LegalTech Architect (ex-Litigator)
Appellate AI Tools · Litigation Risk · LLM Consensus
Former litigator with a decade of courtroom experience turned backend engineer and LegalTech architect. After 10 years of navigating the unpredictability of the court system, I transitioned into software development to build systems that replace legal intuition with computable data. A pragmatic developer (PHP for business logic, Go for high-load parsing, Python for LLM tooling and experimentation) who builds tools to solve the exact problems I faced in practice. Currently developing Lexometrica.com - a predictive analytics engine that evaluates the success probability of appellate court decisions. Recently published research exposing LLM "sycophancy" in legal analysis, proving that AI models artificially inflate a lawyer's chances of winning when fed biased arguments, and designed a multi-model consensus architecture to neutralize it.
An experimental B2B predictive justice system currently in the R&D phase. Designed to eliminate uncertainty in commercial arbitration, Lexometrica functions as high-load legal infrastructure. The core architecture (Go/Python/PostgreSQL) utilizes a multi-LLM consensus pipeline to parse complex judicial patterns, evaluate litigation risk at scale, and build dynamic AI profiles of individual judges based on their past decisions and behavioral tendencies. The goal is to shift legal strategy from subjective human intuition to a statistically calibrated, API-first risk assessment model.
A production B2C AI service [https://neshemyaka.ru] that predicts litigation and appellate risk for individual users and small businesses. It ingests raw case text, classifies case types, and returns a 0–100 risk score with a confidence score, along with structured risk factors and practical recommendations. Built on a Go backend with PostgreSQL and a multi-model LLM layer with fallbacks, the system is optimized for low COGS, robust JSON error handling, and fast UX (magic-link auth, HTMX frontend), making advanced legal risk analytics accessible without any enterprise integration.
Lexometrica Ground Truth is an independent LLM leaderboard built on a closed, static dataset of 30 highly complex cases derived from real Russian court practice. We discard standard memorization metrics to test real legal intelligence within the IRAC (Issue, Rule, Application, Conclusion) logical framework: evaluating the models' ability to identify hidden problems (issue-spotting), apply relevant norms to facts (rule-application), and draw accurate conclusions.
A few valuable lessons from LegalTech archaeology: long before today’s wave of generative AI, researchers were already modeling court systems. In 1968, they simulated the flow of felony defendants through the District of Columbia trial courts, and by 1973, others were proposing a largely algebraic method for simulating legal systems for joint use by engineers and lawyers. Main argument is simple: in law, we should not try to model the lawyer first - we should model the system. Those older papers still feel surprisingly modern because they focus on process, transitions, and structure rather than treating legal work as just text generation, which is exactly the distinction many current Legal AI products still struggle with...
A short reflection on how generative AI is destabilizing the traditional link between author, work, and ownership. Through Barthes, Thaler v. Perlmutter and Edmond de Belamy toward broader conclusions...
Habr
Article (in Russian) where I benchmarked Claude, GPT‑5.2 and Gemini on predicting appellate outcomes using real cases, measured Brier scores and calibration, and demonstrated how LLM sycophancy can dangerously inflate a lawyer’s perceived chances of winning.