How I implemented corporate AI for 15 employees and a continuous customer flow
One of my clients runs an AI platform that supports the daily work of a 15-person team and serves customers through the website and WhatsApp Business. The main model, Gemma 4 31B, is deployed on a rented GPU in the US. The infrastructure and API budget is $921 per month.
What matters to me here is that people use this continuously. An employee opens documents, discusses a task, comes back the next day and all that context needs to be there. Meanwhile customers keep coming in, choosing products and checking prices and availability. So I built the system around these two streams, with separate databases and memory, so each person gets the data they can access and can carry on working.
Company knowledge and each employee's memory
Employees use personal workspaces and Slack. The shared knowledge base contains company documents and policies, while personal memory holds each person's tasks, confirmed decisions and working preferences. Department, project and user permissions determine which materials enter the response.
Documents go through structured parsing, are divided into meaningful passages and are indexed for full-text and vector search. For each request the system checks permissions first, then retrieves and ranks the relevant sources. The answer uses the necessary evidence and gives the employee a link to the original document.
A changed file updates the corresponding records. A new decision in a conversation is saved with its source and owner. Employees can view, correct or delete memory. This keeps work continuous and makes the origin of each fact visible.
Customer assistance and current prices
The website and WhatsApp Business use a separate customer database. It receives approved materials, the catalog and operational data. The company's internal database stays in its own environment.
Product descriptions help retrieve suitable options, while dedicated functions obtain prices and availability from current records. When a customer asks for the cheapest suitable option, the system checks the required specifications first, then compares the total price. A price update changes the catalog record and its associated caches.
Website and WhatsApp conversations retain the customer's requirements and product selection history. A handoff to an operator is recorded in the same workflow, so the person picks up the work already in progress.
Workload routing and management analysis
Separate queues handle employees, customers and background updates. Gemma on the company's server performs the main workload. Peak traffic has its own external inference budget, and department heads and company leadership use a separate API route to advanced analytical models.
A manager selects the question and documents, reviews the outgoing package and starts the analysis. The system checks access permissions, accounts for the cost and saves the result in the corporate workspace.
Costs and control of data
The $921 monthly budget contains $749 for the GPU, $22 for the public server and backup storage, $50 for peak requests and $100 for external management analysis. The GPU is billed every three months at $2,247 per term. Implementation and human administration are separate cost lines.
I also worked through the API costs for 3,000 model calls a day, with 80% of input read from cache and the remaining 20% written into a short-lived cache. At those volumes, the application running on Sol or Sonnet costs $1,672.40 a month, including the application servers and management budget. Against $921, that's $751.40 saved each month, or 44.9%. Sol and Sonnet come out equal because their token tariffs match. The document shows the input, output and cache-write costs separately, along with the Opus and Astra comparisons.
The attached document contains the full implementation, diagrams, settings, prompts and calculations.