GembaShift Inc. (Headquarters: Takasaki City, Gunma Prefecture, Representative Director: Masataka Mayuzumi) has launched "SpecLift," an AI-powered core analysis platform that uses AI and static analysis to decipher existing black-box systems and generate current specifications, business rules, dependencies, impact scope, and AI modernization data with "evidence." The service became available on July 2, 2026.

In large enterprises across finance, manufacturing, public services, and telecommunications, core systems that have undergone decades of modifications and expansions form the backbone of their operations. However, these systems have reached a state where "no one can accurately explain the entire system anymore." The original developers have retired, design documents no longer match the actual system, and even when attempting modernization with AI, there is no "accurate current input to provide." This issue, once warned as the "2025 cliff," intensifies with each postponement of action, becoming a critical management challenge for many companies today that requires immediate attention.

SpecLift addresses the problem of "not understanding the current system" by visualizing the overall current state, hidden business rules, dependencies, and impact scope from source code, and generating the necessary materials for AI modernization with "evidence." Its most significant feature is its ability to publicly disclose the accuracy of its deciphering by comparing the generated current specifications and business rules against a separate set of ground truth data, independent of the analysis target (ground truth data is frozen beforehand, and self-scoring is not performed). In this release, this non-self-scoring verification will be referred to as "independent verification."

AI Core Analysis Platform "SpecLift" - Analysis Accuracy Verified on Banking Core Systems

Analysis accuracy measured on the entire OSS "Apache Fineract" (5,274 files) for a real banking core system. API specifications: 94.5% / Data structure: 83% / Business rules: 82% / Calculation logic: 77%. The core 3 functions (interest, repayment, fees) achieved 100%. This independent verification was conducted on the most unforgiving financial subject matter to ensure both depth and breadth.

Service Site: SpecLift

Background of Development: Wanting to Modernize with AI, but "Not Understanding the Current System"

Large-scale legacy core system modernization projects often fail not due to a lack of technical skill, but because "no one within the company can accurately explain what the current system does." The following three fears keep decision-makers awake at night:

No one understands the overall picture (original developers are gone, design documents don't match the actual system).

Differences in specifications are only discovered during production operation (accidents occur due to missing hidden business rules).

Budgets and timelines are unpredictable (estimating without knowing what exists where).

Generative AI is powerful, but modernization cannot begin without "accurate current input." Providing inaccurate materials will lead to rework without capturing hidden business rules, increasing the risk of accidents. GembaShift emphasizes designing AI not just as an efficiency tool, but as a platform that generates "verified facts" suitable for decision-making. SpecLift was developed to fill this gap of missing "current materials" with evidence, and to verify and publicly disclose its accuracy with numbers.

What is SpecLift: An AI Core Analysis Platform that Visualizes the "Current State" Simply by Inputting Source Code

SpecLift is an AI Modernization Readiness Platform that automatically generates the overall current state, current specifications, hidden business rules, dependencies, impact scope, and data for AI modernization simply by inputting the source code of existing systems (design documents are not required).

The process involves two stages: "machine solidifies facts, AI assigns meaning." First, static analysis is used to graph the source code structure and solidify the facts, after which AI linguistically interprets the business meaning. All outputs are enforced with "file:line" (file name, line number) evidence, eliminating unsubstantiated assertions. Unclear parts are honestly indicated as "requires review," making the design resistant to "plausible falsehoods."

Source Code Input → SpecLift → 6 Deliverables. Evidence of file:line for all outputs

Simply by inputting source code (design documents not required), SpecLift automatically generates a "System Map, Current Specifications Document, Hidden Business Rules, Impact Analysis, Data for AI Modernization, and Reliability Score." The process involves two stages: "machine solidifies facts, AI assigns meaning," with evidence of file:line (file name, line number) attached to all outputs.

Key Features of SpecLift

1. Independently Verifies Analysis Accuracy and "Discloses with Numbers"

SpecLift's most significant feature is its ability to compare the accuracy of AI analysis with independent ground truth data and disclose it numerically. All AI outputs are enforced with evidence (file:line), and unclear parts are marked as "requires review." This is backed by evidence that it is not self-scoring. Adhering to the principle that "unverified accuracy is not accuracy" is crucial for passing engineer reviews in the core system domain, where AI distrust is high.

2. Verified on a Real Banking Core System (The Heart of Finance)

The subject chosen was Apache Fineract, a real core system for banking and microfinance (5,274 source files). Specifically, the most complex financial calculations for interest, repayment, and fees were targeted, and the entire system was measured comprehensively (no cherry-picking of convenient parts). The results demonstrated the ability to decipher the entire system with an accuracy of 80-94% for major items, and 100% for business rules, calculation formulas, and API contracts when focusing on the core 3 functions (with 96.8% accuracy for the quality of calculation formulas).

3. Supports 6 Stacks Including COBOL Mainframes

The supported languages and technology stacks are 6 types: Java/Spring, Python, C#/.NET, PL/SQL, PHP, and COBOL+JCL, all of which have been verified with real open-source software. Even mainframe assets (COBOL/JCL), which have the highest modernization costs, are visualized with evidence solely from source code (including PIC type and digits of fixed-length records, and business states at the 88-level). The 4 stacks of C#/.NET, PL/SQL, PHP, and COBOL+JCL alone completed analysis of 12,306 files, approximately 1.2 million lines, with zero errors and achieved 100% accuracy measured against independent ground truth data.

Supports 6 Stacks Including COBOL Mainframes. Verified with Real OSS

Verified 6 stacks (Java/Spring, Python, C#/.NET, PL/SQL, PHP, COBOL+JCL) with real OSS. Mainframe assets in COBOL/JCL, a difficult area for modernization, were visualized down to fixed-length layouts and business states. Analysis of approximately 1.2 million lines across 4 stacks was completed with zero errors.

4. Secure Design for Confidential Sources

All containers can be deployed entirely on-premises/VPC (private cloud), and data remains within the environment. Operation is read-only (no writing to production DB, no automatic deployment, no automatic code modification). Only masked evidence excerpts are sent externally. LLMs can be replaced with customer-managed ones, and customer data is not used for model training. Tenant isolation (RLS) and audit logs with tampering detection are also implemented, making it suitable for the strictest security and compliance requirements in the financial sector.

5. Generates "Current Materials" for AI Modernization with Evidence

SpecLift generates a System Map, Current Specifications Document, Hidden Specification Catalog, Impact Analysis, and Context Bundles/Work Orders directly usable for AI modernization. Furthermore, it outputs these as Mermaid diagrams, narrative explanations, and automatically generated digest PDFs with diagrams. It serves as a breakthrough for the "pre-modernization" phase when companies want to modernize with AI but cannot proceed.

6. States What It Cannot Do (Honesty = Brand)

It measures the entire system, not just convenient parts, and discloses areas of weakness numerically. It does not pretend to be 100% accurate and clearly indicates uncertain parts as "requires confirmation." Paradoxically, this honesty is believed to be the most effective way to gain the trust of decision-makers in large enterprises with strong AI skepticism.

Numbers Verified by Independent Validation (All Measured Values)

Real Banking Core System Apache Fineract (5,274 files): API specifications 94.5% / Data structure 83% / Business rules accuracy 82%, coverage 81% / Calculation logic 77%

Core 3 Functions (Interest, Repayment, Fees): Business rules, calculation formulas, API contracts all 100% (Calculation formula quality 96.8%)

Real Legacy Shopizer with Zero Design Documents (1,339 files / 124,756 lines / approx. 42,000 nodes): 78 tables / approx. 188 APIs automatically mapped, hidden specification top-tier true positive rate precision=1.00

All 6 stacks verified with real OSS. Among them, 4 stacks (C#/.NET, PL-SQL, PHP, COBOL+JCL) completed analysis of a total of 12,306 files, approximately 1.2 million lines, with zero errors, achieving 100% exact match for table and program extraction. Java achieved 80-94% on banking core Fineract, and Python was also verified with real OSS.

* The figures presented are measured values from independent verification using public sources such as Apache Fineract/Shopizer. Ground truth data prepared separately from AI output was used for scoring, and this was frozen during verification (criteria not changed retroactively). AI self-scoring was not performed. Overall accuracy of business rules is an estimation based on statistical samples (precision/recall extraction judgment) due to the large scale, while core functions were manually cross-checked entirely. Accuracy may vary depending on the system type and condition.

Anticipated Use Cases

Financial Institutions such as Banks, Insurance Companies, and Securities Firms

Visualize current specifications of accounting and insurance calculation systems with complex business logic like interest, repayment, and fees, with evidence. Can be used as a basis for decisions on modernization, system integration, and regulatory compliance.

Large Enterprises in Manufacturing, Distribution, Telecommunications, etc.

Inventory of overall maps and hidden business rules of core systems (production management, inventory, order processing, billing) that have undergone years of modifications. Can be used for prerequisite surveys for cloud migration and package replacement.

Public Sector and Mainframe-Holding Organizations

Visualize mainframe assets in COBOL/JCL solely from source code, down to fixed-length layouts and business states. Can be used for planning the modernization of core systems facing risks of vendor lock-in and aging technical staff.

Major System Integrators and Maintenance Vendors

Generate initial drafts for current investigation, impact analysis, test design, and migration diagnostics in tens of minutes to hours (time required varies depending on the scale and condition of the target). Creates a situation where even junior engineers can contribute to large legacy projects with evidence, which were previously reliant on individual expertise.

Differences from Conventional Approaches

Compared to manual current investigation (weeks to months, person-dependent, prone to omissions), SpecLift offers automatic generation in tens of minutes to hours with references (time varies by scale).

Compared to pasting code into raw LLMs (no references, fabricated content, not suitable for large scale), SpecLift enforces Evidence (file:line), discloses accuracy, and graphs the entire system.

Compared to static analysis tools (which show structure but not business rules or specification text), SpecLift generates structure + AI interpretation + business rule discovery + modernization materials.

Compared to other AI analyses (disclosing accuracy numerically through verification is not yet common), SpecLift performs answer checking, discloses accuracy numerically, and reveals areas of weakness.

To our knowledge, legacy analysis services that "perform answer checking and disclose accuracy numerically" are still few. SpecLift aims to be one of these few.

Pricing and Trials

SpecLift is offered on a custom quote basis depending on the scale, stack, and usage scope of the target system. The following evaluation steps are available for consideration:

Free Smoke Test: Reports on "how accurately it could read" for your source code (small portion acceptable), with numbers from independent verification. Areas that could not be read are also honestly disclosed.

Small-Scale Start: Can begin with one function or one subsystem at a time.

Document Provision: We provide materials including a feature list, security specifications, and sample deliverables (actual analysis digests).

Detailed pricing, contract periods, and on-premises configurations will be provided upon individual inquiry.

What SpecLift Aims For

SpecLift aims for more than just creating AI-generated documents.

The success or failure of legacy modernization is determined by the accuracy of the "current materials." SpecLift provides "verified current materials" suitable for decision-making by solidifying facts with machines, assigning meaning with AI, and independently verifying and disclosing the accuracy with numbers. By adhering to the principle of "stating what cannot be done," it aims to move AI modernization from the "trial" phase to the "actually implementable" phase.

GembaShift's Application-Specific AI Services and Support System

In addition to SpecLift, GembaShift offers application-specific AI services tailored to the business challenges of each company.

Gemba Book: AI for generating illustrated work standards, one page per step, from photos and existing documents. Supports multiple languages, approvals, version management, and audit logs.

SecureCanvas AI: A secure AI document generation platform for company-wide use, featuring Entra ID authentication, automatic PII removal, and audit logs.

BeforeAfter.ai: Visual AI for proposals that generates post-proposal visuals from on-site photos.

GembaShift goes beyond providing individual tools by offering end-to-end support, including inventorying business flows, designing ROI, implementing AI workflows, ensuring on-site adoption, and supporting in-house development, thereby advancing companies' AI utilization from the "trial" stage to the "business transformation" stage.

Representative's Comment

Masataka Mayuzumi, Representative Director, GembaShift Inc.

The biggest reason for legacy modernization failures is not technical skill. It is that "no one within the company can accurately explain what the current system does."

Even in an era of rebuilding with AI, if the "current materials" provided are inaccurate, accidents will occur due to missing hidden business rules. SpecLift fills that gap with evidence and has independently verified and publicly disclosed its accuracy with numbers using an OSS for banking core systems (Apache Fineract) as a target.

We adhere to the principle of "stating what cannot be done." We believe that unverified accuracy is not accuracy. We encourage you to first see "how accurately we can read" your system with numbers through our free smoke test.

Regarding Free Smoke Tests and Information Requests

SpecLift offers a free smoke test where we accept your source code in small units and report on "how accurately it could be read" with numbers from independent verification. We also provide information including a feature list, security specifications, and actual analysis digests (sample deliverables).

We invite inquiries from representatives of companies in finance, manufacturing, distribution, public services, and telecommunications facing challenges in modernizing, migrating, or integrating black-box core systems, as well as from representatives of system integrators and maintenance vendors handling large legacy projects.

Service Site: SpecLift

Inquiries: GembaShift Inquiry Form

About GembaShift Inc.

GembaShift Inc., headquartered in Takasaki City, Gunma Prefecture, is a company dedicated to building on-site-driven AI infrastructure. We provide end-to-end support, including inventorying business flows, designing ROI, implementing AI workflows, ensuring on-site adoption, and supporting in-house development, creating systems that are not just concepts but continuously operate in the field.

Company Name: GembaShift Inc.

Representative: Masataka Mayuzumi, Representative Director

Location: 489-11 Kami-namie-cho, Takasaki City, Gunma Prefecture, 370-0801

Established: January 21, 2026

Capital: ¥1,000,000

Business Activities: AI Utilization Consulting, AI System Development, Corporate Training

Official Website: GembaShift

Inquiries Regarding This Matter

GembaShift Inc.

Inquiry Form: GembaShift Inquiry Form

FACT BOX

  • Source: PR TIMES
  • Category: 製品リリース
  • Organizations: Apache Fineract / Shopizer