Principles
AI systems that do not fail come from structure, not from prompts. These are the principles Almaneo applies to every build.
Hallucination is blocked by architecture
- Numbers come only from the calculation engine
the model does not calculate; it explains.
- Narrative comes only from source quotations
every sentence carries a citation.
- Without evidence, the field stays empty
a blank is better than a lie.
Patterns we apply repeatedly
Hybrid query router
numeric questions go to SQL, narrative questions go to retrieval augmentation, routed automatically
Single source for tool contracts
the schemas of the tools an agent uses are versioned in one place
Golden set regression gate
hundreds of verified question-answer pairs in CI, checked automatically on every change
PII masking pipeline
mask → model → reverse-map; originals never reach the model
Gateway abstraction
swap models or providers and the code stays the same
Human approval gate
at the end of automation, a person approves
Credit metering
measure usage per feature and convert to KRW
Model tier strategy
assign model tiers by task difficulty to control cost
Multimodal processing
documents, images, and audio in the same pipeline
Built in code, handed over as documentation
All infrastructure and pipelines are scripts that are safe to run again. Parameters are gathered in one configuration file, so a new project reuses the same skeleton by changing only that file. Load steps compare source row counts with loaded row counts and fail on mismatch. Points where manual work is unavoidable are documented as a gap list, so uncertainty in quotes and schedules is managed explicitly.
Proven in the Korean environment
Korean PII patterns, KRW billing, electronic tax invoices, public API integration — all with a track record in production.