# Konstantin Sofeikov

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Konstantin Sofeikov](https://cerebralvalley.ai/u/sofeikov)
- **GitHub:** https://github.com/sofeikov/motifql
- **Demo video:** https://www.youtube.com/watch?v=Ew2yWyLin58
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/327

MotifQL is a fuzzy graph search engine for time-ordered streams that lets compliance analysts describe complex financial crime patterns in plain English. Meta's Muse Spark translates human intent into a structured query, and our custom Python backend executes an anchored, backward path-matching engine to track down multi-hop laundering shapes across millions of streaming transactions.

How it Works & Technical Innovation

• AI Compiler & Self-Repair Loop: Built with Muse Code. Muse Spark compiles English to a custom DSL via a formal Lark grammar and typed AST. A programmatic loop feeds syntax exceptions directly back to the AI for automated self-correction.
• Fuzzy Traversal Engine: Replaces brittle boolean rules with continuous scoring (0.0–1.0). If a criminal subtly shifts an amount or introduces a time delay to evade exact thresholds, the engine degrades the score gracefully instead of killing the alert.
• Explainable Graph AI: Delivers a complete "witness" chain with a localized, human-readable narrative identifying the weakest link in the laundering chain.

Empirical Proof (IBM AMLworld Benchmark)

Tested against a strong LightGBM machine learning baseline at a strict real-world budget of 50 alerts/day:
• Cycles (Money loops): MotifQL caught 24 transfers vs LightGBM's 1.
• Layering Chains: MotifQL caught 13 transfers vs LightGBM's 0.
(Provenance: My transaction rule workspace, ezrules, and the IBM dataset existed prior to the hackathon. The entire MotifQL language, Lark compiler, fuzzy path-matcher, self-repair loop, and evaluation framework were built completely from scratch this week.)

## More from Meta Global AI Developer Hackathon

- [Naga Sai](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/324)
- [Apex Labs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/325)
- [BiG - M](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/326)
- [Buildebate](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/328)
- [Uniq minds](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/329)
- [Sisi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/330)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/327. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt