§03 / services / ai
AI Engineering
Applied LLMs that pass review.
We ship LLM systems where you can defend every output: typed schemas, evaluation harnesses, audit trails, on-prem when GDPR demands it.
FIG. 03.0ai engineering
§03.1/ capabilities
What we ship.
- 03.1Applied LLM features you can defend in review: typed schemas, audit trails, evaluation
- 03.2Agentic workflows that automate real operational tasks, not chat demos
- 03.3Retrieval pipelines that answer from your own data, with every claim traceable to source
- 03.4LLM-as-judge auditing and evaluation, drawn from our peer-reviewed research
- 03.5On-premise deployment of open-weight models for strict data-residency or GDPR needs
- 03.6Clean integration of AI features into existing JVM and TypeScript systems
§03.2/ stack
Tools in hand.
LangChainLangGraphOpenAIAnthropicLlamaQwenDeepSeekOllamavLLMPythonJavapgvectorQdrant
§03.3/ artifact
In production.
A sample of what an engagement actually looks like in code.
from langgraph.graph import StateGraph, END
from typing_extensions import TypedDict
class AuditState(TypedDict):
submission: str
rubric: list[Rubric]
scores: dict[str, float]
flags: list[Flag]
graph = StateGraph(AuditState)
graph.add_node("grade", grade_with_ensemble) # 4 local LLMs
graph.add_node("dissent", detect_disagreement) # unanimous == flag
graph.add_node("explain", qualitative_diff)
graph.set_entry_point("grade")
graph.add_edge("grade", "dissent")
graph.add_edge("dissent", "explain")
graph.add_edge("explain", END)
auditor = graph.compile() # GDPR-safe: no data leaves the box