David Cohen
CTO, Cohen & Partners law firm
We moved all of our document work to AI through liracode.dev. The data physically stays with us — clients are at ease and the lawyers are happy.
Sovereign AI Infrastructure · Israel
Frontier AI,
on infrastructure
you own.
Every leading model — Claude, GPT, Gemini, and the rest — runs on hardware you own or control, in your jurisdiction. Your data never leaves your disks. A deed of ownership, not a subscription.
All models, one endpoint
Your data never leaves
Cloud, your servers, or both
Owned, not rented
Every leading model
Claude
GPT-4
Gemini
Llama
Mistral
Qwen
DeepSeek
Powered by
How it works
Three steps to
managed AI
01
Step 1 · Your foundation
Physical disks and databases in a certified Israeli data center. Full data ownership — no AI provider has access.
NVMe · 3.84TB · Encrypted
NVMe · 3.84TB · Encrypted
Backup · Cold
02
Step 2 · We manage
Setup, indexing, RAG, model routing, security, monitoring — we handle all infrastructure for you.
Your request
→
Router
→
Claude
GPT-4
Gemini
03
Step 3 · You get
All leading models through a single subscription. Open models hosted on Nebius GPU — or API access to closed ones. One price, no DevOps.
Claude 4 SonnetAPI · Subscription
Claude 4 SonnetAnthropic
Context200K
TypeClosed API
TasksCode, analysis, RAG
Accessliracode.dev subscription
Anthropic's flagship model. The best choice for legal analysis, code generation and complex reasoning chains.
GPT-4.1API · Subscription
GPT-4.1OpenAI
Context1M
TypeClosed API
TasksGeneral-purpose
Accessliracode.dev subscription
OpenAI's multimodal model. Excellent at long documents, summarization and multilingual tasks.
DeepSeek-R1-0528Nebius · $0.80/1M Input
DeepSeek-R1-0528DeepSeek · via Nebius
Parameters671B MoE
Context128K
Input$0.80/1M
Output$2.40/1M
A reasoning model with chain-of-thought. A rival to o1 on math and logic. Data never leaves your infrastructure.
DeepSeek-V3-0324Nebius · $0.50/1M Input
DeepSeek-V3-0324DeepSeek · via Nebius
Parameters685B MoE
Context128K
Input$0.50/1M
Output$1.50/1M
A fast MoE model for generation, chat and summarization. The optimal balance of price and quality.
Qwen3-32BNebius · $0.10/1M Input
Qwen3-32BAlibaba · via Nebius
Parameters32B
Context128K
Input$0.10/1M
Output$0.30/1M
Compact and fast. Ideal for streaming document processing and routine tasks at the lowest price.
Llama 3.3 · 70BNebius · $0.13/1M Input
Llama 3.3 · 70BMeta · via Nebius
Parameters70B
Context128K
Input$0.13/1M
Output$0.40/1M
Meta's open-source flagship. A great balance for code generation, chat and instructions. Fully on your infrastructure.
We purchase AI tokens in bulk at reduced prices and give you access to all models through a single subscription. You get Claude, GPT, Gemini and other models without needing separate contracts with each provider. Your data stays on your physical disks in an Israeli data center.
Architecture
Three layers
from model to execution
Model Layer
All AI Models
Claude, GPT, Gemini (API) · Qwen, DeepSeek (Nebius) · Your fine-tuned models
Control Layer
Routing & Optimization
routing · fallback · cost optimization · observability
Execution Layer
Compute Infrastructure
Nebius GPU · self-hosted servers · hybrid deployment
Physical ownership
Your data on your physical disks. No AI provider has access to your documents. Client confidentiality is preserved by default.
Enterprise RAG
Semantic search across cases, contracts and documents. AI legal assistant works with your knowledge base on your disks.
Cost transparency
See exactly what you pay for. Platform fee is separate from compute and token costs. No hidden margins.
No lock-in
Your infrastructure is portable by design. Open configs, exportable data, swappable providers. Leave anytime.
Deployment
Deploy your way —
cloud, self-hosted, or both
Cloud (Nebius)
GPU instances, managed scaling, token API. We handle infrastructure — you use models.
Self-hosted
Your servers, your network, full isolation. Complete control over every component.
Hybrid
Mix cloud and on-prem based on workload. Sensitive data stays local, scale bursts go to cloud.
Nebius models & GPU pricing
| Qwen2.5-32B | ~$0.06/1M input tokens |
| DeepSeek-R1 | ~$0.8/1M input tokens |
| DeepSeek-V3-0324 | $0.50/1M input |
| Llama / Hermes 405B | Market rate |
Token API
Pay per use
GPU Compute
Rent by hour
Self-host
Your hardware
No lock-in guarantee
Replaceable by design.
We earn your business every month.
Leave anytime
No contract lock, no exit fees. Month-to-month, cancel when you want.
Full data export
All data, embeddings, and configs exportable at any time. Your data is always yours.
Open configs
YAML/JSON configuration, infrastructure-as-code, fully dockerized. No proprietary formats.
Multi-provider
Not locked to any single cloud. Swap execution layers freely between Nebius, AWS, GCP, or self-hosted.
Portable vectors
Qdrant, Weaviate, Milvus — take your embeddings anywhere. Standard formats, no vendor trap.
Security
We protect
from perimeter to query
Physical data center security
Your disks in a certified Israeli data center. Access control, video surveillance, backup power — all included.
Network protection
WAF · Web Application Firewall
DDoS Protection · Always On
Traffic Filtering · L3–L7
Invisible Servers · Zero Exposure
Zero Trust
We set it up — you work. Every employee is verified by identity and device. No trust by default.
Encryption
Documents and embeddings encrypted at rest and in transit. Inference queries sent to LLM providers via encrypted channels. Encryption keys are yours only.
Data isolation
Configurable access levels for each employee. The vector database returns document chunks only after verifying company and case access rights.
Audit & compliance
GDPR, Israeli law — we maintain logs. Every document access is recorded. Full audit trail for regulators.
For maximum privacy, deploy self-hosted models via Nebius GPU — all inference stays on your infrastructure.
Query path
How a query passes
through the security system
Stage 0
User
>
DDoS ShieldSQL InjectionXSSBot FilterScript Block
Cloudflare · Partner
Result: cleaned, authorized query ready for Query Security Filter
JSON Output
Query Sec. Filter
Checking client role in the company. Reformulating or blocking suspicious queries.
JSON{
"clean-query": "...",
"user-id": "12345",
"tenant-id": "Comp-A",
"user-role": "standard",
"risk-score": 0.05
}
Checks
Access & Threat Detection
flowchart LR
A["Запрос"] --> B{"Роль?"}
B -->|Нет| X["Блок"]
C -->|Вер| X["Блок"]
B -->|OK| C{"Pattern Scan"}
C -->|OK| D["Risk Score"]
C -->|Inject| X
D --> E["Пропуск"]
Embedding Zone
Stage 3
Embedding Service
Capabilities
Preprocessing & Embedding
flowchart LR
A["Текст запроса"] --> B["Нормализация"]
B --> C["Stop-words"]
C --> D["Embedding Model"]
D --> E["Vector 768-dim"]
Local
Qdrant
Weaviate
Milvus
Chromapgvector
Cloud
Pinecone
ZillizWeaviate Cloud
Vector DB
Vector DB Search
flowchart LR
Q["Vector"] --> CS["cosine_sim"]
CS --> TH{"threshold?"}
TH -->|OK| TF{"tenant?"}
TH -->|No| X["Drop"]
TF -->|OK| PL{"perm?"}
TF -->|No| X
PL -->|OK| R["Top-K"]
Filtering
Access-Filtered Results
flowchart LR
DB[("Vector DB")] --> RLS["Row-Level Security"]
RLS --> T["Tenant Filter"]
RLS --> P["Permission Filter"]
T & P --> OUT["Filtered Chunks"]
Stage 5
Chunk-Level Sanitization
flowchart LR
A["Чанк с PII"] --> B["NER Scan"]
B --> C["Phone"]
B --> D["Email"]
B --> E["DocID"]
C & D & E --> F["Чистый чанк"]
050399029...→[Phone]
slepppi@gmail→[Email]
Token Factory
Local LLM Inference
Local LLM Inference
Audit
Query Budgeting & Audit
flowchart LR
A["Ответ"] --> B["Log"]
B --> C["Tokens"]
B --> D["Budget"]
C & D --> E["Audit Trail"]
E --> F["GDPR"]
Control
Rate Limiting + Audit Trail
▸ 2026-03-18T14:32:01Z query_id=a8f3c
▸ tokens_used: 1,847 / budget: 92%
▸ rate_limit: 14/50 req/min
▸ audit_hash: sha256:e4b2...9f1a ✓
Pricing
Simple subscription
no hidden costs
Choose how you pay
Token API
Pay per million tokens, all models
GPU Compute
Rent H100 by the hour for training/inference
Hybrid
Subscription + on-demand compute
Platform fee is visible. We don't hide margins in token markup.
Starter
~₪750/mo
support · for small teams
Setup: from ₪5,000 (one-time)
- All AI models (Claude, GPT, Gemini)
- Basic physical disks
- Up to 3 users
- Basic support
- Encryption and audit
Infrastructure details
Starter
$500 — $1,300/mo
Cloudflare CDNMiddleware ProxyApp Server
Supabase PostgreSQLQdrant Vector DBNebius GPU
Local NVMe 2TBLLM Inference
Database server (Supabase)
$80 — $150
Vector DB (Qdrant)
$80 — $150
GPU inference (Nebius)
$300 — $900
Physical disks (2TB)
$15 — $50
Network (Cloudflare)
$20 — $60
Infrastructure:$495 — $1,310
You pay only ~₪750/mo — we cover everything else
MedOneNebiusCloudflare
API models via token subscription. Open-source models available as add-on.
Recommended
Professional
~₪2,500/mo
support · for law firms
Setup: from ₪15,000 (one-time)
- All AI models + priority access
- RAG and vector search for cases
- Up to 15 users
- Extended support
- AI legal assistant
Infrastructure details
Professional
$1,500 — $5,500/mo
Cloudflare CDN + LBMiddleware ProxyK8s Backend Cluster
PostgreSQL HAQdrant ClusterNebius GPU x2-4RAG Pipeline
Local NVMe 8TB RAIDMulti-Model Routing
Backend cluster
$200 — $400
Vector DB cluster (Qdrant)
$200 — $500
GPU inference ×2-4 (Nebius)
$1,000 — $4,000
Physical disks (8TB RAID)
$60 — $220
Network + LB (Cloudflare)
$100 — $300
Infrastructure:$1,560 — $5,420
You pay only ~₪2,500/mo — we cover everything else
AnanNebiusCloudflareVercel
API + open-source models included. Nebius models available as add-on.
Organization
from ₪7,500/mo
support · for large firms
Setup: from ₪35,000 (one-time)
- Dedicated infrastructure
- Unlimited users
- SLA with guarantees
- Personal manager
- Custom integration
Infrastructure details
Organization
$7,000 — $25,000/mo
Cloudflare EnterpriseLoad BalancerK8s Cluster x4-6
PostgreSQL HA + RedisQdrant HA ClusterNebius H100 x5-20
Enterprise RAGFull Model FleetAudit + Compliance
NVMe 24TB RAID + DRBackup + DR
Backend nodes ×4-6
$500 — $1,500
Vector DB cluster (Qdrant)
$500 — $2,000
GPU ×5-20 (Nebius H100)
$4,000 — $18,000
Distributed cache (Redis)
$200 — $500
Physical disks (24TB RAID + backup)
$250 — $900
Full colocation rack
$500 — $2,000
Infrastructure:$5,950 — $24,900
You pay only from ₪7,500/mo — we cover everything else
AnanBezeqNebiusCloudflareVercel
Full model fleet — API, open-source, and Nebius models. Custom fine-tuned models supported.
One-time setup + ongoing support · No DevOps on your side · all AI models · 24/7 monitoring
Works with
Add-ons
When you are
ready
Model freedom
Run open-source models (Llama, Mistral) or Nebius-hosted models (Qwen, DeepSeek) on your data. Data never leaves your disks. Full control over the model.
GPU time
Rent Nebius H100 compute for fine-tuning and inference. $3–5/hour. Scales to your task. Pay only for actual usage.
Smart routing
Automatic query evaluation and optimal model selection. Frequent query caching. Cost reduction without quality loss.
Private sandbox
Isolated environment for testing models on confidential data. Zero outbound requests. Complete network isolation.
Training available for open-source models (Llama, Mistral) and Nebius models (Qwen, DeepSeek) — data stays on your disks. Claude, GPT, Gemini — access through token API or within your subscription.
Why us
Five reasons
to choose us
@dmitrii.mukomelAvailable for hire
Work TogetherFounder
Built by one engineer
Full-stack engineer who works problem-first: identifies real operational friction, designs an architecture against it, and ships solo to production.
Meet the founder →Testimonials
What our
clients say
Michael Levi
Senior Developer, fintech startup
I used to call the OpenAI API directly. After a data leak hit our competitors, I switched to liracode.dev. Every model, zero trust, our own drives.
Noa Ben-Ari
Head of Operations, logistics
We automated order processing and routing with AI. We save 40 hours a week, and the data never leaves us.
Ron Mizrahi
DevOps Lead, cybersecurity
I reviewed the architecture — Cloudflare + Zero Trust + physical isolation. Finally a managed AI service you can actually trust.
Yael Shapira
CEO, accounting firm
We connected Claude and GPT to analyze financial statements. Client data sits on our own drives in Israel — that is the deciding factor.
Amir Hassan
Full-stack Developer, freelance
One subscription instead of five API keys. Transparent pricing, and no need to think about security — it is all set up already.
David Cohen
CTO, Cohen & Partners law firm
We moved all of our document work to AI through liracode.dev. The data physically stays with us — clients are at ease and the lawyers are happy.
Michael Levi
Senior Developer, fintech startup
I used to call the OpenAI API directly. After a data leak hit our competitors, I switched to liracode.dev. Every model, zero trust, our own drives.
Noa Ben-Ari
Head of Operations, logistics
We automated order processing and routing with AI. We save 40 hours a week, and the data never leaves us.
Ron Mizrahi
DevOps Lead, cybersecurity
I reviewed the architecture — Cloudflare + Zero Trust + physical isolation. Finally a managed AI service you can actually trust.
Yael Shapira
CEO, accounting firm
We connected Claude and GPT to analyze financial statements. Client data sits on our own drives in Israel — that is the deciding factor.
Amir Hassan
Full-stack Developer, freelance
One subscription instead of five API keys. Transparent pricing, and no need to think about security — it is all set up already.
David Cohen
CTO, Cohen & Partners law firm
We moved all of our document work to AI through liracode.dev. The data physically stays with us — clients are at ease and the lawyers are happy.
Michael Levi
Senior Developer, fintech startup
I used to call the OpenAI API directly. After a data leak hit our competitors, I switched to liracode.dev. Every model, zero trust, our own drives.
Noa Ben-Ari
Head of Operations, logistics
We automated order processing and routing with AI. We save 40 hours a week, and the data never leaves us.
Ron Mizrahi
DevOps Lead, cybersecurity
I reviewed the architecture — Cloudflare + Zero Trust + physical isolation. Finally a managed AI service you can actually trust.
Yael Shapira
CEO, accounting firm
We connected Claude and GPT to analyze financial statements. Client data sits on our own drives in Israel — that is the deciding factor.
Amir Hassan
Full-stack Developer, freelance
One subscription instead of five API keys. Transparent pricing, and no need to think about security — it is all set up already.
[ COMMON QUESTIONS ]
Questions teams ask before they switch
What if local models aren't good enough?
liracode uses hybrid routing: sensitive data stays local, non-sensitive goes to cloud. You get the best of both worlds.
What's the migration risk?
We provide managed migration with zero downtime. Hybrid mode lets you transition gradually over weeks, not months.
Do we physically own the hardware?
Yes. liracode procures and deploys hardware in your name. It's your asset on your balance sheet. We manage it.
Can we use our own cloud API keys?
Yes. BYOK (Bring Your Own Keys) lets you use Claude, GPT-4, or Gemini. liracode sanitizes prompts before they reach any provider.
What about scaling?
Add GPU nodes on demand. We handle capacity planning, procurement, and deployment. No cloud lock-in.