Client work, run by the people who build Avatar.
Four practices, priced by phase. Most projects touch two of them. Open a heading for scope, deliverables and how long it usually takes.
A language model is easy to demo and hard to trust, so most of the work sits around it: retrieval that finds the right paragraph, prompts held in version control, evaluations that fail loudly when quality slips.
Recent shape of the work — a support assistant grounded in five years of tickets, a pipeline that reads scanned purchase orders into an ERP, a forecast that names the twelve products to reorder on Monday. Where a query and a report would do the job, we will say so on the first call.
- Working system, in your infrastructure
- Evaluation suite and measured baseline
- Cost model per thousand requests
- Retraining runbook
Avatar, or whichever model measures best on your data.
Whole products, from the schema to the empty states. We keep to a small set of well-understood tools so the code stays legible to the team that inherits it.
Interfaces are designed for the person using them forty times a day: keyboard paths, states that explain themselves, and a first paint that arrives before the coffee lands.
- Product, repository and CI
- Design source and component set
- Written architecture notes
- Recorded handover walkthrough
TypeScript end to end, Postgres, one queue, one cache.
We map the workflow as it is performed, not as the manual describes it, then automate the parts that never needed judgement. Everything an agent does is written to a log a person can read.
Where a mistake would be expensive, the agent stops and asks. Approval thresholds are yours to set and change without a release.
- Automated workflow with audit trail
- Exception queue for humans
- Before-and-after time measurement
- Failure alerts routed to a person
Anything done weekly by three people from the same source data.
Deployments you can repeat from a clean machine, secrets held in a vault rather than a chat thread, and a restore somebody has actually performed while being timed.
Sizing follows the traffic you have today, with the next two steps written down so growth is a decision rather than an emergency.
- Infrastructure as code, reviewed
- Release and rollback process
- Access model and audit log
- Restore drill, timed and documented
Monthly, with a named engineer and a response window.
How a project runs
Four phases, each ending in something you can open in a browser and argue with.
We sit with the people doing the work and write down the decisions your business makes every day.
One thin slice on real data. Projects that should stop, stop here, with the research yours to keep.
Weekly releases to a staging environment you use, in a shared channel with the engineers writing it.
Runbooks, a restore drill, a recorded walkthrough, and three months of support through first month-end.
Not sure which of the four you need?
Describe the problem in plain language. We will tell you which practice it belongs to, and whether it needs software at all.