New opportunity

Founding AI Engineer - Proofline

Meraki-Labs · Bengaluru, India

About this role

Meraki Labs is hiring a Founding AI Engineer for Proofline, a production product with institutional pilots. The engineer will own retrieval and knowledge capabilities end to end, connecting published research and private institutional information for research institutions and enterprise research teams. The role is based in HSR, Bengaluru, five days a week on-site; remote and hybrid arrangements are not offered.

Responsibilities

Own the retrieval pipeline, including document ingestion, chunking, embeddings, hybrid search and reranking.
Build retrieval across published literature and private institutional documents while preserving provenance. Develop citation and DOI/arXiv-class identifier resolution and support onboarding of large corpora.
Build a retrieval evaluation harness, define metrics and demonstrate quality improvements across releases.

Requirements

At least 7 years of experience, with ML or backend engineering experience involving production RAG, search or information retrieval systems.
Strong Python skills and hands-on experience with embeddings, vector stores and rerankers. Candidates must be able to show how they have measured and improved retrieval quality.
Graph database experience and familiarity with scientific or technical document corpora are pluses, not requirements.

Working approach:

The small team uses written decisions, reviews designs before implementation and verifies features on a live system. Pull requests undergo automated review until no findings remain. AI agents perform a substantial share of implementation, so the role calls for judgment in defining problems, settling designs and identifying incorrect results. Production engineering concerns include monitoring, reliability, cost and latency.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

PythonRAGInformation retrievalDocument ingestionDocument chunkingEmbeddingsHybrid searchRerankingVector storesRetrieval evaluationCitation resolutionData provenanceBackend engineeringProduction monitoringReliability engineeringGraph databasesAI agents

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Own retrieval end to end
The role owns document ingestion, chunking, embeddings, hybrid search and reranking. Be ready to explain how design choices across the pipeline affect retrieval quality.
Connect two corpora with provenance
Proofline retrieves across published literature and private institutional documents. Preserving where retrieved material came from is an explicit responsibility.
Resolve citations at scale
The role includes DOI/arXiv-class identifier resolution, citation resolution and onboarding large corpora. Prepare to discuss approaches to document identity and source traceability.
Prove retrieval improvements
The engineer will build an evaluation harness, define metrics and demonstrate quality gains across releases. The posting specifically asks for evidence of measured improvements.
Exercise production design judgment
The team reviews designs before implementation, uses AI agents for substantial implementation and verifies features on a live system. Monitoring, reliability, cost and latency are also named production concerns.

Likely challenges to think through

Balancing coverage with trustworthy attribution
A likely challenge is retrieving useful results from both corpora without losing provenance or misresolving citations. Prepare a design example that explains how you would detect and handle ambiguous identifiers or conflicting sources; this is a preparation hypothesis, not a reported product issue.
Showing gains without hiding operational costs
Improved retrieval metrics may come with higher latency or cost in a production system. Prepare to explain how you would compare releases using an evaluation harness while checking live-system behavior; this trade-off is inferred from the stated evaluation and production duties.
Moving quickly while catching incorrect output
In a small team where AI agents implement substantial work, speed can conflict with design quality and correctness. Prepare an example of how you would frame a problem, review a design and verify an implementation on a live system; the difficulty is inferred from the described working approach.

Practice your answers

Technical practice

How would you design ingestion and chunking for a research-document retrieval pipeline?

What to coverThe role owns document ingestion and chunking across multiple sources. Preparation suggestion: explain how you would preserve document structure and metadata, and how you would test chunking choices against retrieval quality.

How would you choose and operate embeddings and a vector store?

What to coverThe posting calls for hands-on work with embeddings and vector stores. Preparation suggestion: discuss relevance, update behavior, indexing cost and latency, using measured results to justify choices.

When would hybrid search outperform a single retrieval method?

What to coverHybrid search is part of the stated pipeline. Preparation suggestion: compare lexical and embedding-based retrieval for exact research terms versus semantic matches, then explain how you would evaluate their combination.

How would you evaluate whether a reranker improves retrieval?

What to coverThe role includes reranking and requires demonstrated retrieval evaluation rigour. Preparation suggestion: describe where reranking fits in the pipeline and how you would weigh relevance gains against latency and cost.

How would you retrieve across published literature and private institutional documents while preserving provenance?

What to coverDual-corpus retrieval with provenance is an explicit responsibility. Preparation suggestion: explain how results would retain source identity and citation context, and how you would check that provenance survives each retrieval stage.

How would you resolve citations and DOI- or arXiv-class identifiers?

What to coverThe posting explicitly calls for citation and identifier resolution. Preparation suggestion: discuss normalization, ambiguous or missing identifiers, duplicate records and tests for incorrect matches.

How would you onboard a large corpus without degrading an existing retrieval system?

What to coverLarge-corpus onboarding is a stated responsibility, and the product is already live. Preparation suggestion: cover staged ingestion, validation, reprocessing and monitoring, with trade-offs among freshness, reliability and cost.

What would you put in a retrieval evaluation harness?

What to coverThe role owns the evaluation harness and its metrics. Preparation suggestion: describe representative queries, relevance judgments, corpus-specific slices and repeatable metrics that expose failures rather than only an aggregate score.

How would you demonstrate retrieval improvement across releases?

What to coverThe posting asks for quality improvements to be proved release over release. Preparation suggestion: compare against a fixed baseline, inspect regressions by query type and distinguish meaningful quality gains from changes in latency or cost.

How would you troubleshoot a production retrieval-quality regression?

What to coverThe posting emphasizes live-system verification, monitoring, reliability, cost and latency. Preparation suggestion: trace the regression through ingestion, indexing, retrieval and reranking, then explain what evidence would justify a fix or rollback.

Behavioural practice

Tell me about a time you turned an ambiguous problem into an end-to-end engineering plan.

What to coverThe role emphasizes ownership and turning messy problems into concrete solutions. Preparation suggestion: use a real situation, your actions to define the problem and make trade-offs, and the resulting outcome.

Tell me about a design decision that changed after critical review.

What to coverThe team describes written decisions and adversarial design review before implementation. Preparation suggestion: give a real example covering the initial proposal, how you handled objections and the result.

Tell me about a time review findings or live verification exposed a flaw in your work.

What to coverThe posting describes automated pull-request review and verification on a live system. Preparation suggestion: explain the situation, what you changed in response and how you established the final result.

Tell me about a time you caught an incorrect result from an AI-assisted implementation.

What to coverThe posting says AI agents perform a substantial share of implementation and stresses knowing when they are wrong. Preparation suggestion: use a real example that shows how you detected the error, corrected it and verified the outcome.

Tell me about a production trade-off you made under time pressure.

What to coverThe role highlights a fast-moving environment and production concerns including reliability, latency and cost. Preparation suggestion: describe the situation, the options and evidence behind your action, and the observed result.

Sharpen your knowledge

Topics to prepare

Retrieval pipeline design
The role owns ingestion, chunking, embeddings, hybrid search and reranking. Sketch the full pipeline and explain where each stage can improve or degrade retrieval quality.
Dual-corpus retrieval and provenance
The posting specifies retrieval across published literature and private institutional documents with provenance. Trace a hypothetical result back to its source document and explain how that attribution survives ingestion and ranking.
Citation and identifier resolution
The role includes citation and DOI/arXiv-class identifier resolution. Work through examples of incomplete or duplicate references and describe how you would resolve them, including uncertain matches.
Retrieval evaluation
The engineer must set metrics and demonstrate improvement across releases. Define a small query-and-relevance test set, choose metrics appropriate to it, and explain how you would compare two pipeline versions. Specific metric choices are preparation suggestions, not stated requirements.
Production and design tradeoffs
The posting emphasizes monitoring, reliability, cost, latency, written decisions and review before implementation. Practice explaining a retrieval design choice, its tradeoffs and how you would verify its behavior on a live system.

Skills to revise

Implement a Python retrieval pipeline
Strong Python and hands-on experience with embeddings, vector stores and rerankers are required. As practice, build a small pipeline that ingests and chunks documents, retrieves candidates using hybrid search, then reranks them; explain each stage's output.
Build a retrieval evaluation harness
The role explicitly owns the harness and release-over-release quality evidence. As practice, write a repeatable evaluation script that runs the same queries against two pipeline versions and reports per-query results and aggregate changes.
Resolve references and retain source records
Citation resolution and provenance are stated responsibilities. As practice, implement a small routine that matches sample DOI/arXiv-class references to records and carries source identifiers into returned results; flag ambiguous matches rather than silently choosing one.
Review and verify an implementation
The team reviews designs before code, drives automated PR review findings to zero and verifies features on a live system. As practice, write a brief design decision for a pipeline change, review its failure cases, then specify the live checks that would show whether it worked. This is a preparation exercise, not a claim about a particular interview task.

Your preparation checklist

Prepare a production retrieval walkthrough
If you have a relevant RAG, search or IR project, prepare a diagram and concise account of your Python work, including ingestion, chunking, embeddings, vector stores, hybrid search and reranking. Distinguish what you owned from what the team owned.
Sketch a dual-corpus retrieval design
As a preparation exercise, draw how published literature and private institutional documents could be ingested, searched and returned with provenance. Note design choices and unresolved questions rather than presenting them as Proofline’s implementation.
Work through citation resolution
Prepare a small example showing how you would handle citations and DOI/arXiv-class identifiers, including ambiguous or missing matches and a plan for onboarding a large corpus.
Bring evidence of retrieval improvement
If available, assemble a before-and-after example with evaluation data, metrics, the change made and its measured effect. The role explicitly calls for an evaluation harness and proof of improvement across releases.
Practice production trade-offs
Prepare a short design review explaining how you would measure and balance retrieval quality, latency, cost, reliability and monitoring for a live system. Label proposed metrics or thresholds as your suggestions.
Prepare a written decision example
If you have one, bring a concise example of a design decision you documented, reviewed and verified in a live system. Be ready to explain how you caught incorrect results, including in AI-assisted implementation; these reflect the team’s stated working approach.
Draft role-specific questions and confirm logistics
Ask how Proofline currently evaluates retrieval across its two corpora, handles provenance and verifies releases during institutional pilots. Separately, confirm whether five days on-site in HSR, Bengaluru works for you; the posting does not offer remote or hybrid work.

Availability last checked: 5 Nov 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in India. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…