012 wks
Scope
- Ranked use-case list with value and risk
- Data inventory and access plan
- Architecture and hardware options
- Cost model with break-even month
LLM.co · Private AI & Custom AI DevelopmentHow we work
Five phases, each with fixed scope, named deliverables and a go / no-go. You can stop after any phase and keep everything built so far.
FIG. 06 — IBM 7090 room, NASA AmesLLM.co
Process
From scoping call to running in your rack.
Fixed-scope phases with a go / no-go at the end of each one. You can stop after any phase and keep everything built so far.
012 wks
Scope
Pick the workflow worth automating and prove the numbers.
- Use-case ranking
- Data audit
- Cost model
023–4 wks
Prototype
A working build on your real data, scored against an eval set.
- Eval set v1
- Model bake-off
- Pilot UI
034–6 wks
Harden
Make it safe to give to hundreds of people and an auditor.
- Red-team
- SSO + RBAC
- Audit logging
041–2 wks
Deploy
Installed on your hardware or in your cloud account.
- GPU sizing
- Runbooks
- Handover
05Ongoing
Operate
Monitoring, model upgrades and retraining as the work changes.
- Drift alerts
- Model swaps
- Quarterly evals
How our AI development process runs
Every custom AI development engagement at LLM.co runs in five phases: Scope, Prototype, Harden, Deploy and Operate. Scope takes two weeks. Prototype takes three to four weeks, Harden four to six and Deploy one to two. A focused first system usually reaches production in eight to twelve weeks.
Each phase has a fixed scope, named deliverables and a go / no-go meeting at the end. We quote one phase at a time. If you stop, you keep everything built so far, including code, documents, evaluation sets and any trained models. This AI development process is built so that you never commit to more than the next step.
Who takes part from your side
Three roles matter most, plus a pilot group during Prototype. Their time is heaviest in Scope and Harden and lighter in between.
- A business owner for the workflow, who decides what good output looks like and makes each go / no-go call.
- A subject-matter expert who knows the data and can judge answers, such as a senior adjuster, technician, paralegal or analyst.
- A security or IT contact, who approves data access, identity integration and where the private AI system will run.
- During the Prototype phase, a small group of pilot users who try the system on real work and report what breaks.
How quality is measured
Quality is measured with an evaluation set. During Prototype we work with your subject-matter expert to collect real past cases with known correct answers. These include easy cases, hard cases and cases the system should refuse or send to a person. Each case gets a scoring rule, such as exact field match, correct citation or a rubric for drafted text.
The system is scored against that set before the pilot and again before launch. Candidate models are compared on the same set in a bake-off, so hardware choices follow from measured results. After launch the set is re-run on every model, prompt or data change, and you receive a quarterly evaluation report during Operate.
How scope changes are handled
Scope changes are a normal part of any AI development process once people use a working system. Small changes that fit the current phase are absorbed and noted. Larger changes, such as a new data source, a new integration or a second workflow, are written up with their effect on the evaluation set, schedule and cost. You decide whether to add them now, defer them to a later phase or drop them. Nothing is added to a quote without your approval.
What handover includes
The AI development process ends with handover. At the end of Deploy the system runs on your hardware or in your own cloud account, and everything needed to run it lives in your repositories.
- Source code, prompts and configuration.
- Evaluation sets, scoring scripts and the latest reports.
- Infrastructure as code for servers, GPUs, networking and monitoring.
- Runbooks, alerts and training for your operators.
- Security and data-flow documentation for your reviewers.
- Any fine-tuned model weights.
Deliverables
What you receive at each phase.
023–4 wks
Prototype
- Working system on your real data
- Evaluation set and baseline scores
- Model bake-off results
- Pilot with a small group of users
034–6 wks
Harden
- Red-team findings and fixes
- SSO, roles and document permissions
- Audit logging to your SIEM
- Security and data-flow documentation
041–2 wks
Deploy
- Hardware or cloud sizing
- Infrastructure as code
- Runbooks and alerts
- Training for your operators
05Ongoing
Operate
- Monitoring and drift alerts
- Model upgrades on your schedule
- Quarterly evaluation reports
- Optional support agreement
Questions
Asked on every scoping call.
Can we stop after a phase?
Yes. Every phase ends with a go / no-go decision. If you stop, you keep everything built so far, including code, documents, evaluation sets and models. Builds are quoted one phase at a time, so you are never committed beyond the current phase.
Who needs to be involved from our side?
A business owner for the workflow, a subject-matter expert who knows the data, and a security or IT contact. Their time is heaviest in Scope and Harden. During Prototype, a small group of pilot users tries the system on real work and reports problems.
How do you measure quality?
With an evaluation set built from your real past cases, each with a known correct answer and a scoring rule. The system is scored before the pilot and before launch, and the set is re-run on every model, prompt or data change. You see the scores before each go / no-go decision.
What happens in the two-week discovery sprint?
We rank candidate use cases by value and risk, audit the data each one needs, outline the architecture and hardware options, and build a cost model with a break-even month. You finish with a recommendation for the first build and a quote for the Prototype phase.
What if requirements change mid-project?
Small changes that fit the current phase are absorbed. Larger ones, such as a new integration or data source, are written up with their effect on the evaluation set, schedule and cost. You decide whether to add, defer or drop each one before any quote changes.
Does the process change for private AI on air-gapped networks?
The phases stay the same. Deployment and updates change. Models and software arrive as signed bundles that pass through your change process, and our engineers work from synthetic or approved sample data so they do not need access to controlled systems.
What support is available after handover?
Your team can run the system with the runbooks and training we deliver. If you prefer, LLM.co can handle monitoring, drift alerts, model upgrades and quarterly evaluations under a support agreement. The code, models and data remain yours either way.
Start here
Submit a job card.
Tell us the workflow and where the data lives. An engineer, not a salesperson, replies within one business day with a first take on architecture and cost.