Callicode | Enterprise AI that proves its answers

Work

Real cases.

For lawyers

Consultations on legal questions

Ask in plain words, get the answer with the clause it comes from. Precision comes first: tuned agentic and hybrid retrieval with answer quality control.

Under the hood

Agentic RAGHybrid RAGgrounding checksPythonFastAPIReactJavaSpringKafkaSeaweedFSQdrantOpenSearchMongoDBRedisPostgreSQL

For field engineers

Manuals, guides, product data sheets

Engineers who install equipment ask and get the answer with its source. The document base updates itself.

Under the hood

Hybrid RAGPythonFastAPIReactKafkaTemporalSeaweedFSQdrantOpenSearchMongoDBRedis

For government

A data lake that scores risk

About 100 TB from many sources in one air-gapped lake: a three-tier architecture with machine-learning pipelines that score compliance risk.

Under the hood

KafkaSpring Cloud Data FlowMongoDBPostgreSQLElasticsearchMinIORedisKubernetesTerraformArgoCDELKPyTorchtransformersCatBoostXGBoostPyCaretOptunascikit-learn

How I work

From the first sketch to a measured launch.

I design the architecture, build it with your team and answer for the result. Directly, or inside an integrator's team.

01
DiscoverUse cases, documents, constraints, and who answers for a wrong answer.Scope and risk map
02
ShapeArchitecture, estimate and the architecture part of the proposal.Target architecture
03
Measure firstA golden set of real questions, agreed before any code is written.Golden dataset
04
BuildRetrieval, agents and guardrails, deployed inside the perimeter.Working system
05
ProveEvery question scored and compared run to run until quality holds.Quality report
06
Hand overRunbooks, tracing and team training, from someone who taught architecture at university.Runbook and tracing

What you get

A working assistant in your networkPeople ask in plain words and see the document behind every answer.
A quality score before launchEvery test question is checked: right answer, right source, or not. Repeated on every release.
An architecture your team can ownDesign, decisions and the reasons behind them, written down.
A team that runs itRunbooks, tracing and training, handed over before I leave.

Discuss a project on LinkedIn →

Stack

What I build and what it runs on.

Direction

RAG

Answers from your documents, with the passage they came from.

Types: Hybrid RAGGraph RAGAgentic RAGSelf-RAGFederated RAGAdaptive RAGPathRAGCRAG
Ranking: reranker
Voice: speech recognitionspeech synthesis

Direction

Agentic AI

Agents that reason across your systems and follow one set of rules.

Agents: LangGraphSpring AILangChain
Tools: MCP
Workflows: TemporalKafkaRabbitMQ

Platform

OCR

VLM-based recognition for a heavy flow of PDF, DOCX and scans.

Recognition: VLM OCRPaddleOCR
Parsing: Docling
Labeling: Label Studio
Pipeline: TemporalKafka

Platform

Ingest

Turns your data into indexes RAG can search: chunked, embedded and linked.

Index: QdrantOpenSearchNeo4jBM25
Embeddings: sentence-transformers
Chunking: fixedsentenceparagraphstructure-aware
Graph communities: LeidenLouvain

Principle

Sovereign AI

Models and data stay inside your perimeter, down to air-gapped delivery.

On-prem models: vLLMOllamaLiteLLM
Platform: KubernetesTerraformArgoCDGitLab
Data: MinIOPostgreSQLMongoDBRedis
Access: Keycloak

Principle

Responsible AI

Guardrails, grounding checks and measured quality on every release.

Guardrails: NeMo GuardrailsGuardrails AI
Evaluation: RAGASDeepEvalCausa RAG
Tracing: LangfuseELK

Foundation

PythonFastAPIPyTorchtransformersscikit-learnCatBoostXGBoostJavaSpringReactAWS

Tools I use today. Models change every few months, so they are not listed.

Open source

My open-source projects.

Causa RAG

Shows which questions a change fixed, and which it broke. Runs on your machine.

github.com/laputski/causa-rag →

RAG World

A self-updating registry of RAG architectures, each with a maturity level from evidence.

ragworld.org →

Contact

Let's talk architecture.

Alexander Laputski

AI Architect

AI architect for enterprise and government. I build AI whose answers show their source, whose data stays in your network, and whose quality is measured before launch. Before AI, I spent years on distributed systems; most recently I delivered a sovereign AI platform with a team of 20.