A private AI knowledge engine running on the client's own GPU
US
On-prem / self-hosted
Security & compliance
RAG / knowledge engine
.png)
On-premise
Self-hosted, zero data egress
Architect-led
Discovery to production
2025-now
Partnership ongoing
Business context
A US property management firm with 20+ years in market was losing ground at industry roadshows. Buyers kept asking the same question — "do you have AI?" — and competitors were already demoing it. The CEO had a dozen ideas. He needed one shipped.
.avif)
“Most AI ideas die in R&D. The ones that survive have three things: clean evaluation criteria, a data set you can actually trust, and a user whose workflow changes if it works. We picked the one that had all three."
.avif)
The constraint that
shaped the stack
The client's data couldn't leave their infrastructure. Property records, financial reports, client documents, none of it could touch a cloud API. On-prem rewrites the stack: self-hosted models, local embeddings, parsing pipelines built for whatever the client throws at them
Privacy
Sovereignty
On-premise only. No cloud, no external APIs.
Parcing
Adaptability
PDFs, DOCX, JSON, and scanned images handled.
Consistency
Personalisation
Answers tuned to each user's terminology.
Discovery to production, in five phases
A partner-level Solution Architect ran the project: in the discovery room, in the standups, in the production decisions. Certified engineers executed alongside
01
Discovery
One workshop. A dozen AI ideas on the table. We left with one: a RAG-powered knowledge engine. The criteria were strict: high impact for business users, fast to validate, viable fully on-prem.
02
Proof of Concept
The hard problem wasn't the LLM. It was getting messy real estate documents into a structure a model could reason over. Our Solution Architect and full-stack engineer tested parsing pipelines and model combinations against the client's actual data, not synthetic samples.
03
MVP
Concept proven, we built the system the business users would actually touch. Document ingestion, semantic search, role-based access, lifecycle controls, all running on the client's own GPU. This became the foundation of the production engine.
04
Expansion
The MVP proved the pattern. We extended it to financial reports — balance sheets, income statements, liability breakdowns — and the engine moved from a single use case to a platform managers rely on day to day.
05
Production at scale
The engagement is ongoing. Our engineer continues under the Solution Architect's coordination, tightening retrieval accuracy, expanding report coverage, and prepping the engine for company-wide rollout.
Results achieved with Radency’s AI team
3 months for PoC and MVP
A small team delivered a working on-premise knowledge engine in just three months of active development + R&D.
~99% faster data lookups
Finding a figure in a 100-page report could take up to 30 minutes. Now semantic search retrieves it instantly, with the local LLM returning answers in seconds.
100% privacy compliance
All models, embeddings, and rerankers run on the client’s own GPU server, with no data leaving their infrastructure. Expiration rules automatically clear old records.
10 report types covered
Beyond PDFs, the engine now processes ~10 financial report formats, giving managers instant access to data once buried in manual reviews.
“With the RAG engine, managers can get answers in seconds instead of spending half an hour scrolling through reports.”
.avif)
Max Honcharuk
Partner & Solution Architect at Radency
A working product, not a demo
The knowledge engine was built as a fully on-premise RAG system, keeping every document inside the client’s infrastructure. It connects to the client's platform via standard REST APIs. Our team continues to refine AI engine's accuracy by improving parsing pipelines and fine-tuning the model
.avif)
Key features shipped
01
On-premise AI stack
Everything runs locally. Embeddings, reranking, and generation sit on a self-hosted Llama 3.3 model. No data leaves the servers, which keeps it compliant and cheaper than running cloud LLMs.
02
Document parsing pipelines
Different processing pipelines handle different file types: PDFs, tables, scans, multi-report packets. OCR cleans up messy scans. Metadata routing sends files to the right parser.
03
Semantic search & chat
Users can just ask questions instead of scrolling through 100-page reports. A query like “Compare assets with liabilities this quarter” gets an instant, context-aware answer.
04
Role-based access
Access is locked down by role. Right now only managers and operators can use it. Rollout starts with internal ops teams before opening up further.
05
Data lifecycle manager
Handles history and cleanup. Chat sessions are stored, collections can be resynced, and old data is automatically removed on schedule.
06
Financial report support
Covers around 10 report types: balance sheets, income statements, liability. New data is added as business needs evolve.
From zero AI to Knowledge Engine in 90 days
The engine is in full rollout. The client is scaling the AI team, moving from RAG into agents and more advanced models
Before
Competitors demoing AI. Buyers asking for it
Up to 30 min to search through reports
Cloud AI ruled out by data residency
After
A working AI product on their own servers
Seconds. Same number. Same report
Self-hosted Llama 3.3. Zero data egress
Increase your team velocity
Book a free consultation to see how our certified engineers can be embedded into your product development process
If you have more engineers like this, I'll make space for them.
.avif)
I'm delighted to say that deciding to go with Radency has taken away a key area of stress in my life!
.avif){kind=icon}
They have exceeded expectations in terms of value for money, professionalism, and quality.
.avif)
They really took the time to understand our project and asked the right questions.
Tell us about
your ideas
Oops! Something went wrong while submitting the form.
.avif)
.svg){kind=icon}
.avif)