Internal knowledge with RAG
Assistants that retrieve policies, procedures and private documentation with semantic search and traceable sources.
Enterprise AI integration · Backend-first
I integrate AI into your systems without turning your backend into a black box.
I connect AI agents to APIs, documentation and internal processes through RAG, MCP and controlled tools — with security, permissions, auditability and an architecture your team can maintain.
Focused on integrations over .NET and Java architectures.
RAGMCP.NETJavaJWT / OAuthPostgreSQL + pgvectorTerraform
Problems worth automating
AI connected to the business, not an isolated chatbot.
Calling a model is the easy part. The real work is connecting it to real systems without losing control, security or traceability.
Agents with real tools
I connect models to APIs and services through MCP or function calling so they can query live data and execute controlled actions.
Permissions and auditability
Scopes, tool-level authorization, deterministic validation and action logs so the model is never the final authority.
Integration over your architecture
I do not replace what already works: AI capabilities are layered over existing backends while business logic stays outside the model.
Approach
The model reasons. Your software decides what may execute.
AI interprets intent and requests tools. The backend validates permissions, executes operations and keeps authority over data and business rules.
01 User
02 API / Agent Service
03 Model
04 RAG + MCP
05 Internal systems
Technical case study
AtlasSupply: an enterprise supply-chain assistant.
An end-to-end demo combining operational data, internal knowledge and actions protected by capability scopes. The agent decides when to use RAG, MCP or both.
- Multilingual RAG on PostgreSQL + pgvector.
- Remote MCP for suppliers, orders and incidents.
- JWT scopes that filter and authorize agent tools.
- Persistent audit logs for allowed and denied tool calls.
- Angular + .NET + OpenAI with a clear split between intelligence and authority.
How I work
Start small, validate quickly, harden afterwards.
Discovery
Understand the process, systems involved and where AI can create measurable value.
Scoped PoC
Build the smallest useful case over real data and APIs to validate value, cost and risk.
Integration
Connect tools, knowledge, permissions, observability and user experience.
Hardening
Security, auditability, tests, deployment and iteration until the solution is maintainable.
POLICYdefault-denyTool execution requires explicit capability scope.
orders.delayed.readALLOW
incidents.createDENY
Enterprise AI ≠ unrestricted access
Security in the architecture, not patched on afterwards.
The model does not receive credentials or direct database access. It requests capabilities; the backend decides what is allowed and executes each action under deterministic rules.
- Authentication and capability scopes.
- Tools filtered before reaching the model.
- Authorization validated again before execution.
- Audit trail of actions and outcomes.
- Secrets outside source code and environment-specific configuration.
When it makes sense to talk
A strong fit when you already have systems and data AI should be able to use.
.NET / Java backends with existing APIs.
Teams with scattered or difficult-to-search documentation.
Internal processes that require finding data and then acting on it.
AI PoCs that need to become secure integrations rather than demos.
FAQ
No hype: what I do and what I do not do.
Do I need to replace my backend to integrate an agent?+
No. The approach is to expose controlled capabilities over the existing architecture and keep business logic in software rather than in prompts.
Do RAG and MCP solve the same problem?+
No. RAG fits relatively stable knowledge; MCP/tools fit live data and actions. Real systems often benefit from combining them.
Does the model get direct access to my database?+
It should not. The model requests tools; the backend validates and executes them, returning only the result it needs.
Can we start with a small PoC?+
Yes. A concrete workflow, one knowledge source and a small tool set is the right way to validate value before expanding scope.
Do you only work with OpenAI?+
No. The model provider should be abstracted so the business is not tightly coupled to a single vendor.
Do you have a process that could benefit from AI connected to your systems?
We can start with a 20-minute conversation.
No endless sales deck: tell me the context, we identify whether there is a real use case, and decide what a sensible next step looks like.