Services
CodeBailey gathers scattered data, reads documents and turns them into knowledge, and puts multiple agents to work, owning the design, build and operation end to end.
How it flows
See how seven services fit together, from data collection to multi-agent systems, with tuning, private infrastructure and security beneath them. Adopt only the stages you need.
Service 1 / Data collection
Databases, ERP, groupware, the web, email, legacy systems: data piles up separately in every department. Collection agents gather it on a schedule and by rule, then keep syncing only what changed.
Data kept by each department differs in format and access rights, so people move it by hand and no one can be sure which version is current.
Connectors and collection agents link databases, ERP, the web and mail, sync only what changed, and log every pull so each record can be traced to its origin.
Connect databases, ERP, groupware, and SaaS through connectors and pull only what changed (CDC)
Collect portals, disclosures, and boards on schedule, filtering out duplicates and changes
Group mail and attachments by task and hand them off to the next step
Record what was collected, when, and from where, so the source stays traceable
From personal health record data exchange to real-time collection of production, quality and equipment data, these pipelines were built into systems in daily operation.
Built and operate a personal health record (PHR) MyData work system
Built a real-time environment for collecting and integrating production, quality, and equipment data
Service 2 / Document parsing
PDF, HWP, Office, and scanned documents become data while headings, tables, and paragraph structure stay intact. Then we pull the fields you need and classify and summarize them.
HWP files and scanned PDFs resist even basic text extraction, broken tables strip numbers of their meaning, and staff end up retyping key fields by hand.
Heading hierarchy and table structure survive conversion, scans are read with OCR, and fields such as contract amounts and deadlines are extracted and validated.
Converts documents while keeping heading hierarchy, table rows and columns, and even footnotes
Recognizes scanned documents, images, drawings, and handwriting
Pulls defined fields such as contract amount, deadline, and parties, and validates their format
Sorts document types and summarizes the key points so search and agents can put them to use
Service 3 / Embedding and RAG
We split documents into meaningful chunks, embed them, and store them in a vector database. Keyword and semantic search work together to find what matters, and every answer carries its source.
General LLMs don't know internal rules, answers without sources can't be used at work, and many organizations simply can't send documents to an outside cloud.
Documents pass through chunking, bge-m3 embedding, hybrid search and local LLM generation, and every sentence of the answer links back to its original source.
Gathers PDF, DOCX, HWP, PPTX, XLSX, and scanned images (OCR) while preserving tables and layout structure
Splits documents into meaning units and tunes parent-child structure and overlap to preserve context
Converts chunks into vectors with the multilingual embedding model bge-m3 (1024 dimensions)
Stores vectors in PostgreSQL's pgvector HNSW index, searching large collections of chunks in milliseconds
Fuses keyword search (BM25) with vector search using RRF, cutting down on missed documents
Feeds retrieved evidence in as context, and a local LLM generates the answer through Ollama
Attaches a source marker [n] to each sentence, and a click reveals the original evidence
Attention, cosine similarity, BM25, RRF and MMR: the math that governs retrieval and generation quality, and where each one is put to work.
F.01 · Transformer Core
Attention(Q, K, V) = softmax( QK⊤ ⁄ √dk ) · V
The core computation behind every LLM inference. Q·K⊤ scores are normalized by √dk to stabilize training, and FlashAttention with KV-cache cuts inference latency.
F.02 · Vector Search
sim(q, d) = E(q) · E(d) ⁄ ( ‖E(q)‖ · ‖E(d)‖ )
Finds documents that are semantically close in embedding space. HNSW indexes search large vector collections approximately, keeping responses to milliseconds.
F.03 · Lexical Score
BM25(D, Q) = Σi IDF(qi) · Ti Ti = tfi(k₁+1) ⁄ ( tfi + k₁(1−b+b·|D|⁄avgdl) )
A keyword-matching method built on TF-IDF. It pairs with vector search to catch signals such as domain abbreviations and identifiers that semantic embeddings tend to miss.
F.04 · Hybrid Fusion
RRF(d) = Σi 1 ⁄ ( k + ranki(d) ) k = 60 (standard heuristic)
Fuses results from different scoring scales using rank alone. Without normalization, it reliably combines BM25 and vector search results into a single ranking.
F.05 · Diversification
MMR = arg maxd [ λ·sim(d, q) − (1−λ)·maxj sim(d, dj) ]
Balances relevance and diversity when selecting passages. Removing duplicate chunks makes the context window more efficient and raises the answer's information density.
F.06 · RAG Eval
Faithfulness = | claims grounded in ctx | ⁄ | claims generated |
Measures hallucination numerically. Every claim in an answer is broken down and checked against how much of it is backed by the retrieved context.
A simulation that reproduces how the real product handles three screens: Q&A, bulk research and security.
Automatically finds the stakeholder perspectives on a topic.
Writes in-depth questions per perspective to widen the search.
Gathers evidence from internal documents with BM25 and vector search.
Designs the report structure from the evidence gathered.
Writes each section, citing its evidence.
Closes with a summary of the key points.
※ This demo is a simulation of how the real product, Bailey RAG, works.
LLM and RAG programs for KISTI, an LLMOps platform and Bailey RAG, the company's own air-gapped RAG platform, were all designed and built in-house.
An air-gapped research platform with document structuring, text, table, image and formula extraction, and a standard API linking LLM, VLM, RAG and agents (ongoing)
Training-data generation and integrated search for a science-specialized LLM, LLMOps platform features, and an AI agent development and deployment platform
Built an LLMOps platform to support large language model (LLM) development and operations
An installable RAG that answers from internal documents with citations, with access filters by document class and audit logs of queries and uploads
Service 4 / Multi-agent systems
Research, drafting, review, approval. We design agents with roles split like a human organization, hand them your internal systems as tools, and route important decisions through human approval.
A chatbot that only answers can't complete work that spans several systems, and it leaves no record of what the AI did or on what basis.
Each step gets its own agent and permissions, internal systems connect as MCP tools, and sensitive steps such as payments or external mail wait for approval.
Breaks work into steps and assigns each step's agent and permissions
Connects internal systems and APIs as MCP tools so agents can query and execute directly
Splits work across agents, reassigns it on failure, and collects the results
Routes important steps, such as amounts and external communications, through human review before proceeding
Two 2026 patent filings scale processing modules with load and, when one fails, hand its task queue and state to healthy modules so the work carries on.
Based on resource profiling in a shared state store (Application No. 10-2026-0110008)
Across multiple autonomous task-processing modules (Application No. 10-2026-0109943)
Service 5 / Fine-tuning
We build a domain dataset, adapt the model with LoRA and DPO, then score accuracy and faithfulness as numbers, comparing before and after on the same standard.
Off-the-shelf models misuse industry terms and document formats, prompts alone can't stabilize quality, and there is no way to tell whether a change actually helped.
Datasets come from your work documents, models are tuned with LoRA and DPO, and RAGAS metrics compare results before and after on the same yardstick.
Builds training and evaluation data from work documents and query logs
Trains only a small set of parameters, adapting the model even on a single GPU
Matches answer style using pairs of good and bad responses
Measures accuracy and faithfulness with RAGAS and compares before and after changes
LoRA trains only a tiny fraction of the parameters so tuning fits on a single GPU, while DPO aligns answer style directly from preference pairs with stable training.
F.07 · PEFT
W = W₀ + (α ⁄ r) · BA A ∈ ℝr×k, B ∈ ℝd×r, r ≪ d
Adapts a model to a domain by training 104× fewer parameters than full fine-tuning. Combined with QLoRA's NF4 quantization, it can fine-tune a 65B-class model on a single 48 GB GPU.
F.08 · Alignment
ℒDPO = − log σ ( β·Δw − β·Δl )
Δy = log π(y|x) ⁄ πref(y|x), y ∈ { yw, yl }
Aligns a model directly from preference pairs, without a reward model or PPO. It keeps training stable while staying compute-efficient.
An LLMOps platform supporting LLM development and operations, plus training-data generation and LLMOps capabilities for KISTI's KONI LLM.
Built an LLMOps platform to support large language model (LLM) development and operations
Service 6 / Private AI
We design on-premise model serving and GPU infrastructure so LLMs run even in network-segmented, air-gapped environments, mixing in public cloud when it's needed.
Public, financial and medical data can't leave the premises, and the burden of running GPUs and keeping up with model changes stalls adoption further.
Open-source LLMs run on internal servers with GPUs sized to actual use, and a hybrid setup keeps sensitive data inside while the rest runs on public cloud.
Serves open-source LLMs on internal servers through Ollama and similar runtimes
Designs GPU configuration and placement to match actual usage
Keeps sensitive data inside, and runs everything else on AWS, Azure, or GCP
Lets you swap the model alone, without changing the application
As an NVIDIA Inception member, CodeBailey runs its own air-gapped RAG platform, Bailey RAG, and deploys to separated networks with offline bundles and Kubernetes.
A global partnership for GPU-accelerated AI development
Designed, built, and operated in-house
Service 7 / Operations and security
We log what agents were asked and how they answered, block prompt attacks and personal-data leaks, and watch accuracy and cost every day.
Once live, prompt attacks, data leaks, unnoticed drops in answer quality and unpredictable costs become the real challenges.
Input and output guardrails, automatic masking of sensitive data, query and access audit logs, and monitoring of accuracy, latency and cost keep operational risk under control.
Blocks prompt injection and prohibited topics at both input and output
Automatically hides sensitive data like resident registration numbers and phone numbers
Records who asked what, and which documents they saw
Tracks accuracy, latency and cost, and alerts when any of them crosses a threshold
The National Tax Integrated System (NTIS) has been kept running without interruption, and a K-Health data safe zone for analyzing de-identified medical data was built with Konyang University Hospital.
Built the data safe zone portal for analyzing de-identified medical data securely
Ensured stable operations for a core national tax system
How we work
We assess your data and workflows, prove accuracy with numbers in a PoC, connect the result to internal systems, and stay on to run and secure it.
01
We look at your data and work to pick the problem agents should solve
02
We pick a core workflow and verify accuracy with numbers
03
We connect it to your internal systems and put it into real work
04
We run monitoring, evaluation, and security together with you
Contact
Start with a 30-minute consultation on turning your scattered data into agents, and see the impact yourself in a pilot within 2 weeks.