Support for the design, controlled implementation, execution and evaluation of a pilot project for IBS

Description

The Inter-American Development Bank (IDB) calls on consulting firms to support ENCAT in the development of a detailed technical study for the elaboration, training and evaluation of two Proofs of Concept (PoCs) of Small and Specialized Language Models (SSLMs). The objective is to apply artificial intelligence in the tax classification of products and in the interpretative standardization of the Tax Reform (IBS) from 2026.

1. Technical dimension and AI architecture
Data methodology: Guidelines for XML file extraction, curation, cleaning, anonymization, and labeling of tax bases for training and data validation.

Specialized models: Evaluation of high-agility and low-cost techniques, such as semantic search (RAG) and lightweight fine-tuning (LoRA/QLoRA), measuring ROI compared to generalist models.

Infrastructure and governance: Recommendation of requirements for on-premise, cloud, or hybrid environments, ensuring data sovereignty and information security.

Technical training: Conducting alignment workshops and knowledge transfer on SSLM and LLM architectures to ENCAT's technical staff.

2. Functional, fiscal and economic analysis
Risk mitigation: Identification of risks associated with heterogeneity of interpretation between states and estimation of the economic impacts of classification errors.

Synergy with the reform: Alignment of SSLMs with the macro processes of the Tax Reform and mapping of potential new AI use cases for state tax authorities.

Feasibility study: Economic comparison of SSLMs against generalist LLMs and traditional technologies (regex, classic NLP) to estimate the economics of national adoption.

3. Pilot scope and exclusions
Discovery Phase: Conducting interviews to gather functional and legal requirements, in addition to defining metrics for evaluation and agreement of the PoC backlog.

Contractual limitations: Explicit exclusion from productive deployment, definitive system development, model training from scratch, definitive classification, or automated decisions.

Note: Interested companies should review the requirements on the BEO Bidders Portal.