OFFAS

An AI application meant to assist office workers and also disclose information base on their RBAC, to avoid disclosing classified info to lower ranked employees in accordance with uploaded data

Comment

Inspiration

Inspiration

Artificial Intelligence has made remarkable progress, but for many African businesses, it remains inaccessible. Most AI-powered customer support platforms depend on cloud infrastructure, continuous internet connectivity, and recurring subscription fees. While these assumptions may work for large organizations, they create significant barriers for small and medium-sized enterprises (SMEs), which form the backbone of Africa's economy.

During our research, we found that businesses such as mobile money agents, microfinance institutions, cooperatives, and local retailers still spend considerable time answering repetitive customer questions. Information such as transaction limits, account registration procedures, dispute resolution, and fraud prevention already exists in documents and manuals, but finding the right information quickly remains difficult. Existing AI solutions either require sending sensitive business data to third-party cloud providers or incur ongoing operational costs that many organizations cannot justify.

The Africa Deep Tech Challenge 2026 inspired us to approach this problem differently. Instead of asking "How can we build another chatbot?", we asked:

Can enterprise-quality AI run entirely on a standard laptop without internet access, cloud services, or API costs?

That question became the foundation of OFFAS AI (Offline AI Support Assistant).

Our mission was to prove that intelligent customer support should not depend on expensive infrastructure. By combining modern Retrieval-Augmented Generation (RAG), lightweight embedding models, and quantized language models, we set out to build a practical AI assistant capable of operating completely offline while remaining fast, reliable, and affordable.

What it does

OFFAS AI is a fully offline, enterprise-ready customer support and knowledge assistant designed to run on the ADTC Standard Laptop profile (8 GB RAM, integrated graphics, Ubuntu 22.04).

Organizations can upload their internal knowledge base—including FAQs, policy documents, product manuals, operating procedures, and customer support documentation—and immediately begin asking natural language questions.

When a user submits a query, OFFAS AI:

  1. Searches a local vector database for the most relevant document passages.
  2. Retrieves the most useful context.
  3. Uses a locally hosted quantized Large Language Model (LLM) to generate an accurate response.
  4. Returns a grounded answer together with references to the source documents used.

Because every component runs locally, OFFAS AI requires:

  • No internet connection
  • No cloud infrastructure
  • No API keys
  • No subscription fees
  • No external data transmission

This makes it particularly suitable for organizations operating in bandwidth-constrained environments, privacy-sensitive industries, or locations where reliable internet access cannot be guaranteed.

How we built it

From the beginning, we wanted OFFAS AI to be more than a proof of concept. We wanted to build it like an enterprise software product.

To achieve this, we adopted Clean Architecture, separating the application into independent layers:

  • Domain Layer – business entities and interfaces
  • Application Layer – core use cases such as document ingestion and question answering
  • Infrastructure Layer – integrations with llama.cpp, SQLite, and sentence-transformers
  • Presentation Layer – FastAPI endpoints exposing the application

This dependency-inverted architecture allows us to replace the language model, embedding model, or vector database without changing the business logic, making the system flexible and maintainable.

Our Retrieval-Augmented Generation (RAG) pipeline works as follows:

  1. Enterprise documents are processed and divided into semantic chunks.
  2. Each chunk is converted into an embedding using a lightweight embedding model.
  3. Embeddings are stored in a local SQLite-based vector store.
  4. User queries are embedded and matched against stored document vectors.
  5. The most relevant document chunks are retrieved.
  6. These retrieved passages are supplied as context to a locally hosted quantized LLM.
  7. The model generates a grounded response with citations to the original knowledge base.

To satisfy the competition's hardware constraints, we carefully selected lightweight models and optimized inference for CPU-only execution. Environment-variable configuration allows deployment across different systems while preserving sensible defaults for the ADTC hardware profile.

We also invested heavily in software quality. Application-layer logic is tested against in-memory implementations of domain interfaces, allowing our unit tests to execute in milliseconds without requiring AI model weights or GPU resources.

Challenges we ran into

Building an enterprise AI application under strict hardware constraints presented several technical challenges.

Operating entirely offline

Many modern AI frameworks assume internet connectivity for model downloads, hosted APIs, telemetry, or cloud inference. Ensuring that every stage of the application—from document ingestion to answer generation—worked entirely offline required careful evaluation of open-source tooling and significant configuration effort.

Working within an 8 GB memory budget

Running an LLM locally on commodity hardware meant every megabyte mattered. We evaluated multiple quantized models before selecting one that balanced response quality, inference speed, and memory consumption.

Balancing accuracy and performance

Retrieving more document chunks often improved answer quality but increased latency and memory usage. We repeatedly benchmarked different retrieval configurations to find an optimal balance between responsiveness and reliability.

Preventing hallucinations

Enterprise users need trustworthy answers. Instead of allowing the model to rely solely on its pretrained knowledge, we grounded every response using Retrieval-Augmented Generation so that answers remain tied to the organization's own documentation whenever possible.

Building for maintainability

Introducing Clean Architecture increased the initial complexity of development, but it resulted in a system that is modular, testable, and easier to extend as new enterprise requirements emerge.

Accomplishments that we're proud of

We're particularly proud that OFFAS AI demonstrates that practical enterprise AI does not require cloud infrastructure.

Some achievements include:

  • Successfully building a customer support assistant that operates 100% offline.
  • Running entirely on the ADTC Standard Laptop profile without requiring a dedicated GPU.
  • Designing a modular architecture where major infrastructure components can be replaced independently.
  • Implementing a complete Retrieval-Augmented Generation pipeline using only local resources.
  • Returning grounded responses supported by source citations to improve user trust.
  • Creating an open-source project that others can study, extend, and deploy.

Beyond the technical implementation, we're proud that the project addresses a genuine need faced by many African businesses: affordable AI that respects privacy while remaining accessible.

What we learned

This project reinforced that successful AI systems are not defined solely by the intelligence of the language model.

We learned that:

  • Careful software architecture is essential for building maintainable AI applications.
  • Retrieval quality often has a greater impact on answer accuracy than simply using a larger model.
  • Quantized models make practical offline AI possible on ordinary hardware.
  • Benchmarking and optimization are as important as feature development.
  • Building for constrained environments encourages better engineering decisions and produces solutions that are more accessible to real users.

Most importantly, we learned that AI can be both powerful and practical without relying on cloud infrastructure.

What's next for OFFAS

OFFAS AI represents the first step toward a broader offline enterprise knowledge platform.

Our roadmap includes:

  • Permission-aware document retrieval using role-based access control.
  • Integration with enterprise identity providers such as Active Directory and LDAP.
  • Support for additional enterprise document formats.
  • Hybrid semantic and keyword search.
  • Multilingual customer support for African languages.
  • Administrative dashboards for managing users, documents, and system analytics.
  • Improved benchmarking and automated evaluation of retrieval quality.
  • One-click deployment for organizations without dedicated IT teams.

Our long-term vision is to make secure, affordable, enterprise-grade AI available to organizations across Africa without requiring cloud services, recurring subscription costs, or high-end hardware.

We believe intelligent customer support should be accessible to every business—not just those with the largest budgets.

Share this project:

Updates