Retrieval that still works on your worst documents
Our engineers have built retrieval for oncology trial matching, ESG research and cruise booking. They work in your repo and your standups, and you interview them before anyone starts.
What you are getting
4 things that decide this
- 01Senior engineers only, each with retrieval systems already running in production for named clients you can read about on this site.
- 02You interview every engineer before they join, so nobody is assigned to your team from a profile you never saw.
- 03They embed in your tools and your standups rather than working behind a ticket queue, and they stay on the same project instead of rotating.
- 04If the fit is wrong, say so and we replace the engineer. Every line they write is yours from the first commit.
What a RAG engineer actually does here
Very little of this job is prompting. Most of it is data work and permissions.
On TrialTriage they built ranked trial matching over de-identified patient data, with a nurse approving every result and an audit trail covering 23 tracked action types. On Greenlight they built research across more than 50 ESG topics, pulling 10 to 15 independent sources per topic so scores never rest on a company's own reporting. On Military Cruise Deals they retrofitted conversational search into a live WordPress site using FastAPI and LangGraph.
Three different industries, one repeated skill: getting the right passage back, for the right person, and proving it stayed right.
What they take off your roadmap
Ingestion for documents you did not choose
Scans, exports, tables pasted into Word. They build a pipeline per source instead of forcing one template over all of them.
Retrieval tuning against real questions
Chunk sizes, hybrid keyword and vector search, re-ranking. Adjusted against the questions your users actually ask, not a sample set.
Multi-tenant isolation
Tenant tags at ingestion and filters inside the query, plus a test that tries to cross the boundary and must fail.
Evals that survive you
A scored question set in your repo, running on every release. It is the artefact that keeps working after the engagement ends.
- Scoping callFree. What you need built
- ShortlistEngineers matched to the work
- You interviewYour process, your bar
- EmbedYour repo, standups, tools
- ReviewSwap if the fit is wrong
The interview is yours. A marketplace that assigns a vetted profile is skipping the only step that predicts fit.
Retrieval over documents we did not choose
“Hashlogics' fast turnarounds stand out.”
Sil van der Woerd · Studio Birthplace
How hiring works
- 01
Tell us what is stuck
A free call about the system, the documents and the deadline. If retrieval is the wrong answer to your problem, you will hear that on the call.
- 02
Meet the engineers
We shortlist people who have tuned retrieval against real user questions, and you interview them. Say no and we go back to the shortlist.
- 03
They embed
Your repo, your standups, your ticket system. One of our engineers owns the retrieval quality and is named as the person accountable for it.
- 04
They hand over
The ingestion pipeline, documentation and the scored eval set, plus someone on your team trained to run it. Where a client would rather we kept watching accuracy, we do that under a service level we agree.
Stack
Models and orchestration
Data and infrastructure
Practices
Tell us what your assistant keeps getting wrong
Bring the questions it fails and the documents behind them. The scoping call is free, and you will leave it knowing whether this is a retrieval problem or a data problem.
01How is this different from a freelancer or a job board hire?
You get an engineer who has already shipped retrieval to production, plus a company accountable if it goes wrong. A job board hands you a candidate and a hiring risk you carry alone. If our engineer is not right, we replace them, which is not a conversation you can have with a contractor you found yourself.
02Can we swap the engineer if the work shifts?
Say the word and we swap them out. You interviewed them, so this is rare, and it stays our problem rather than becoming a hiring cycle you have to run again.
03How much overlap do we get with our working day?
A daily overlap with your hours, fixed before anyone starts. Standups, reviews and pairing happen while both sides are awake, because retrieval debugging is a conversation rather than a ticket. We agree the window during scoping and it does not move without you.
04Who owns the index and the eval set?
All of it is yours from the first commit. That includes the ingestion pipeline, the index and the scored question set, which is the piece that tells you months later whether answer quality has drifted.
05What actually drives the cost of a retrieval build?
The state of your documents, more than anything else. Clean text with consistent structure is quick; scanned PDFs, mixed languages and tables pasted into Word are the work. Scoping calls are free. Where we have to get into an existing codebase before answering, a paid two-week diagnostic produces a fixed price rather than a guess.
