New opportunity

Senior Data Scientist- Conversational AI

HireVeda · Bengaluru, India

About this role

Build and own a conversational AI research assistant for retail investors in Indian capital markets. The assistant will explain financial instruments and macroeconomic analysis to non-expert users and respond to incomplete, real-world questions.

Responsibilities

Design multi-agent pipelines with tool calling, memory management and multi-turn conversations.
Develop retrieval-augmented generation pipelines, including source curation, chunking, embedding quality, retrieval tuning, reranking and citation-grounded answers.
Integrate live financial data from NSE/BSE feeds, Screener, Tickertape, news APIs and company filings as callable tools.
Build guardrails and evaluation processes for domain scoping, hallucination mitigation, confidence scoring and ongoing accuracy monitoring. Fine-tune or adapt LLMs when retrieval alone is insufficient.
Develop machine learning models for user behaviour, personalization and financial insights that support the conversational assistant.

Requirements

4–8 years of hands-on experience in data science, machine learning or applied AI; strong Python proficiency, including scikit-learn, TensorFlow or PyTorch.
Understanding of statistical methods, machine learning algorithms, wealth technology applications, data wrangling, visualization and analysis. Familiarity with LLMs such as Llama and Anthropic Claude, and cloud platforms such as AWS Bedrock and GCP Vertex AI.

Preferred

Capital Markets or WealthTech experience is highly preferred; experience building AI research assistants for financial products or strong personal investing or trading knowledge is also relevant.
Capital-market use cases, recommender systems, personalization, production model deployment, embedding and retrieval improvements, a data science portfolio or open-source contributions, or experience at an AI-first startup.

Location: Bengaluru. The posting lists two openings.

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

Skills for this role

Conversational AIData ScienceMachine LearningPythonscikit-learnTensorFlowPyTorchLarge Language ModelsLlamaClaudeAWS BedrockGoogle Cloud Vertex AIMulti-agent systemsTool callingRetrieval-Augmented GenerationEmbeddingsRerankingLLM fine-tuningRecommender systemsPersonalizationStatistical methodsData wranglingData visualizationModel evaluationFinancial data integration

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 the retail-investor assistant
The role centers on building a conversational research assistant for non-experts in Indian capital markets. Prepare to explain how you would design multi-agent tool calling, memory and multi-turn flows for incomplete user questions.
Deliver citation-grounded retrieval
The posting calls for source curation, chunking, embeddings, retrieval tuning and reranking. Prepare to discuss how you would measure whether retrieved evidence supports an answer and its citations.
Connect live market information
The assistant is expected to use NSE/BSE feeds, financial-data services, news APIs and company filings as callable data sources, not just static documents. Prepare to explain how an answer would distinguish live data from retrieved background material.
Keep answers accurate and within scope
The role includes domain scoping, hallucination mitigation, confidence scoring and ongoing accuracy monitoring. Prepare to describe evaluation cases involving unsupported claims, incomplete questions and uncertain answers.
Extend the assistant with applied ML
The posting includes LLM adaptation when retrieval is insufficient and ML models for user behaviour, personalization and financial insights. Prepare to explain when each approach would add value to the conversational experience.

Likely challenges to think through

Interpreting incomplete financial questions
A likely design challenge is balancing helpful follow-up questions against answering too confidently from an ambiguous retail-user query. Prepare examples showing when the assistant should clarify, answer with stated assumptions or stay within its domain scope.
Reconciling live data with citations
A likely difficulty is producing a coherent answer when tool-fetched market data and retrieved documents differ in timing or context. Prepare a way to evaluate data freshness, source attribution and whether each claim is supported.
Choosing the right improvement lever
When answers are weak, the trade-off may be between improving retrieval, changing tool use and adapting the LLM. Prepare a diagnostic approach that uses error categories and evaluation results before selecting an intervention.

Practice your answers

Technical practice

How would you design a multi-agent research assistant with tool calling?

What to coverThe posting calls for multi-agent pipelines that answer retail investors’ questions about Indian capital markets. Preparation suggestion: explain how you would divide responsibilities, choose when to call a tool, and control complexity when a simpler flow would suffice.

How would you handle an incomplete question across multiple turns?

What to coverThe assistant is expected to handle messy retail-user queries using memory and multi-turn conversation. Preparation suggestion: discuss when to ask a clarifying question, what context to retain, and how to avoid treating an unstated assumption as fact.

How would you curate and chunk sources for financial RAG?

What to coverThe posting names source curation and chunking as parts of the RAG architecture. Preparation suggestion: explain how you would assess source reliability, preserve context such as dates and instrument names, and test whether chunk boundaries hurt answer quality.

How would you improve retrieval quality?

What to coverEmbedding quality, retrieval tuning, and reranking are stated responsibilities. Preparation suggestion: compare approaches using representative investor questions and discuss the trade-off between finding relevant evidence and adding latency.

When should the assistant use live financial data rather than retrieved documents?

What to coverThe role involves callable NSE/BSE feeds, Screener, Tickertape, news APIs, and company filings. Preparation suggestion: describe how you would select a source, check its timestamp and provenance, and handle missing or conflicting data without presenting stale figures as current.

How would you produce a citation-grounded explanation for a non-expert investor?

What to coverThe posting calls for trustworthy, citation-grounded responses about financial instruments and macro analysis. Preparation suggestion: explain how claims would map to supporting evidence, how you would communicate uncertainty, and what you would do when the sources do not support an answer.

How would you evaluate guardrails and monitor accuracy over time?

What to coverDomain scoping, hallucination mitigation, confidence scoring, and ongoing monitoring are explicit responsibilities. Preparation suggestion: outline test cases and measures for unsupported claims, out-of-scope requests, and poorly calibrated confidence, including how failures would inform changes.

When would you adapt an LLM instead of improving retrieval?

What to coverThe posting mentions fine-tuning or adapting LLMs where retrieval alone is insufficient. Preparation suggestion: distinguish a knowledge-access failure from a model-behaviour failure and discuss evidence, evaluation, and maintenance trade-offs before choosing adaptation.

How would you build and assess personalization that feeds the assistant?

What to coverThe role includes ML models for user behaviour, personalization, and financial insights; recommender systems are listed as good to have. Preparation suggestion: describe suitable signals, a baseline and evaluation approach, and how you would avoid mistaking inferred preferences for a user’s stated intent.

How would you take a Python-based model from analysis to a dependable conversational feature?

What to coverThe posting expects Python, statistical and ML methods, data wrangling, visualization, and analysis; production deployment is listed as good to have. Preparation suggestion: walk through data checks, model evaluation, integration with the assistant, and how you would detect degraded performance after release.

Behavioural practice

Tell me about a time you owned an applied AI or data science problem end to end.

What to coverThe role calls for building and owning the research assistant. Preparation suggestion: use a real situation, your decisions and actions, and the result, making clear which parts you personally owned.

Describe a time you made a technical finding understandable to a non-expert audience.

What to coverThe proposed assistant must explain financial instruments and macro analysis to retail users. Preparation suggestion: give the situation, how you adjusted the explanation without losing important caveats, and evidence of the result.

Tell me about a time you found and addressed an accuracy problem in an AI or data product.

What to coverThe posting emphasizes hallucination mitigation, evaluation, and ongoing accuracy monitoring. Preparation suggestion: describe how the issue surfaced, what you changed or escalated, and how you checked the result.

Describe a time incomplete requirements or ambiguous user input changed your approach.

What to coverHandling messy, incomplete questions is central to the advertised work. Preparation suggestion: explain the situation, how you identified the missing information, the action you took, and what happened.

Tell me about a time you chose between competing model or system approaches.

What to coverThe role spans retrieval tuning, LLM adaptation, and ML models that feed a conversational assistant. Preparation suggestion: give a real situation, the evidence and trade-offs behind your decision, your actions, and the observed result.

Sharpen your knowledge

Topics to prepare

Indian capital-market concepts for retail users
The assistant is intended to explain financial instruments and macroeconomic analysis to non-experts; capital-markets or WealthTech experience is highly preferred. Suggested revision exercise: outline how you would explain an instrument and a macroeconomic change in plain language, separating established facts from uncertainty.
Multi-agent conversation design
The posting calls for tool-calling, memory management and multi-turn flows that handle incomplete queries. Suggested revision exercise: sketch a flow that asks for missing context, decides when to call a tool and retains only relevant conversation history.
Retrieval quality and grounded answers
The RAG work spans source curation, chunking, embeddings, retrieval tuning, reranking and citations. Suggested revision exercise: compare two chunking and retrieval approaches for a company filing, then explain how you would check whether an answer's citations support its claims.
Live data versus static retrieval
The role integrates NSE/BSE feeds, Screener, Tickertape, news APIs and filings as live callable sources. Suggested revision exercise: map which parts of a sample investor question need current data, which need retrieved documents and how you would disclose missing or stale data.
Guardrails and ongoing evaluation
The posting specifies domain scoping, hallucination mitigation, confidence scoring and accuracy monitoring. Suggested revision exercise: draft a small test set of ambiguous and out-of-scope questions, with criteria for grounded answers, clarification and uncertainty.
Model adaptation and personalization decisions
The role includes adapting LLMs when retrieval is insufficient and building ML models for user behaviour, personalization and financial insights. Suggested revision exercise: explain when you would improve retrieval, adapt an LLM or use a separate predictive model. The posting also names familiarity with Llama, Claude, AWS Bedrock and GCP Vertex AI.

Skills to revise

Python data analysis
Strong Python proficiency and experience in data wrangling, visualization and analysis are stated expectations. Suggested revision exercise: clean a sample financial-data table, investigate missing values and produce a chart with a short explanation of its limitations.
Statistical and ML modelling
The posting expects statistical and ML understanding and names scikit-learn, TensorFlow and PyTorch. Suggested revision exercise: build a small user-behaviour or personalization baseline with a relevant library, choose an evaluation metric and check for data leakage.
Build and tune a citation-grounded RAG prototype
RAG architecture and retrieval-quality improvement are central to the work. Suggested revision exercise: index a small set of sample filings, test chunk sizes and reranking, and inspect whether generated claims are supported by retrieved passages.
Implement a multi-turn tool-calling flow
The assistant must handle messy retail-user queries with tools and memory. Suggested revision exercise: prototype a conversation using a mocked financial-data tool; test a follow-up question, an incomplete request and an unavailable tool response.
Validate live financial-data integration
The posting calls for live financial sources as tool-callable data layers. Suggested revision exercise: define and test a sample tool response with a source, timestamp and missing-data case, then ensure the assistant does not present unavailable values as current.
Run guardrail and accuracy checks
The role includes confidence scoring and monitoring for accuracy over time. Suggested revision exercise: run a small question set through a prototype, record unsupported claims and incorrect citations, and propose checks that would detect those failures after deployment.

Your preparation checklist

Map your evidence to the role
Prepare a one-page evidence sheet covering the stated 4–8 years of applied AI or data science experience, Python and ML libraries, statistical methods, and any relevant capital-markets or WealthTech work. Mark preferred experience separately.
Rehearse a relevant project walkthrough
Choose an actual conversational AI, financial research, or production ML project and explain your decisions, trade-offs, evaluation, and results. If you lack a close example, label a retail-investor assistant walkthrough as a proposed design rather than past work.
Sketch a multi-turn assistant design
Draw a proposed flow for an incomplete retail-investor question, showing how agents clarify intent, call tools, manage memory, and explain instruments or macroeconomic context to a non-expert.
Prepare a citation-grounded RAG exercise
Outline source selection, chunking, embeddings, retrieval tuning, and reranking for company filings or other financial sources. Include a sample answer with citations and a small test set for retrieval and answer quality.
Design a live-data tool call
Show how a proposed assistant would request and present data from a source named in the posting, such as an NSE/BSE feed or news API. Cover timestamps, unavailable data, and how the response distinguishes live data from retrieved documents.
Define guardrails and evaluation
Prepare example tests for out-of-scope requests, unsupported claims, confidence scoring, and accuracy over time. Explain how you would decide whether retrieval improvements suffice or LLM adaptation merits testing.
Outline a personalization model
Draft a Python-based approach to user-behaviour or personalization signals that could feed the conversational layer. Identify candidate features, a baseline model, evaluation measures, and a way to inspect errors; present it as a proposal unless drawn from your work.
Write questions for the employer
Ask which investor questions are in scope, what financial data and citation sources the team can use, how accuracy is measured, and what ownership of deployment and monitoring the role entails.

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…