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.
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.
- Data stays inside your perimeter
- Open weights, no vendor lock-in
- One measurable pilot before scale
- Humans approve the learning loop
- Every generation is auditable
- Rollback is always available
Stack
Boring where it should be boring; modern where it matters.
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.