Lost Demand Intelligence — a search analytics engine built for e-commerce
Every "no results" event is classified before the user presses Enter. The engine separates typos from genuine B2B queries — queries that signal missing inventory, not broken UX.
When a user searches for a product that doesn't exist in the catalog, then buys an alternative anyway — that's a Gold Signal. Confirmed purchase intent on an untracked product. The highest-value training signal in the system.
The same engine runs on automotive parts and consumer electronics without retraining the underlying model — only the domain knowledge layer (brands, categories, feature extractor) gets rebuilt per vertical, not a zero-effort transfer. 91/100 correct classifications on the automotive domain — 91% accuracy on a new vertical.
Every query is stored as a labeled record: intent confidence, semantic validity flag, reward score, Gold/Positive/Negative classification. Ready to export as JSONL fine-tuning data.
Not raw logs. A structured PDF report: top missing products by search volume and estimated revenue impact, with prioritized restocking recommendations. Decision-ready output.
Python · Flask · SQLite · WebSocket · custom semantic validation pipeline · reward scoring engine · JSONL fine-tuning export · real-time debug dashboard. Built and maintained by one engineer.
Before LDI: built Skankran from scratch in 4 months with no prior programming experience — a water quality analysis platform integrated with Google Gemini API, trained on WHO/EPA medical documentation. First platform of its kind in Poland. That project is what led to this one.
Interested in the architecture or a conversation?
lukasz@adeptai.pl