, fintech, and enterprise AI deployments.
The algorithm gets the credit. The features do the work.
I built a credit scoring model expecting brilliance. The metrics told the truth. On measuring ML model quality honestly.
Read new published articles in your RSS reader. No account needed.
Good software starts with better questions.
Explore my work