AI developer · private deployments

George Shkliar

I build AI systems that run inside a company’s own infrastructure: local models, retrieval over internal knowledge, fine-tuning for brand voice, and agents that do real work — with humans in control.

GS
FocusPrivate LLM · RAG · fine-tuning
DeploymentOn-prem · air-gapped
Working stylePilot first, then scale

What I do

A narrow set of things, done properly: private inference, retrieval, adaptation, and the operational layer around them.

Local LLM deployment

Open-weight models on your GPUs with vLLM or Ollama: sizing, quantization, routing, and capacity planning.

Retrieval over your knowledge

Documents, CRM, tickets, and wiki become a permission-aware knowledge base with citations, not a hallucination machine.

Fine-tuning & brand voice

LoRA/QLoRA adapters and style profiles trained on approved examples: terminology, tone, templates, and domain logic.

Evaluation & operations

Golden sets, regression checks, guardrails, audit logs, and versioned rollbacks — so the system stays trustworthy after launch.

How I work

The constraints are part of the design, not an afterthought.

Stack

Boring where it should be boring; modern where it matters.

python · vLLM · ollama · llama / qwen / mistral
embeddings · vector search · rag pipelines
lora / qlora · evaluation harnesses · guardrails
docker · kubernetes · linux · observability

Engagements

Short audit, focused pilot, production handover, ongoing support. No long discovery theater.

Audit

Workflows, data readiness, hardware, constraints. A prioritized list of use cases with expected effect.

Pilot

One process, real data, one metric. A working system and an honest read on whether to scale.

Production

Integration, permissions, monitoring, and training for the team that will own it.

Support

Model updates, quality reviews, and fixes as your data and processes change.

Tell me the process that costs you the most time.

I will reply with a concrete pilot plan: data needed, hardware, timeline, and how success will be measured.

Telegram @george059