How much does AI consulting cost?

BlokAI's engagements range from $5,000 to $75,000 depending on scope. Simple automations and LLM integrations start around $5K–$15K. Custom AI agent systems and full data infrastructure projects run $30K–$75K. We scope every project before committing to a price.

What does BlokAI build?

BlokAI builds AI agents, LLM applications, RAG systems, automation pipelines, and data infrastructure for startups and SMBs. We take projects from architecture to production — not demos, not prototypes.

Is BlokAI based in Japan?

Yes. BlokAI is a solo studio run by Kent Tran, based in Tokyo, Japan. We work with clients globally in English.

What is RAG and why does my business need it?

RAG (Retrieval-Augmented Generation) lets AI systems search your own documents, databases, or knowledge bases before answering — instead of relying only on training data. It powers AI customer support, internal search tools, and sales enablement systems that are accurate and up-to-date.

Llama 2 Is Here — Open Source AI Just Changed the Game for SMBs — BlokAI Blog

"Llama 2 is now available for free for research and commercial use. We believe open access is critical to the responsible development of AI." — Meta AI Blog, July 18, 2023

Meta just released Llama 2. For anyone paying attention to the AI infrastructure space, this is a big deal.

Let me explain why, and more importantly, what it actually means for businesses that are thinking about AI implementation.

What Is Llama 2?

Llama 2 is a family of large language models released by Meta on July 18, 2023, in partnership with Microsoft. Available in 7B, 13B, and 70B parameter sizes. Crucially: it's open source, free to download, and licensed for commercial use.

The previous version — Llama 1, released in February 2023 — was research-only and leaked in a way that made commercial use legally murky. Llama 2 fixes that. It's explicitly available for commercial applications, with some restrictions (primarily that if you're serving more than 700 million monthly active users, you need a separate license — not a problem for most businesses).

The performance on the 70B model is competitive with GPT-3.5 on many benchmarks. Not GPT-4 — but GPT-3.5, which is what's powering the free ChatGPT and a significant portion of production AI applications.

Why This Changes the Calculus for SMBs

Up until now, if you wanted to build a production AI application, you were choosing between:

OpenAI API — best quality, API key, send your data to their servers, pay per token
Anthropic Claude — strong alternative, similar model with different tradeoffs, same data-leaving-your-infrastructure issue
Azure OpenAI — OpenAI models on Microsoft infrastructure, better compliance story but still not on-premise
Self-hosted smaller models — various options existed but capability was limited

Llama 2 at the 70B level changes option 4. For the first time, "run the model yourself, on your own hardware or private cloud infrastructure" is a genuinely competitive option for a wide range of business use cases.

The Privacy Argument Is Now Real

Here's a conversation I've been having with clients for months: "We want to use AI, but we can't send our data to OpenAI."

For some businesses — healthcare, legal, finance, anything with sensitive IP — this is a hard constraint. Some Japanese enterprises in particular have been very conservative about data sovereignty concerns, and for good reason.

With Llama 2, you can now:

Run the model on your own AWS/Azure/GCP private infrastructure
Run it on-premise if you have the hardware
Build RAG systems where your documents never leave your environment
Fine-tune the model on your proprietary data without sharing that data with anyone

This isn't purely theoretical. We're already scoping projects where Llama 2 or a derivative is the right call specifically because data cannot leave the client's environment.

The Fine-Tuning Opportunity

Because Llama 2 is open weights (you get the actual model parameters, not just API access), you can fine-tune it. That means taking the base model and continuing to train it on your specific domain.

Fine-tuned Llama 2 models are already appearing for medical, legal, and coding applications. The same approach is available to any business with:

A clear domain where the base model underperforms
A dataset of examples showing the right outputs
Access to GPU compute for the fine-tuning process

The cost of fine-tuning is coming down rapidly. Running a fine-tuning job on a 7B parameter model is now achievable in hours on a single A100 GPU — roughly $50-100 in cloud compute for a meaningful training run.

What You Need to Run Llama 2

This is the honest part: running Llama 2 properly isn't plug-and-play.

Hardware requirements: The 70B model requires significant GPU memory — typically 4x 24GB GPUs minimum for inference. This is not a laptop project. Cloud deployment on something like a p4d instance on AWS or equivalent is the realistic path for most businesses. Serving infrastructure: You need a model serving layer (Hugging Face Text Generation Inference, vLLM, or similar) to handle API requests reliably at scale. More moving parts than just calling OpenAI's API. Quantisation: Reduced-precision versions of the model (Q4, Q5, Q8 quantisation) can run on less hardware with modest quality tradeoffs. The 7B model quantised can run on a single consumer GPU. Useful for experimentation and specific use cases. The tradeoff: More infrastructure complexity in exchange for data control, cost predictability, and model ownership. For the right use cases, this is absolutely worth it.

The Ecosystem Is Moving Quickly

Llama 2's release has kicked off a wave of derivative models and tooling. We're already seeing fine-tuned variants focused on specific domains, quantised versions optimised for different hardware configs, and improved inference optimisation.

By the end of this year, running a capable open-source model will be meaningfully more accessible than it is today. The tooling will catch up. The patterns will standardise.

Our Take: A Tool in the Toolbox, Not a Replacement

At BlokAI, we're not switching everything to Llama 2. For most general-purpose applications, GPT-4 still delivers better results for the complexity of reasoning involved. For cost-sensitive, high-volume applications where data privacy isn't a concern, GPT-3.5-turbo remains the right call on pure economics.

But Llama 2 fills a real gap: the privacy-sensitive, custom-deployment, self-hosted use case. For those projects, it went from "theoretically possible" to "actually viable" this month.

If data sovereignty is a constraint for your AI project, this changes what's possible.

Want to implement AI in your business? Book a 30-minute architecture call at blok-ai.com.