Hire engineers who fix retrieval end to end.
Most LlamaIndex failures start at ingestion: a parser that flattened the tables, chunks that lost their headings, metadata nobody extracted. Fixing them means working the whole path (node parsing, index choice, query routing, response synthesis) and measuring retrieval separately from generation. Our engineers have shipped LlamaIndex systems in production, and harder stacks than it.
What they own.
- Build ingestion pipelines where node parsing preserves the structure the answer depends on
- Extract and use metadata so filtering does work retrieval cannot do alone
- Choose between vector, keyword, and hybrid indexes per corpus instead of by default
- Route queries and decompose multi-part questions rather than retrieving once and hoping
- Tune response synthesis modes against a measured recall target
- Evaluate retrieval separately from generation, so the failing half is known
What actually breaks, and what they do about it.
The parser destroyed the answer
A table flattened into prose or a heading dropped during parsing removes the exact structure the answer depended on, and no amount of reranking recovers it. Ingestion is where most retrieval bugs are actually introduced, and where they have to be fixed.
Metadata nobody extracted
Filtering on document type, date or owner does work that semantic similarity cannot, and it is cheap. Corpora that skip metadata extraction end up retrieving the right topic from the wrong year.
One index for every question
Vector search misses exact identifiers users type verbatim; keyword search misses paraphrase. Choosing per corpus, and keeping both where the questions vary, is a decision that has to be measured rather than assumed.
Retrieval and generation measured together
When only the final answer is scored, a retrieval failure and a generation failure look identical, and teams tune the prompt to fix an indexing bug. Recall has to be measured on its own before generation quality means anything.
What you can ship with them.
Tools they reach for in production.
- Framework
- LlamaIndexNode parsersQuery engines
- Index
- VectorKeywordHybrid
- Stores
- pgvectorQdrantWeaviate
- Measure
- Recall@kFaithfulnessEval suites
Seniority: Engineers who have already worked on this stack, and on harder ones than it.
Five stages.
The top 3% remain.
Every stage asks the same question: can they keep AI running once real customers are using it? Getting something started is the easy part, and it is not what we screen for.
400 applicants
3%of applicants reach
your shortlist
What removes them
We start with something they built
100% → 20%One real system, pressed hard. How much traffic did it take? What broke first? Who got the call when it did?
We break something and watch them fix it
20% → 9%A working system with a bug hidden inside it. Anyone can build a demo in a weekend. Fixing code you have never seen is the actual job.
How will you know it is working?
9% → 5%Before they write anything, they have to tell us how they would test it, and what they would do when it gets an answer wrong.
Make it fast without running up the bill
5% → 4%We give them a speed target and a budget, then ask them to explain the tradeoffs they made to hit both.
It is late and the AI got it wrong
4% → 3%What do you do first? How do you find out what happened, undo it, and explain it to the customer in plain words?
Often hired together.
RAG & Retrieval Engineers
Retrieval specialists who ground every answer in your data.
Data Engineers for AI
The pipelines that turn raw data into retrieval-ready, training-ready fuel.
AI Agent Engineers
Builders of agents that plan and take real actions in your systems, safely.
Find the people to accelerate your roadmap.
You don’t need more resumes. You need proven AI engineers embedded in your workflow and ready to build from day one. Tell us what’s missing and we’ll line up a shortlist.