AI consulting

What the service covers

ByteInk designs and builds AI features that run inside the software a business already has. The work falls into four kinds of problem:

• Document retrieval (RAG). Search and question answering over your own documents, with every answer citing the passage it came from. Built on embeddings, a vector index such as pgvector in PostgreSQL, and a model instructed to answer only from what was retrieved.
• Agents. Models that call tools such as your APIs, databases and internal services, with scoped permissions, a log of every action, and a person approving any step that carries risk.
• Process automation. Extracting structured data from documents, emails and forms, and drafting replies or reports for a person to review.
• Classification. Routing tickets, tagging records and flagging items that need attention.

An AI feature also needs what keeps it working after launch: an evaluation set of real examples with known correct answers, logging of inputs and outputs, a fallback for low-confidence results, and a written record of where data goes.

Who it is for

The service fits a company with a working product or internal process and a specific task that people do by hand at a volume that keeps growing.

It fits a team that has built an AI pilot which works in a demo and has not made it into production.

It fits data that cannot leave the company. Open-weight models can run on your own servers or on the user's device, with no request sent to an outside provider. ThinkByte, ByteInk's iPhone assistant, runs its models entirely on the phone for the same reason.

If AI is one part of a new product, the work runs as product engineering. If the open question is whether an AI approach works at all, start with a prototype.

How an engagement runs

Assessment. We define the task in one sentence, name its owner and choose the number that should move. We review the data available and the privacy rules that apply to it, such as UAE data protection law. The output is a written brief with the use case, the metric and the recommended approach. Sometimes that approach is a rule or a template with no model at all.

Evaluation set. Before building, we collect real examples with their correct outputs. Every later change to prompts, retrieval or models is scored against this set.

Pilot. We build the smallest version that runs on real data inside your system. It starts with prompting a hosted or open model, and adds retrieval or fine-tuning only when the scores require it. The output is a working feature behind a flag, with its results against the evaluation set.

Production. The feature gets monitoring, cost tracking, fallbacks and human review points, then rolls out to all users. The output is the live feature and a runbook for provider outages and model changes.

After launch. Depending on the project, ByteInk keeps running and improving the feature, or hands it over with the evaluation set, the runbook and the code.

The approach is set out in more detail in Adding AI to an existing product and The real cost of "just add AI".

Technical scope

• Hosted models. OpenAI and Anthropic APIs, called through each provider's own SDK.
• Local models. Open-weight models served through Ollama, LM Studio or llama.cpp on your own hardware, or run on iPhone as in ThinkByte.
• Provider independence. Model calls sit behind a thin interface so the provider can change without a rewrite. aish, ByteInk's open-source AI shell assistant, runs against Ollama, LM Studio, OpenAI or Anthropic from one codebase.
• Speech. Speech-to-text with whisper.cpp behind an OpenAI-compatible API, as in voiced, a local voice gateway for Apple Silicon.
• Retrieval. Document parsing and chunking, embeddings, PostgreSQL with pgvector, and source citations in every answer.
• Integration. TypeScript and Go services, REST APIs and background job queues, deployed with Docker on Kubernetes.

More of ByteInk's public code is on the open source page.

Questions

What is RAG and when does it fit?

Retrieval-augmented generation answers questions from your own documents. Relevant passages are found first, and the model answers from them with a citation. It fits when the answers live in material a general model has never seen, such as policies, contracts or product manuals.

Can AI run without sending our data to OpenAI or Anthropic?

Yes. Open-weight models can run on your own servers or on the user's device, so the data never leaves your infrastructure. The trade-off is that smaller local models are less capable than the largest hosted ones.

Do we need to train our own model?

Usually not. Most business tasks are handled by an existing model with good prompting and retrieval. Fine-tuning pays off when you have both a task general models handle poorly and the data to teach it.

Where is ByteInk based?

Dubai, in the UAE. Work with teams outside the UAE runs remotely. To discuss an AI feature, get in touch.