Service
Private LLM, RAG & Agents
Open-weight models, retrieval and agents running inside your network: evaluated, monitored and access-controlled.
A chatbot demo is easy. A private assistant your security team approves is not.
Most enterprise knowledge can't leave the network: contracts, engineering documents, customer records. Public AI APIs are often ruled out, and internal pilots stall on the same questions: which model, on what hardware, who may see which documents, and how do we know the answers are right?
We build the assistant inside your environment, connected to the systems where the knowledge lives, and hand it over with the evaluation and monitoring your team needs to run it.
You may need this if
- Your data can't be sent to a public AI API, or only after a long approval.
- A RAG pilot answers well on a demo set, but users don't trust it.
- Search results must respect the document permissions you already have.
- You want agents to take actions in internal systems, safely.
What we deliver
01
Model selection and serving
Open-weight models sized to your hardware, served behind an authenticated, OpenAI-compatible API.
02
Permission-aware retrieval
Ingestion from your document stores and databases, hybrid search and reranking, filtered by each user's existing access rights.
03
Evaluation you can repeat
A question set from your domain experts, automated scoring and a release gate, so changes don't silently make answers worse.
04
Agents with guardrails
Tool-using agents with read-only defaults, approval steps for actions, and audit logs.
05
Monitoring and feedback
Latency, cost, retrieval quality and user feedback on one dashboard your team owns.
06
Handoff
Source code, infrastructure-as-code, runbooks and training for the team that will run it.
Reference architecture
A reference design, not a fixed product. Discovery adapts it to your systems and constraints.
Users and channels
- Web chat
- Messenger bot
- Internal apps
Application and guardrails
- Assistant API
- Agent tools
- Input and output filters
Retrieval
- Hybrid search
- Reranker
- Permission filter
Models
- LLM server
- Embedding model
Your data
- Document stores
- Wikis and tickets
- Databases
Operations
- Evaluation set
- Monitoring
- Audit logs
- Single sign-on
How an engagement usually runs
- Discovery2 weeks
Use cases ranked by value and feasibility, data sources and permissions mapped, model and hardware options, and a go/no-go.
- Pilot6–8 weeks
One assistant in production for a real user group, with an evaluation set and monitoring.
- PodMonthly
More sources and use cases, agents, and a platform other teams can build on.
Constraints we design for
- Network separation (망분리) and air-gapped networks
- Offline installation from an internal registry, with no calls to external APIs.
- Personal and confidential data
- Data stays in your environment; logging follows PIPA, DPDP and your retention rules.
- GPU budget
- Model size and quantization chosen for the hardware you have or can buy.
- Security review
- A threat model and prompt-injection test results prepared for your security team.
Typical stack
- vLLM
- Open-weight LLMs
- PostgreSQL
- pgvector
- OpenSearch
- Kubernetes
- Python
- OpenTelemetry
We use the tools you already run where they fit. Names are for reference only.
Tell us about your systems
An engineer reads every message and replies within one business day.