Data Scientist + Applied AI Engineer

I build data products that make difficult decisions easier to inspect.

I work across applied AI, data science, analytics, and data engineering—turning uncertain questions into useful systems with clear boundaries, measurable behavior, and interfaces people can actually reason about.

View selected work Read my profile
Portrait of Vaibhav Khurana
Vaibhav KhuranaData Scientist + Applied AI Engineer

Selected work

The question comes first.

Each case study begins with the decision a stakeholder needs to make, then shows the product, architecture, trade-offs, and limits.

View all work
01Working technical prototype

Financial services

Guarded Text-to-SQL

A review-first interface for natural-language analytics

My roleApplied AI engineer · Data engineer

The question

How can a data team offer natural-language exploration without letting model-generated structured query language (SQL) execute unchecked?

The answer

Separate model proposal from execution, validate every proposal with deterministic policy, and require human approval before a bounded read-only preview can run.

Demonstrated valueTurns generated SQL into a reviewable proposal and prevents direct model-to-database execution.

PythonFastAPIDuckDBSQLGlotAzure OpenAI
02Deployed technical prototype

Legal technology

Legal Discovery Graph

Evidence-linked investigation across documents and entities

My roleAI engineer · Data engineer

The question

How can an investigator move from a large discovery record to evidence they can inspect and trace?

The answer

Combine vector and graph retrieval with cited evidence, structured entities, timelines, and calibrated refusal.

Demonstrated valueConnects documents, people, events, and citations in one inspectable investigation workflow.

PythonFlaskLangChainsentence-transformersONNX Runtime
03Deployed technical prototype

Legal technology

Legal Document RAG

Grounded research over a controlled public corpus

My roleAI engineer · Data engineer

The question

How can legal research return an evidence-linked answer without wandering beyond an approved public corpus?

The answer

Register and checksum sources, preserve document structure, retrieve from a promoted index, and require citations or refusal.

Demonstrated valueTurns public legal documents into a searchable corpus with evidence-linked answers and an explicit refusal path.

PythonFlaskAzure Document IntelligenceAzure OpenAIAzure AI Search

How I make decisions

Good applied AI work is mostly judgment made visible.

I start by clarifying the user and the decision, then choose the smallest data and product system that can be evaluated honestly.

  1. Frame the decision

    Clarify the user, source boundary, and decision the system must support.

  2. Structure the data

    Design durable ingestion, retrieval, and data models around the source material.

  3. Evaluate the system

    Measure behavior against known evidence before trusting a polished interface.

  4. Deliver with clarity

    Ship a useful product with observable tradeoffs, documentation, and disclosure.

Capabilities

Tools connected to work, not a logo wall.

Technology matters here because of the role it played: generating, structuring, validating, retrieving, or delivering something a person could use.

Azure OpenAIApplied AI

I use models inside explicit product boundaries: curated context, inspectable output, and deterministic checks where they matter.

Guarded Text-to-SQL
SQLGlotEvaluation and controls

I turn risk statements into executable policies and bounded evaluations instead of relying on prompt instructions alone.

18 deterministic policy cases
PythonData systems

I build the ingestion, orchestration, service, and evaluation layers that connect analytical questions to usable products.

Legal Discovery Graph
Neo4jConnected analysis

I combine document retrieval with structured relationships when the decision depends on people, events, and evidence paths.

Graph-expanded investigation

Experience

Currently a Data Analyst at Morgan & Morgan, P.A..

My path runs from business analytics and teaching into legal operations, data products, and applied AI systems. The common thread is making complex information usable without hiding its limits.

View the full timeline