New opportunity

AI Researcher: Entity Resolution & Knowledge Graphs

Auric AI · Bengaluru/Mumbai, India

About this role

Auric AI seeks an AI Researcher to own entity resolution and knowledge-graph research for an intelligence system built to help India connect information across documents and identify emerging threats. The central challenge is determining whether different mentions refer to the same person without creating a confident but false identity. The posting lists an experience range of 0–10 years and places the full-time role onsite in Bengaluru or Mumbai.

Responsibilities

Develop approaches to resolving identities across scripts, transliterations, abbreviations, honorifics and descriptions without names. Standard edit-distance, phonetic and multilingual-embedding methods may fail when records lack comparable forms or distinct people share name components.
Design candidate generation and decision thresholds at a scale where all-pairs comparison is impractical. Account for the much higher cost of a false merge than a missed match, including when very few labels are available.
Extract dependable entities from mixed-language, inconsistently formatted reports and OCR'd handwriting, with trustworthy hallucination detection.
Build a growing knowledge graph that retains full provenance, revisits earlier merges and makes the consequences of an incorrect merge traceable when it is undone.
Identify structural patterns, anomalies against learned baselines and signals that can be surfaced before someone asks a specific question. Rare-event base rates make overall accuracy insufficient; performance at the top of a ranked list matters.

Qualifications and application:

Able to decompose ambiguous problems, explain tradeoffs and design solutions; understands embeddings and retrieval mechanically and has worked with messy real-world data. The role requires attention to errors whose consequences are disproportionate to their frequency.
Direct experience with entity resolution, knowledge graphs or multilingual NLP is a strong signal, not a prerequisite. Backgrounds in linguistics, information retrieval, database theory or formal methods are also of interest. This is not a role focused on RAG over a document store or maintaining an existing ontology.
Apply by emailing a resume and up to two pages describing how to match two records without labelled pairs when false matches are far costlier than missed ones, including threshold selection and evidence that the approach works, to careers@auricai.in.

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

Skills for this role

Entity resolutionKnowledge graphsMultilingual NLPEmbeddingsInformation retrievalEntity extractionOCRHallucination detectionAnomaly detectionCandidate generationLinguisticsDatabase theoryFormal methods

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

Resolve identities across inconsistent mentions
The role centers on deciding whether mentions across scripts, transliterations, abbreviations, honorifics and nameless descriptions refer to the same person, without falsely combining distinct people.
Make cautious decisions at scale
Design candidate generation when all-pairs comparison is impractical, and set merge thresholds that reflect the much higher cost of a false match, even with very few labels.
Extract dependable entities from difficult sources
Work with mixed-language, inconsistently formatted reports and OCR'd handwriting; the posting emphasizes trustworthy detection of invented or unsupported extraction results.
Maintain a correctable, provenance-rich graph
Build a growing knowledge graph that preserves source provenance, revisits earlier merges and makes downstream effects traceable when a bad merge is undone.
Surface useful signals before a query
Find structural patterns and anomalies against learned baselines. The posting stresses performance near the top of a ranked list rather than overall accuracy.

Likely challenges to think through

Candidate coverage versus comparison cost
Preparation hypothesis: cheap blocking could miss the very cross-script or nameless matches the role targets, while broad retrieval could produce too many pairs. Prepare to explain a staged candidate-generation design and how you would examine missed candidates.
Thresholds without reliable labels
Preparation hypothesis: scarce labels make a seemingly precise score hard to justify, especially when false merges cost more than missed matches. Prepare a defensible way to choose conservative thresholds, defer uncertain cases and gather evidence that the decisions work; this is also the stated application exercise.
Correcting a merge without hidden graph damage
Preparation hypothesis: splitting one mistaken identity may require revisiting relationships and conclusions derived from it. Prepare to describe how source-level provenance and dependency tracking would make those effects findable and reviewable.
Ranking rare signals amid noise
Preparation hypothesis: when relevant events are rare, a system can look accurate overall yet put mostly false alarms in its highest-ranked results. Prepare to discuss evaluation at the top of the list and how uncertain extraction or identity links should affect signal confidence.

Practice your answers

Technical practice

How would you decide whether mentions in different scripts refer to the same person?

What to coverPreparation suggestion (inference): Discuss transliteration ambiguity, abbreviations and honorifics, and explain why a shared spelling cannot be assumed. Describe what corroborating or contradictory evidence you would seek before merging identities.

How would you generate candidate matches when all-pairs comparison is impractical?

What to coverPreparation suggestion (inference): Propose ways to retrieve plausible candidates even when names have no comparable surface form. Explain the trade-off between reducing comparisons and missing a true match before the decision stage.

How would you assess a possible match when one record describes a person without naming them?

What to coverPreparation suggestion (inference): Separate evidence about identity from details that could apply to many people. Explain when the available evidence warrants a match, further review or leaving the records separate.

Where could multilingual embeddings or phonetic matching mislead an identity-resolution system?

What to coverPreparation suggestion (inference): Explain mechanically how similarity can arise without shared identity, particularly when distinct people share name components. Compare what these methods might contribute to candidate retrieval with the stronger evidence needed for a merge.

How would you set a merge threshold with no labelled pairs and a high cost for false matches?

What to coverPreparation suggestion (inference): Make the asymmetric costs explicit and describe a conservative decision rule, including an option to abstain. Explain what assumptions or limited checks could inform a threshold, and what uncertainty would remain.

What evidence would convince a sceptic that an entity-resolution approach works?

What to coverPreparation suggestion (inference): Describe how you would inspect likely false merges and missed matches when labels are scarce, without treating an unverified sample as ground truth. Discuss how you would test sensitivity to rare matches and report limitations alongside results.

How would you extract dependable entities from mixed-language reports and OCR'd handwriting?

What to coverPreparation suggestion (inference): Address uncertain text recognition, inconsistent formatting and the risk of inventing an entity not supported by the source. Explain how each extracted claim could retain its source evidence and be flagged for verification.

How would you design a growing knowledge graph so an incorrect merge can be undone?

What to coverPreparation suggestion (inference): Describe how provenance and merge decisions could be represented so affected relationships and downstream conclusions are traceable. Explain the trade-off between convenient consolidated identities and preserving enough history to revisit them.

How would you surface unexpected structural patterns or anomalies in the graph?

What to coverPreparation suggestion (inference): Explain how a learned baseline might distinguish unusual connections from ordinary variation as the graph grows. Address how uncertain identities or extraction errors could create misleading signals.

How would you evaluate a ranked list of rare-event signals?

What to coverPreparation suggestion (inference): Explain why overall accuracy can conceal poor performance when relevant events are rare. Discuss performance among the highest-ranked results, the cost of false alerts, and how you would examine whether apparent patterns come from graph errors.

Behavioural practice

Tell me about a time you decomposed an ambiguous research problem.

What to coverPreparation suggestion (inference): Use a real situation, the steps you took to identify assumptions and trade-offs, and the result. Focus on how you chose a testable direction without claiming the problem had a known answer.

Describe a time you handled an error with disproportionate consequences.

What to coverPreparation suggestion (inference): Give the situation, what you did to detect or limit the error, and the result. Explain how its potential cost changed your decision-making rather than relying on an aggregate success rate.

Tell me about a time messy data changed your technical approach.

What to coverPreparation suggestion (inference): Describe the data problem, the action you took to investigate or adapt, and the result. Distinguish what you verified from what remained uncertain.

Describe a time you explained a difficult technical trade-off to others.

What to coverPreparation suggestion (inference): Set out the decision and competing costs, how you communicated the evidence and uncertainty, and the result. A relevant example could involve balancing coverage against the risk of a consequential false positive.

Tell me about a time new evidence made you revisit an earlier decision.

What to coverPreparation suggestion (inference): Explain the original decision, the new evidence, what you changed, and the result. Include how you identified any work affected by the earlier decision.

Sharpen your knowledge

Topics to prepare

Identity resolution across scripts
The posting centers on deciding whether mentions with different scripts, transliterations, abbreviations, honorifics or no name refer to one person. Suggested exercise: compare three ambiguous record pairs and explain where edit distance, phonetics and multilingual embeddings could produce false merges.
Asymmetric decisions with few labels
A false merge is much costlier than a missed match, and the role involves setting thresholds with very few labels. Suggested exercise: work through a decision rule with a conservative merge threshold, an uncertain category and a plan to test how conclusions change as the assumed error costs change.
Candidate generation and retrieval mechanics
The posting rules out all-pairs comparison and warns that candidates may lack comparable surface forms; it also expects mechanical understanding of embeddings and retrieval. Suggested exercise: sketch a candidate-generation approach, estimate its comparison count and identify true matches it might omit.
Extraction reliability in messy documents
Source reports may mix languages, vary in format and include OCR'd handwriting; hallucination detection is an explicit concern. Suggested exercise: trace how one uncertain OCR reading could become an unsupported entity or identity claim, and specify what evidence would prevent that promotion.
Knowledge-graph provenance and correction
The graph must retain full provenance, revisit merges and expose downstream consequences when a merge is undone. Suggested exercise: draw a small graph with two merged identities, document-level evidence and dependent relationships, then trace what must be reconsidered after a split.
Rare-event signals and ranked evaluation
The role calls for structural patterns, anomalies against learned baselines and useful signals near the top of a ranked list; overall accuracy is insufficient at rare-event base rates. Suggested exercise: explain how two rankings with similar overall accuracy could differ sharply in the value of their highest-ranked items.

Skills to revise

Draft an evidence-based matching decision
Suggested practice: write a decision record for two conflicting person mentions, separating supporting evidence, contradictions and unknowns before choosing merge, do not merge or defer. This exercises the posting's core identity-resolution problem without treating a name similarity as proof.
Sketch a scalable candidate pipeline
Suggested practice: specify how records enter a candidate set, where retrieval narrows comparisons and how you would inspect missed candidates. This addresses the stated scale constraint and the failure of surface-form-only generation.
Audit entity extraction against source text
Suggested practice: annotate a short, invented mixed-language report with extracted mentions, source spans and uncertain OCR readings; reject any entity claim without supporting text. This rehearses dependable extraction and hallucination detection, not a prescribed employer workflow.
Walk through a merge reversal
Suggested practice: record the documents and decisions behind an invented graph merge, then list every affected relationship and derived conclusion when it is reversed. This directly tests the posting's requirements for provenance and traceable correction.
Evaluate conservative decisions and ranked signals
Suggested practice: propose a provisional threshold without labelled pairs, state its assumptions, and describe evidence that would challenge it; separately compare the top-ranked anomaly candidates against a baseline. This connects the application prompt's threshold question with the posting's emphasis on costly false merges and top-of-list performance.
Write the requested design memo
Suggested practice: draft no more than two pages explaining a no-labelled-pairs record matcher, threshold choice and what evidence would convince a sceptic. This mirrors the posting's application exercise; present any proposed validation or review procedure as your suggestion, not an established company process.

Your preparation checklist

Prepare a project walkthrough
Choose a real example involving messy data, retrieval, entity resolution or graphs, if available. Explain the problem, design choices, failure cases and evidence of results; if using a hypothetical example, label it clearly.
Design candidate generation
Sketch how to narrow possible identity matches without all-pairs comparison, including mentions across scripts, abbreviations, honorifics and records with no name. Explain where edit distance or embeddings could produce false matches.
Define a conservative merge decision
Write down the evidence that would support a merge, trigger review or keep records separate. Propose a threshold-setting approach for no labelled pairs that reflects the posting’s much higher cost of false merges.
Plan a sceptical evaluation
Specify what evidence would test the matching approach despite scarce labels, such as targeted review of proposed merges and hard cases involving shared name components. Address false-merge risk rather than relying on overall accuracy.
Trace an entity from document to graph
Walk through a proposed design for extracting entities from mixed-language or OCR’d reports, flagging unsupported extraction, retaining source provenance and finding downstream effects when a merge is undone.
Sketch a ranked-signal exercise
As an interview preparation exercise, describe one way to detect graph anomalies against a baseline. Explain how you would inspect the top-ranked signals when genuine events are rare.
Prepare questions about research tradeoffs
Ask how the team would judge evidence for a merge, revisit disputed identities and evaluate top-ranked signals. Keep the questions open: the posting does not specify these processes.
Finish the requested application materials
Prepare a resume and up to two pages answering the posting’s no-labelled-pairs matching prompt. Cover the proposed build, threshold choice and evidence that would convince a sceptic; check that the claims match work actually done.

Availability last checked: 4 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…