AI systems, engineered where your data lives.
Locus designs agentic workflows, adapts open-weight models to your domain, and operates the dedicated hardware they run on — all inside environments you control.
- DUAL RTX PRO 6000 · BLACKWELL
- UBUNTU 24.04 LTS
- OPEN-WEIGHT MODELS
- ON-PREM · VPC · HYBRID
- HUMAN-IN-THE-LOOP
AGT · workflow orchestration
What we build
Four capabilities. One controlled stack.
Agents, adapted models, dedicated hardware, and the research behind them — operated as one system inside your boundary.
Workflows where AI does the work, not just the talking — research, document processing, and internal operations handled by agents that call your tools and answer to your rules.
Technical layer
- Multi-agent orchestration
- Retrieval-augmented generation
- Private tool integration
- Human-in-the-loop approval
Representative outputs
- Internal knowledge assistants
- Research and analysis agents
- Document-processing pipelines
- Domain-specific copilots
Deployment context
Runs inside your network boundary. Agent actions are logged and traceable by design; approval gates sit wherever your governance requires them.
Workflows where AI does the work, not just the talking — research, document processing, and internal operations handled by agents that call your tools and answer to your rules.
Technical layer
- Multi-agent orchestration
- Retrieval-augmented generation
- Private tool integration
- Human-in-the-loop approval
Representative outputs
- Internal knowledge assistants
- Research and analysis agents
- Document-processing pipelines
- Domain-specific copilots
Deployment context
Runs inside your network boundary. Agent actions are logged and traceable by design; approval gates sit wherever your governance requires them.
How an agentic workflow runs
From request to result — every step inside your boundary.
Illustrative architecture. Actual workflows are designed per engagement.
Custom model adaptation
From a general model to your model.
Open-weight models are capable generalists. Adaptation makes one of them a specialist in your domain — running on your hardware, at a size your budget likes.
Base model selection
Candidate open-weight models are screened against your task, latency, memory, and licensing constraints.
Proprietary knowledge
Your documents, records, and domain expertise — the material no public model was trained on.
Curated training set
Raw knowledge becomes training pairs: cleaned, deduplicated, formatted, and reviewed for quality and leakage.
Adaptation
Supervised fine-tuning, adapters, or preference optimization — the method follows the task, not the fashion.
Evaluation
A task-specific evaluation set measures whether the adapted model beats the base model on your work — before it ships.
Quantization
Weights are compressed to fit your memory and latency envelope, with quality re-measured after every compression step.
Deployment
The model goes live on your infrastructure behind the same serving, monitoring, and rollback discipline as any production system.
Managed AI infrastructure
Serious hardware, without the operations burden.
Locus leases dual NVIDIA RTX PRO 6000 Blackwell server configurations — installed, tuned, monitored, and maintained as a managed service inside your environment.
Tap a layer to inspect it
GPU
Dual RTX PRO 6000 Blackwell
Two workstation-class Blackwell GPUs per node, provisioned for local inference and fine-tuning workloads.
- Local inference
- Model serving
- Secure network integration
- Observability
- Optimization
- Maintenance
- Deployment assistance
- Technical support
Final configurations are determined after a workload assessment. We size hardware to measured requirements, not the other way around.
Choosing the right architecture
Local is a decision, not a religion.
Local infrastructure wins on some dimensions and loses on others. This is the framework we actually use — the honest version.
| Dimension | Local / managed by us | API / public cloud |
|---|---|---|
| Data location | Sensitive workflows stay inside client-managed environments. | Data transits and is processed on provider infrastructure. |
| Workflow control | Model versions, prompts, and behavior change only when you decide. | Provider updates can change behavior without notice. |
| Customization | Full access to weights enables fine-tuning and deep adaptation. | Limited to prompting and provider-supported tuning options. |
| Recurring cost | Fixed lease and operations cost, independent of token volume. | Per-token pricing scales directly with usage. |
| Latency | On-network round trips; no public-internet dependency. | Dependent on provider load, region, and network path. |
| Frontier capability | Bounded by current open-weight models — strong, but not always the ceiling. | Access to the largest proprietary models on the market. |
| Operational ownership | Hardware and serving need operating — that is what the managed service is for. | Zero infrastructure to run; the provider carries operations. |
| Elastic scale | Capacity is fixed until hardware is added. | Bursts absorb instantly, at a price. |
Local
Sensitive workflows stay inside client-managed environments.
API
Data transits and is processed on provider infrastructure.
Local
Model versions, prompts, and behavior change only when you decide.
API
Provider updates can change behavior without notice.
Local
Full access to weights enables fine-tuning and deep adaptation.
API
Limited to prompting and provider-supported tuning options.
Local
Fixed lease and operations cost, independent of token volume.
API
Per-token pricing scales directly with usage.
Local
On-network round trips; no public-internet dependency.
API
Dependent on provider load, region, and network path.
Local
Bounded by current open-weight models — strong, but not always the ceiling.
API
Access to the largest proprietary models on the market.
Local
Hardware and serving need operating — that is what the managed service is for.
API
Zero infrastructure to run; the provider carries operations.
Local
Capacity is fixed until hardware is added.
API
Bursts absorb instantly, at a price.
Our position
The right architecture may be local, private cloud, hybrid, or selectively API-assisted. We recommend what the workload justifies — including, sometimes, not us.
The R&D Lab
We test before we recommend.
The Lab is where open-weight releases meet our hardware. Its findings decide what we deploy for clients — and what we quietly decline to.
Agent reliability
How do multi-step agent workflows fail, and which orchestration patterns keep failure contained?
Local model evaluation
Which open-weight models hold up on real enterprise tasks — not leaderboard prompts?
Inference optimization
How much throughput can a dual-GPU node deliver per workload class before quality degrades?
Model memory
What memory architectures let an assistant stay useful across weeks of context without leaking scope?
RAG quality
Which retrieval and chunking strategies actually improve answer accuracy on messy internal corpora?
Fine-tuning efficiency
How small can an adapted model get before it stops beating the general-purpose baseline?
Open-weight integration
Which emerging open-weight releases are production-credible, and what does it take to serve them well?
Research tracks describe active internal investigation areas — not published papers or completed studies.
FN-01 · DEPLOYMENT POST-MORTEM · July 2026
156 GB of DeepSeek on two GPUs.
A 156 GB MoE model built for NVLink clusters, served from a desk-side workstation pair over plain PCIe — full 128K contexts behind an authenticated tunnel. What broke, what fixed it, and the exact numbers.
72.1 tok/s
single stream at 128K context
108.8 tok/s
across four concurrent streams
~1.2 s
to first token, warm CUDA graphs
11+ h
of serving with flat memory, zero OOM
single stream at 128K context
across four concurrent streams
to first token, warm CUDA graphs
of serving with flat memory, zero OOM
Measured in production — July 2026.
How an engagement runs
Six stages. A deliverable at every one.
Architecture discovery
A working session on your use case, data boundaries, constraints, and what success has to look like.
You receive
Written architecture brief with a candid feasibility read.
Data & workflow assessment
We map the data sources, access rules, and process steps the system will actually touch.
You receive
Assessment report: data map, gaps, and sizing inputs.
Prototype & evaluation
Go / no-go gateA working prototype against your real tasks, measured with an evaluation set built for your domain.
You receive
Prototype, evaluation results, and a go/no-go recommendation.
Local deployment
Hardware installed and configured, models served, workflows wired into your environment.
You receive
Production system running inside your boundary.
Integration & validation
The system meets your real users, tools, and governance. We validate against the acceptance criteria set in stage one.
You receive
Signed-off validation report and runbooks.
Managed operation
Monitoring, maintenance, model updates, and improvement cycles — the system keeps earning its keep.
You receive
Ongoing operations reporting and improvement roadmap.
Your AI stack should fit your operation.
An architecture session covers your use case, data boundaries, model requirements, infrastructure options, and a candid read on whether local deployment is the right call for your workload.
45–60 minutes. Technical staff welcome. No slideware.