About Humai
Humai builds AI you can vouch for. Founded in Dubai in 2025, we exist for one problem: companies have adopted AI but will not rely on it for work with consequences. Our answer is one accountability model across two products. Huscribe for what AI does: it checks agentic workflows against what their owner approved. Mairit for what AI produces: it puts a qualified, accountable name on the work.
Audited to SOC 2 Type II and certified to ISO/IEC 27001:2022, both products in scope. Salesforce and Microsoft partner. Small team, founder-led, no layers.
The role
You build the AI systems behind Mairit, starting with the engine behind our certified legal translation, and you bring the same discipline to Huscribe.
This is applied work. Multi-step pipelines with checks and human gates, evaluation that decides what ships, model choices made on evidence, and an MCP surface that runs inside the agents our customers already use. You work alongside a senior backend engineer and mentor a junior AI engineer.
On-site in Dubai.
Why takes this seat
- Scope. You work across the AI of both products: models, evaluation, pipelines, and the trade-offs between them.
- The problem. Every output ends with a person who puts their name on it. The quality bar is real and your work is measured against it.
- Models and tools. Frontier models across providers, open models you fine-tune and serve, and MCP. We build with AI every day and use agents we built ourselves inside the company.
- Trajectory. This seat grows into the AI lead role as the team grows. You start by mentoring the junior engineer and get there by earning it.
- Pay. A salary set for a senior engineer, a UAE employment visa, and medical insurance.
- Access. A short line to the founder and to the reviewers who use what you build. Decisions take a conversation, not a quarter.
What you'll do
- Design and ship LLM pipelines with checks, retries and human gates, using LangGraph or similar
- Build the evaluation that gates every change: test sets, graders, regression suites, blind comparisons. Measure critical errors and reviewer time, not impressions
- Choose models per step across providers on quality and cost, with versions pinned and the trade-offs written down
- Build retrieval for terminology, exemplars and policies (PostgreSQL with pgvector)
- Fine-tune and serve open models where they beat hosted ones on cost or control
- Handle documents properly: OCR, vision cross-checks, layout and rendering
- Ship and maintain Mairit as an MCP tool for the agent clients our customers use
- Contribute evaluation and scoring to Huscribe
- Respect data controls end to end: provider terms, no-training, region, and our ISO 27001 and SOC 2 obligations
What good looks like after a year
- Mairit's translation engine ships certified work with fewer reviewer corrections, and you can show the numbers
- No prompt, model or pipeline change lands without passing the evaluation you built
- Mairit runs as an MCP tool inside the agents our customers use every day
- Model spend is a deliberate choice, per step, with the reasoning on record
- The junior AI engineer you mentored owns pipelines of their own
What you bring
- 4+ years of software engineering, including 2+ years shipping LLM systems that real users depended on
- Production agents: you have built and run multi-step, tool-using systems, not just prompts
- MCP: you have built or integrated servers or clients
- Evaluation: you have built harnesses that caught regressions before customers did
- Strong Python and production habits: tests, tracing, cost and latency budgets
- Retrieval and vector search in production
- You can explain a model choice to a founder and defend it to an engineer
Nice to have
- Fine-tuning and serving open models (LoRA, vLLM)
- Prompt optimisation frameworks (DSPy)
- OCR and document AI
- Arabic reading ability, for our legal translation work
- Salesforce or Agentforce experience
Not for you if
- You want to research models rather than ship them
- You judge a pipeline by its demo rather than by its evaluation
- You have only used LLMs through a chat window
- You are looking for a 9 to 5 job
- You are looking for a remote job
- You don't keep up-to-date with the latest AI developments
How we hire
- Three conversations: a short intro call, a working session on a real evaluation or pipeline problem, and a final conversation with the founder. Apply on LinkedIn with a link to something you built.
