Hire AI and LLM engineers
RAG, agents, evals, Vercel AI SDK
Engineers who have taken language-model features past the demo: retrieval that returns the right chunk, evaluations that catch a regression, and a cost line that does not startle your finance team in month three.
Getting a language model to do something impressive once is a weekend's work. Getting it to do the same thing reliably for ten thousand users, at a cost you can forecast, with a way to tell whether last week's prompt change made things worse, is the actual job.
Our AI engineers work on that second part. Retrieval comes first: chunking that respects document structure, embeddings chosen by measurement rather than by blog post, hybrid search with a reranker, and retrieval quality evaluated separately from generation quality. That separation matters, because when a RAG system gives a wrong answer it is almost always retrieval that failed, not the model. Then evaluation: a golden set of questions, automated scoring, regression runs in CI, so a prompt or model change becomes a measured decision instead of a hunch somebody defends in a meeting.
They are pragmatic about agents, sometimes to a client's disappointment. Most problems described as needing an agent are better served by a scoped tool call or a deterministic workflow with one model step inside it. Where an agent genuinely fits, they build it with explicit tools, guardrails, budget caps and traces you can replay after a bad run.
Cost and latency are treated as requirements rather than as things to discover later: model routing by task, caching, streaming, batching, and a dashboard showing spend per feature.
When a dedicated AI engineer is the right call. You have a corpus, documentation, tickets, contracts, catalogue, that people search badly today. You shipped a prototype and cannot make it dependable. You need somebody to judge whether an AI feature is worth building at all. Or the feature works and nobody knows what it costs per user.
When it is not. If the work is really connecting SaaS tools to each other, an automation engineer costs less and ships sooner. If you need model training rather than application engineering, that is a research hire and not what this bench does.
The rate reflects a genuinely scarce skill set. We would rather quote it honestly than fill the seat with a Python developer who has read the documentation twice.
What they do
Retrieval-augmented generation
Chunking, embeddings, hybrid search and reranking over your own documents, with retrieval scored separately from the answers.
Evaluation harnesses
Golden question sets, automated scoring and regression runs in CI, so a prompt or model change is a decision rather than a gamble.
Agents and tool use
Scoped tools with typed schemas, guardrails, budget caps, and traces you can replay when a run goes somewhere unexpected.
Chat and assistant interfaces
Streaming interfaces on the Vercel AI SDK, citations shown properly, feedback captured, and a sane fallback when the model is down.
Cost and latency engineering
Model routing by task, prompt caching, batching, token budgets, and a spend dashboard broken down by feature.
Deployment and safety
Rate limits, PII handling and redaction, prompt-injection defences, and logging that satisfies your own data policy.
Skills
Seniority & rates
Sample profiles
Engagement models
Hourly
For bursts of work and part-time needs. Minimum 40 hours. Time tracked in Hubstaff, reported weekly.
- Pay for logged hours only
- Scale up or down weekly
- Same senior engineer throughout
Dedicated monthly
One engineer, full-time, inside your team and your tools. Eight hours a day, five days, in your timezone window.
- 15-day risk-free trial
- Free replacement, no argument
- Daily stand-up, sprint reporting
- NDA + 100% IP assignment
Dedicated pod
Tech lead, three engineers and QA, run as a Scrum team with a delivery manager included in the price.
- Fixed sprint cadence
- Delivery manager at no cost
- Architecture and code review built in
How to hire
Describe the corpus and the questionDay 0
What documents exist, who asks what of them today, and how you would know the answer was good. That last one is the hard part.
Profiles, and our verdict on the idea48 h
Two CVs, plus a written read on whether we think the use case will work. Roughly a third of the AI briefs we get come back with a recommendation not to build it as described.
Interview, and ask about a failure2–4 days
Ask what their last system got wrong and how they found out. Anyone without a story there has not run one in production. No charge at this stage.
A baseline you can trust or rejectDays 1–15
A thin end-to-end slice in the fortnight: real documents, real retrieval, an evaluation set and a measured baseline. You judge numbers rather than a demo.
Watch quality and spend togetherOngoing
NDA and IP before the first commit, evaluation runs in CI, and a cost dashboard per feature from the first month. Pause on 30 days' notice.
FAQ
How do we know the AI feature will actually work before spending months on it?
You build the evaluation set first. Twenty to fifty real questions with known good answers, scored automatically, gives you a baseline inside a fortnight. If that baseline is poor and does not move with retrieval work, you have a cheap and honest no rather than an expensive slow one.
Which model should we use?
It depends on the task, and usually it is more than one: a strong model for reasoning, a small fast one for classification and routing. Our engineers benchmark on your evaluation set rather than repeating vendor claims, and build so you can switch when the market moves again.
Can our data stay private?
Yes. Enterprise API tiers from the major providers do not train on your data, and open-weight models can run in your own cloud where policy demands it. Data handling, retention and redaction go in writing before a single document is indexed.
What will it cost to run each month?
That is part of the design rather than a discovery in month four. Expect a per-request cost model within the first weeks, routing and caching to bring it down, and a dashboard showing spend by feature so you watch it move rather than reading it on an invoice.
Do we need an agent?
Probably not, and we will say so before you have bought anything. Most briefs work better as a deterministic workflow with one or two model steps: cheaper, faster, far easier to debug at 4 p.m. on a Friday. Agents earn their keep when the sequence genuinely cannot be known in advance.
Why is this rate higher than your other backend roles?
Because the supply of engineers who have taken language-model features to production rather than to a prototype is genuinely small, and their salaries reflect that. The rate is what they cost us plus our normal margin, not an AI premium bolted on because the category is fashionable.
Tell us the role. Profiles in 48 hours.
No retainer to see CVs, and no invoice until the trial ends and you say yes.