About the role
A document assistant can give a convincing answer using a policy that expired yesterday. As Senior AI Engineer, you’ll own how a question moves from the search index to the answer on screen, and how a reviewer can check which version of the policy supported it.
You’ll work with the client’s document owners to establish what counts as an authoritative answer. That definition has to survive revised policies, conflicting passages and questions for which the collection has no answer. You’ll carry it through retrieval, generation and the interface a reviewer uses to check the result.
Working at ENNPRO
ENNPRO is an applied machine learning consultancy. We work alongside client engineers and the people who use their systems. Depending on the engagement, a client might need a service they can operate, a benchmark they can reproduce or enough evidence to decide whether to release a model. Their team should be able to use the result after handover.
A document owner may know instantly why a certain appendix overrides the main policy. The engineer building the assistant rarely starts with that knowledge. You’ll sit with both people, learn how they settle a disputed answer, and decide which parts of their reasoning belong in the index, the application or the review interface. You’ll also work with the person responsible for access to each collection. An answer can be accurate and still be inappropriate for the person asking.
As the senior engineer, you’ll divide the work into changes that teammates can test independently. Expect to review ingestion code, pair on a difficult search failure and explain a release decision to someone who works with the documents every day. The client’s team should be able to investigate the next bad answer using the tools you leave them.
Responsibilities
- Build a test collection from real questions. Work with reviewers to mark the passages that support an answer and the cases where the assistant should ask for clarification.
- Compare lexical, vector and hybrid retrieval on the same questions. Inspect which passages disappear at each stage, including during chunking, filtering and reranking.
- Design a citation that resolves to the passage and version the model used. Make it possible to inspect an incorrect answer even after the source document has been revised.
- Ship the serving path with the client’s engineers. Measure response time and cost alongside evidence quality, and agree on a fallback when retrieval or generation fails.
- Build an update path for revised and withdrawn documents. Define when a new edition becomes searchable, when the previous one stops appearing, and how to demonstrate that the index reflects the source collection.
- Make access scope part of retrieval. Write tests with two users who have different permissions and the same question. Check the returned passages, citations and cached responses for each user.
- Own the operating costs of a proposed change. Estimate the effect on indexing time, tokens per answer and waiting time, then use measured traffic to check the assumptions behind the estimate.
Day to day
One week might include an indexing change and an afternoon spent reviewing twenty poor answers with a client specialist. Bring the original questions to that session. Ask which words the user would recognize, which document an experienced colleague would open and what evidence would make them change their mind. Those notes make a better starting point for the next evaluation than a collection of plausible questions written from scratch.
For implementation, expect to move between Python, search indexes, document parsers and ordinary application services. Keep a small replay that runs without a live model call so a reviewer can reproduce an indexing defect and an engineer can test a correction in a pull request.
Projects and first 90 days
A representative assignment
A policy changes, but an assistant keeps quoting the earlier version. Start with one disputed answer and replay the original request. Check whether the current document reached the index, whether the old passage ranked higher, and whether the answer cited the passage it actually used.
A good handoff includes the failing query, a regression test that distinguishes the two versions, and a trace a support engineer can follow when a similar report comes in.
There is also a product decision in this example. If two policies carry the same effective date, the interface needs a way to present the conflict. Work with the owner of the material to agree on the wording and decide which detail the reviewer needs to see before choosing an edition.
How your first engagement could unfold
First month. Choose one document collection and follow a real question through it with a reviewer. Set up a local replay, map where the text and permissions come from, and annotate a small set of questions with the passages that answer them. Take ownership of one failure that colleagues already recognize.
Second month. Compare a targeted change against the current service. It could be a better split for tables, a version filter or a change in ranking. Share representative wins and regressions with the owner of the collection and agree on the conditions for a limited release.
Third month. Put the change through a release rehearsal. Invite another engineer to diagnose a deliberately stale index using your traces and recovery notes. Use the points where they need your help to improve the handover, then identify which unanswered questions belong in the next round of work.
Qualifications
Core experience
- You’ve operated a search or language application and traced a user-visible failure across at least two of its components.
- You can build a Python service, test it with realistic inputs and explain how batching, timeouts or caching affect its behavior.
- You can lead an error review with a domain specialist and turn a disagreement about an answer into a test the two of you trust.
- You understand at least one retrieval stack well enough to explain how documents are indexed, where ranking occurs and what a query log can tell you about an unexpected result.
- You have made release or rollback decisions for a service used by other people and can explain the operational trade you accepted.
Useful additional experience
- Experience parsing tables, scanned pages or multi-column documents where reading order changes the meaning of the text.
- Familiarity with relevance judgments, rerankers or experiments comparing search results across languages and collections.
- Experience maintaining versioned document repositories or application permissions, particularly where users need an audit trail for a decision.
How to apply
Describe one query that exposed a weakness in a search or language system. What did you inspect first, what changed, and how did you check the fix?
A useful example could be a failing search you reconstructed, a release you postponed after evaluation, or a citation problem that turned out to originate in parsing. Describe your own contribution and what changed for the person trying to use the answer. An anonymized query and a few sentences about the evidence are enough.
You can use an anonymized example for confidential work. The application form includes space for your account of the project, an optional résumé and an optional link to a work sample. It currently lets you prepare and review a draft in your browser.
Equal opportunity and accessibility
ENNPRO welcomes applicants of all backgrounds. We consider candidates on the skills and experience relevant to the role, including experience gained through different educational and career paths. If you need an accessible version of these materials or an adjustment to the application process, write to contact@ennpro.pro.
Start application