ANKR Evolution Engine and Multi-LLM Competitive Routing: Self-Improving AI Systems Without GPU Infrastructure

Read full text Summarise with VIRAT
Authors: Anil Sharma
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This document is submitted as a defensive publication. No patent rights are asserted. The submission date establishes prior art for each mechanism described. We describe two interrelated innovations developed in the course of building AnkrStack — a production AI platform comprising 192 registered services operating across maritime, logistics, legal, and compliance domains. The first innovation is the ANKR Evolution Engine: a method for achieving continuous AI self-improvement using only API-based inference calls, with no GPU hardware, no labeled training dataset, and no ML engineering team. The core mechanism applies self-consistency scoring — running the same domain-specific prompt N times and computing the agreement rate p̂ across responses — to identify "frontier" tasks where the model's knowledge is neither fully mastered (p̂ > 0.8) nor beyond reach (p̂ < 0.3). Tasks falling in the frontier band (0.3 ≤ p̂ ≤ 0.8) are automatically captured as high-value training examples. An epoch-based improvement cycle generates fine-tuning datasets from production traffic at a measured cost of $0.0003 per improvement cycle — compared to $200,000+ for GPU-based reinforcement learning approaches. The second innovation is the Multi-LLM Competitive Routing system: a production-deployed architecture in which 15+ LLM providers are evaluated per request according to a strategy cascade (free_first → cheapest → quality → primary), with per-provider health monitoring, rate limit tracking, sliding-window request accounting, and automatic failover. A JudgeAgent pattern enables multiple providers to compete on identical queries with quality-based winner selection. Measured blended cost is $3,000/month at volume where single-provider premium routing would cost $90,000/month — a 97% reduction. Each mechanism is described with sufficient implementation detail, algorithmic specificity, and production evidence to constitute a complete prior art disclosure.

Fields of Study & Subjects

Artificial Intelligence Applications Topic Modeling Natural Language Processing Techniques

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
25 March 2026
Language
EN
Document type
Preprint