Technology
How we build with AI and automation
The technology we use in real projects, shown with diagrams. We design the principle first, then build with proven tools.
01AI development · MCP
Letting AI work with real systems
A language model on its own only writes text. Connected to your data and functions through MCP, it can look things up, process them and record the result.
- Request
- AI model
- MCP tool call
Functions become tools
Each capability AI may use is defined as a tool connected to your systems.
Same functions as people
Screens and AI tools call the same functions, so people can review what AI did.
Limited scope
What AI may call and see is restricted, and outside access is blocked.
02RAG · Retrieval-augmented generation
AI that answers from your documents
RAG first finds the documents relevant to a question and hands them to the AI. Instead of inventing an answer, it answers from evidence and shows the source.
Index
- Documents
- Chunking
- Embedding
- Qdrant
Query
- Question
- Embedding
- Similarity search
- AI model
- Answer + sources
Sources shown
Every answer carries the documents and passages it came from.
Permission-aware
Search is filtered by department or level; restricted documents never reach the answer.
Measured with a question set
Common questions and expected answers are kept and re-run after every change.
Keyword
- Matches on
- Same words
- Strong when
- Product codes · names
- Weak when
- Asked in other words
Vector
- Matches on
- Close meaning
- Strong when
- Asked in other words
- Weak when
- Exact numbers · codes
Hybrid
- Matches on
- Both, merged
- Strong when
- Most business documents
- Weak when
- Needs tuning
| Method | Matches on | Strong when | Weak when |
|---|---|---|---|
| Keyword | Same words | Product codes · names | Asked in other words |
| Vector | Close meaning | Asked in other words | Exact numbers · codes |
| Hybrid | Both, merged | Most business documents | Needs tuning |
03Chunking · Splitting documents
Search quality is decided at chunking
Chunking splits long documents into pieces the AI can find well. Too large and unrelated content mixes in; too small and the context breaks.
Slight overlap
Neighboring chunks overlap so meaning survives a cut.
Titles and sources attached
Each chunk stores its document title, section and date.
Tables and code kept whole
Tables and code blocks are never split in the middle.
Fixed size
- Splits by
- Characters · tokens
- Suits
- Uniform documents
- Watch for
- Cuts mid-sentence
Structural
- Splits by
- Headings · clauses
- Suits
- Rules · manuals · contracts
- Watch for
- Uneven lengths
Semantic
- Splits by
- Topic changes
- Suits
- Minutes · reports
- Watch for
- Costly to split
| Method | Splits by | Suits | Watch for |
|---|---|---|---|
| Fixed size | Characters · tokens | Uniform documents | Cuts mid-sentence |
| Structural | Headings · clauses | Rules · manuals · contracts | Uneven lengths |
| Semantic | Topic changes | Minutes · reports | Costly to split |
04Qdrant · Vector database
A database that searches by meaning
Qdrant stores text as numeric vectors and quickly finds what is closest in meaning. It can be self-hosted so your data never leaves your servers.
Nearest in meaning
Questions and documents share one vector space; results are ranked by distance.
Filtered by condition
Department, date and permission are stored alongside for filtered search.
Self-hosted
Installed with Docker on your own servers.
05n8n · Workflow automation
Automation that links business steps
n8n connects nodes — when, what, how and where to — into a working flow. It can run on your own server.
- Schedule
- Fetch data
- AI summary
- Branch
Runs on your server
Self-hosted, so business data stays in-house.
Alerts on failure
Failures reach your messenger; healthy runs stay quiet.
Flows backed up with keys
Without the encryption key, restored flows lose their credentials.
Suits
- n8n workflow
- Repetitive work between services
- Custom development
- Core work with complex rules
Speed
- n8n workflow
- Fast — connect nodes
- Custom development
- Needs design and build
Changes
- n8n workflow
- Edited on screen
- Custom development
- Code change and deploy
Limits
- n8n workflow
- Complex logic · bulk
- Custom development
- Higher initial cost
| n8n workflow | Custom development | |
|---|---|---|
| Suits | Repetitive work between services | Core work with complex rules |
| Speed | Fast — connect nodes | Needs design and build |
| Changes | Edited on screen | Code change and deploy |
| Limits | Complex logic · bulk | Higher initial cost |
06Video automation
Tools for repetition, people for the finish
Repetitive editing steps — scene splitting, caption drafts, clip cutting — are automated with our own tools; flow and finish are judged by people.
- Source
- Scene split
- Caption draft
- Cut · short
- Render
- Review
Frame-accurate cuts
Keyframe offsets are prevented and output length is measured.
Source quality kept
Needless conversion is avoided and the received quality is shown.
Tools bundled
External tools ship with fixed paths, so every PC gives the same result.
Reduced by swapping in lighter components (September 2026).
Stack
Built on proven, standard technology
Web
ReactNext.jsJavaSpring
Data
PostgreSQLQdrant
Infrastructure
DockerLinux
AI
ClaudeMCPPython
Automation
n8n
Video
FFmpegIn-house video tool