Hire AI developers, from $1,800 a month

LLM features inside a product you already run

Summaries, drafting, classification, pulling fields out of free text. The model call is the small part. Most of the work is the software around it: validating input, checking that structured output parses, timeouts and retries, caching, logging each prompt with its response, and a plain fallback for when the model fails or is slow.

Retrieval over your own data (RAG)

The application finds the relevant passages in your documents first, then asks the model to answer from them and cite them. The engineering sits in the first half: how documents are split, indexed and ranked, and how access permissions carry through so nobody sees a passage they could not open directly. We have no published retrieval project, so test this in the interview: questions 2 and 3 below are written for it.

Agents that take actions, with guardrails

For Eatverse (Vision Foods), Qilin Lab engineers built custom AI agents for catalog validation, price integrity and supply chain anomaly detection. They run across 12 brands and thousands of SKUs, on Next.js, Python and LLM agents. The agents check data at ingestion and surface only the issues that need human attention. Ask for that pattern in any agent: narrow permissions, an action log, and a person who approves anything that cannot be undone. OWASP lists "excessive agency" as risk LLM06 in its 2025 Top 10 for LLM applications.

NLP and ML pipelines

For FPT Software, a team of 6 engineers built a semantic analysis platform for call-centre QA with Python, PyTorch, Transformers and Kafka. It transcribes, classifies and scores every call, and tags topic, intent, sentiment and compliance markers. The speech-to-text stage was tuned for Vietnamese, including calls that switch between Vietnamese and English. QA coverage went from manual sampling to every call analysed.

Evaluation and monitoring

An AI feature without a test set is a demo. A senior engineer collects real inputs with known good outputs before tuning anything, reruns that set on each prompt or model change, and tracks quality, cost per request and latency once the feature is live.

Not on this list: computer vision. Image and video models are a different specialism and we have no shipped vision project to show you, so we do not staff for it. If that is your problem, hire elsewhere for it.

These are failures of ordinary engineering, not of the model. Risk codes are from the OWASP Top 10 for LLM applications, 2025.

1. No evaluation set. The team tries a few prompts by hand and ships. The fix: a few dozen to a few hundred real examples with expected outputs, agreed with the person who owns the outcome.
2. Prompt changes shipped without regression tests. A prompt is code. It lives in version control, the evaluation set runs in CI when it changes, and the model version is pinned so a provider upgrade arrives as a tested change.
3. Cost per request ignored. Log tokens per request from day one, use the smallest model that passes the evaluation set, cache repeats, cap output length and rate-limit callers. OWASP calls the unmanaged version "unbounded consumption" (LLM10).
4. No fallback when the model is wrong. It will be, sometimes (LLM09, misinformation). Show sources, send low-confidence cases to a person, and let users correct the output.
5. Private data sent to a third-party API without review. Before the first call, list which fields leave your network and read the provider's retention and training terms. Strip what the task does not need, and get sign-off from whoever owns data protection (LLM02, sensitive information disclosure). Treat retrieved documents and user text as untrusted input, because prompt injection is LLM01.
6. Latency. Set a budget early, stream partial output, run independent calls in parallel, and move slow work to background jobs.

None of this is exotic. It is senior engineering applied to a component that guesses.