Work
Five systems. One engineer.
Each project started as a problem no off-the-shelf tool could solve.
Products // shipped & running
Systems solving a concrete business problem — live and open to try.
Flagship · Live demo
LDI — Lost Demand Intelligence
An analytics engine that classifies every e-commerce search query before the user presses Enter. It separates typos from genuine B2B intent, and detects the Gold Signal — when a user searches for an unavailable product then buys an alternative anyway. That's confirmed purchase intent on untracked inventory. The highest-value training signal in the system.
Origin Story · Public
Skankran
Built from zero — before I could code. Poland's first water quality analysis platform: digitized municipal water station data, trained a conversational AI on WHO/EPA medical documentation, and connected it to the Gemini API for contextual health analysis. The project that started everything. If I handled government PDF chaos, your product catalog is trivial.
Memory & retrieval // the RAG research line
One thread: giving AI a memory that doesn't reset — and the tools to see inside it.
Innovation · Production VPS
ANIMA — RAG Memory Engine
A long-term memory system for AI companions. Three-channel RAG architecture: episodic memories, behavioral character core, external knowledge — all re-ranked by recency decay, semantic similarity, and keyword boost. Supersede logic rotates stale facts without losing history. Powers a personal AI companion running 24/7 on a private VPS with scheduled emotional insights.
Research · Groundwork for ANIMA
Context structure vs. model response
An early study that became the groundwork for ANIMA: keep the question fixed, vary only how the supplied context is structured, and compare what the model answers. The observation: the structure of the context alone changed the answers — which is why ANIMA treats prompt layout as part of the retrieval pipeline, not an afterthought.
New · Memory observability
Amnezja — RAG Debugger
An X-ray for AI memory. Type a query and watch every stage of retrieval — candidate pool, reranking, temporal filter, MMR, final prompt — exactly as the production model sees it. Plus a sandbox: how the model actually responds to that context, dry-run, without touching the live system. And time-travel — scrub the date forward to see what your agent will remember in 30 days.
Most RAG work asks "why didn't it retrieve this?". The more expensive question is "why was it never stored?" — and it stays invisible until you instrument the write path. So I did.
What it found in production
It surfaced what no top-line metric flagged: a milestone monoculture flooding every prompt. The detox that followed cut junk-memory creation 6.5 → 0.5 / day (−92%) over 1,751 audited entries, and brought memories per prompt from 2.0 down to 0.65.
On the write path it answered in two seconds what had puzzled me for weeks: the most important message of the year never entered memory at all, killed by a 4-word minimum. A month-wide audit put a number on it — 41% of messages left no trace, and of those stored, 52% carried an expiry date.
Then it caught a live one: a long-lived personal fact routed into a DATE bucket
that expires after 7 days, losing to the permanent FACT label by
0.01 of similarity. One hundredth of a cosine deciding whether the
system still remembers it next month.
Same tooling, same session: an acronym search returning zero of 60 matching entries, because the embedding model tokenises a three-letter project name into noise. Measured, then fixed with a lexical channel — 0 → 3 in the prompt.