README / README.md
drmartin's picture
Update README.md
4f5a5d1 verified
---
title: README
emoji: 🏃
colorFrom: yellow
colorTo: gray
sdk: static
pinned: false
---
# Martin Technologies LTD — Sovereign Large Language Models
**Website:** [martintech.co.uk](https://martintech.co.uk)
**Regions:** UK & EU
**Focus:** Training, deploying, and operating **sovereign** Large Language Models (LLMs) with full data control, real-time performance, and cost efficiency.
---
## Mission
We build and operate **sovereign LLMs** for organisations that require **full ownership, auditability, and control** over their AI stack—without compromising on **state-of-the-art capability** or **real-time latency**. Our systems are optimised for **dedicated hardware** to reduce unit economics while delivering predictable performance and strict data boundaries.
---
## What “Sovereign” Means Here
- **You own the runtime:** Dedicated single-tenant deployments (cloud, edge, or on-prem) with **no shared inference plane**.
- **You govern the data:** Hard data boundaries, private networking, and explicit opt-in for any data retention. **No training on your prompts** by default.
- **You decide the geography:** Compute and storage pinned to the **UK or EU** with optional **air-gapped** configurations.
- **You can inspect & reproduce:** Open model families, transparent configuration, deterministic builds, and reproducible evaluation pipelines.
---
## Models & Training
We specialise in **state-of-the-art open-source** model families and customise them to your domain and latency/throughput constraints:
- **Base & Instruct Models:** General chat, RAG-optimised instruction models, coding, and tool-use variants.
- **Fine-Tuning & Adaptation:** Lightweight LoRA/QLoRA, adapters, and full-stack fine-tuning for domain language, terminology, and stylistic constraints.
- **Alignment & Safety:** Multi-objective RLHF/DPO where required; policy gradients for content filters; evaluation suites aligned with your risk profile.
- **Evaluation:** Task-specific evals (exact-match, BLEU/ROUGE, factuality, hallucination risk, tool-use success), latency SLOs, and cost/quality Pareto frontiers.
> We prioritise openly auditable model families to preserve portability and long-term independence.
---
## Real-Time Optimisation on Dedicated Hardware
Our inference stacks are engineered for **low-latency, cost-efficient** operation:
- **Kernel-level acceleration:** FlashAttention-class attention kernels, fused ops, paged KV cache, and continuous batching.
- **Quantisation:** INT8/INT4 & mixed-precision pipelines tuned per layer to balance perplexity vs. latency.
- **Parallelism strategies:** Tensor, pipeline, and context parallelism with NUMA-aware placement.
- **Speculative & constrained decoding:** Speculative decoding, prefix caches, grammar-constrained decoding for structured outputs (JSON/SQL).
- **Memory topology:** KV cache pinning, CPU-GPU offload, NVLink/PCIe bandwidth planning, and pinned host memory for surge loads.
**Outcome:** predictable p50/p95 latency under load, reduced cost per million tokens, and stable throughput on **dedicated single-tenant** hardware.
---
## Deployment Options
### 1) Managed Cloud (UK/EU)
- **Single-tenant** VPC deployments in the UK or EU, private subnets, customer-managed keys (CMK) optional.
- Hard residency guarantees and private endpoint exposure (PrivateLink/private service connect).
### 2) Physical Edge Compute
- Ruggedised nodes for **branch, factory, vessel, or field** environments.
- **Store-and-forward** telemetry, offline-first inference, and sync when connectivity returns.
### 3) On-Premises (Air-Gap Optional)
- Delivered as **appliance** or **reference build** (rack spec + BOM).
- Offline provisioning, **no outbound network** requirement, and fully local observability.
---
## Access Patterns
- **API Access:** OpenAI-compatible endpoints for chat/completions, embeddings, tool calls, and JSON-mode.
- **gRPC & SSE:** Streaming tokens for real-time UX; back-pressure aware.
- **RAG Tooling:** Connectors for document stores, vector DBs, and safety classifiers.
- **Multi-Tenant at Your Edge:** You define tenants; we enforce strict isolation per tenant within your sovereign boundary.
**cURL**
```bash
curl -X POST "$BASE_URL/v1/chat/completions" -H "Authorization: Bearer $MARTINTECH_API_KEY" -H "Content-Type: application/json" -d '{
"model": "martintech/sovereign-llm",
"messages": [{"role": "user", "content": "Summarise our latest policy in 5 bullets."}],
"temperature": 0.2,
"stream": true
}'
```
**Python**
```python
import os, requests, sseclient
BASE_URL = os.getenv("BASE_URL", "https://api.your_instance_url.co.uk")
API_KEY = os.getenv("MARTINTECH_API_KEY")
payload = {
"model": "martintech/sovereign-llm",
"messages": [{"role": "user", "content": "Draft a GDPR-compliant notice."}],
"temperature": 0.0,
"stream": True
}
headers = {"Authorization": f"Bearer {API_KEY}", "Content-Type": "application/json"}
with requests.post(f"{BASE_URL}/v1/chat/completions", json=payload, headers=headers, stream=True) as r:
client = sseclient.SSEClient(r)
for event in client.events():
print(event.data)
```
**JavaScript (Fetch)**
```js
const res = await fetch(`${BASE_URL}/v1/chat/completions`, {
method: "POST",
headers: {
"Authorization": `Bearer ${API_KEY}`,
"Content-Type": "application/json"
},
body: JSON.stringify({
model: "martintech/sovereign-llm",
messages: [{ role: "user", content: "Generate a JSON receipt." }],
response_format: { type: "json_object" }
})
});
const data = await res.json();
console.log(data.choices[0].message.content);
```
> The API is **OpenAI-compatible**, so most existing SDKs and clients work with only a **base URL and key** change.
---
## Security & Compliance
- **Data Handling:** No prompt or completion retention unless explicitly enabled. Configurable TTLs and redaction.
- **Encryption:** TLS in transit; at-rest encryption with customer-managed keys optional.
- **Network:** Private networking, IP allow-lists, and optional mTLS between services.
- **Isolation:** Per-tenant logical isolation; dedicated hardware optional for physical isolation.
- **Observability:** Privacy-preserving logs and metrics; structured audit events with redaction.
- **Governance:** DPA addendum, data residency controls (UK/EU), and support for customer risk assessments.
---
## Cost Optimisation
- Right-sized model families per use case (tiny → large) with **policy-based model routing**.
- Quantisation and continuous batching to reduce **cost per million tokens**.
- **Cache-aware RAG** to minimise context length and I/O.
- Performance budgets and autoscaling tied to your **SLOs** rather than best-effort throughput.
---
## Typical Use Cases
- **Private Assistants** for regulated teams (legal, finance, public sector).
- **RAG over Sensitive Corpora** with strict data residency.
- **Structured Generation** (JSON/SQL) into downstream systems.
- **Edge Autonomy** for low-connectivity scenarios (manufacturing, maritime, defence).
- **Developer Copilots** confined to internal codebases.
---
## Hugging Face Integration
- **Org Repos:** Model cards, adapters, and eval reports published under our Hugging Face organisation for **transparent provenance**.
- **Spaces & Demos:** Private Spaces for stakeholder testing; gated access with audit logs.
- **Artifacts:** Tokenisers, prompt templates, and guardrail grammars for reproducible pipelines.
> Ask us about publishing **redacted eval sets** and **prompt grammars** alongside each model variant.
---
## Getting Started
1. **Choose a deployment:** UK/EU managed cloud, edge appliance, or on-prem.
2. **Select a model class:** General chat, code, RAG-optimised, or constrained-output.
3. **Provide domain data (optional):** We prepare adapters or full fine-tunes with strict handling.
4. **Integrate the API:** Swap your base URL and key; keep your existing SDKs.
5. **Validate:** Review eval dashboards, latency/cost reports, and guardrail policies.
Contact: **martin@martintech.co.uk**
---
## Support & SLAs
- **Production SLAs:** Custom p95 latency, availability targets, and incident response windows.
- **Runbooks:** Operator playbooks for **air-gapped** and **edge** scenarios.
- **Training & Enablement:** Developer workshops, RAG patterns, and prompt-engineering for structured outputs.
---
## Why Martin Technologies LTD
- **Sovereignty by design:** Data, runtime, and geography under your control.
- **Open models, no lock-in:** Auditability and long-term portability.
- **Real-time, cost-efficient:** Systems engineering that meets product UX and budget constraints.
- **UK/EU Native:** Residency, procurement, and compliance aligned with your jurisdiction.
---
### Legal
© Martin Technologies LTD. All rights reserved.
Data residency options available in the **United Kingdom** and the **European Union**.
Model licences and third-party attributions are documented per-artifact in their respective repositories.