● Data & AI Engineer / 5+ years / United States

I build the layer models depend on.

I build the layer underneath analytics and AI in financial services: ingestion from APIs, CRM systems and vendor files, governance that holds, and the tests that keep it honest. Most recently I started Prep Me, a GenAI client-meeting copilot now used by Capital Group SFMs. I took the initiative on it. Opinionated about data quality, because it is the same discipline that separates good training data from noise.

Get in touch GitHub

Jyoshna Goshikahover to focus

Lineage — upstream to production 5 nodes / 0 breaks

2019 — 2021VirtusaHealthcare data platforms, compliance-grade validation

2023 — 2024American ExpressGCP ETL: GCS, Pub/Sub, Spark, BigQuery

2024 — 2025Northern TrustLayered ELT, Neo4j graph models, federated GraphQL

2025 — nowCapital GroupIngestion, Medallion pipelines, dbt on Databricks

FeaturedPrep MeSelf-started GenAI copilot, now used by SFMs

Self-directed 0 → 1 work

Built on my own initiative · used by Capital Group SFMs

Prep Me

GenAI meeting-prep copilot / internal product

A GenAI copilot I took the initiative to build, because I could see SFMs losing time to meeting prep that the firm's own records could have done for them. It is now used by Capital Group SFMs to schedule meetings and build an agenda. An SFM picks their initials, picks a buying unit, and Prep Me surfaces the revenue that SFM has built for the firm, returns the relationship history, drafts the follow-up email to the buying unit contact, and attaches the relevant bios pulled from Salesforce — all from firm data.

Salesforce is the same source system I ingest and model on the data platform side of my job. Prep Me is what that work looks like pointed at a single user: the same records, conformed and trustworthy enough that a language model can compose from them without inventing the parts it cannot find.

Data layer

Engineered the DynamoDB layer feeding LLM-powered retrieval, so the model answered from firm records rather than guesswork.

Grounding data

Curated and structured the datasets used to ground and fine-tune model responses — hands-on with what good training data actually looks like.

Portfolio analytics

Built views surfacing each SFM's book of business and revenue contribution, turning raw records into decision-ready context inside the tool.

Took the initiative

I scoped it directly with the SFMs who would use it, made calls on incomplete information, and carried it far enough that people now rely on it day to day.

How it works — walk through it yourself

Sign in, pick an office, walk in prepared

An SFM signs in and their initials resolve automatically — no lookup, no picking themselves from a list. From there they choose a buying unit and the office they are meeting, and Prep Me returns every meeting held with that client and what was discussed. A new relationship returns nothing, which is the honest answer rather than a fabricated history. The follow-up is drafted, the bio is attached, and the slot is blocked in Outlook and Teams so nobody else books over it. The bios come from the "tell me about yourself" section of each Salesforce record — content that already existed but nobody could reach at the moment they needed it.

Step 01Signed in

Step 02Buying unit & office

Step 03History, email, invite

Output

Signed in.
Choose a buying unit to continue.

Illustrative rebuild of the flow with invented people, buying units and bio text. No Capital Group data, records, Salesforce IDs or model output appears anywhere on this page.

Research & academic projects

Vision systems that hold up in bad conditions

Graduate work on problems where the clean-dataset version is solved and the real-world version isn't. Both panels are interactive.

Masters project — drag the slider

Number plate recognition in rainfall

Plate recognition degrades badly in rain: glare, streaking, motion blur and partial occlusion break character reads that work fine in clear weather. I built a pipeline that recovers plate identity across frames during rainfall and traces the vehicle through the sequence — using LLM-based models to resolve the ambiguous characters that vision alone got wrong.

Texas · USA

JG 5284

Rainfall intensity0 mm/hr

Vision onlyJG 5284

+ LLM resolveJG 5284

Frames to lock1

Schematic illustration of the failure mode, not live model output. Swap in real frames and measured accuracy once you have them.

Capstone project — press run

Classroom attendance by facial recognition

A system that takes attendance from camera input instead of a roll call — detecting faces in a live classroom, matching them against enrolled students, and logging attendance automatically. The hard part is the room itself: angled faces, students behind students, and uneven lighting from the windows.

0 / 24 seats

Idle — press run

Matched0

Low confidence0

Seat-level schematic. The low-confidence flags are the point: occluded and angled faces are where a classroom system actually fails, and they need a human check rather than a silent guess.

AI & LLM practice

What I've actually shipped with models

Every line here maps to something I built and could walk you through the code of.

Retrieval over enterprise data

Structured the DynamoDB layer behind Prep Me's LLM retrieval so responses compose from firm records. The engineering problem is upstream of the model: what you retrieve determines what it can say.

Dataset curation for grounding

Curated and structured the datasets used to ground and fine-tune Prep Me's responses. Deciding what counts as a clean example is the same judgment as writing a data quality test, applied one layer up.

LLM-assisted correction in vision

In the rainfall plate project, language models resolved character reads that the vision model got wrong — using format priors to disambiguate confusable characters rather than guessing.

dq-mcp — an MCP server

Built and published a Model Context Protocol server that gives a language model real data-quality tools instead of guesses: schema inference, profiling, and not-null, uniqueness, referential-integrity and accepted-value checks, named after the dbt tests they mirror. Source on GitHub →

Data quality for AI workloads

dbt test frameworks, schema standards and lineage across pipelines feeding analytics and model workloads. An ungoverned pipeline into a model produces confident, traceable-to-nothing answers.

Philosophy

How I work

I believe meaningful work begins with curiosity, integrity and a genuine desire to create value. Every challenge is an opportunity to learn, grow and make a lasting impact. I approach every project with authenticity, collaboration and continuous improvement — building solutions that are both innovative and purposeful.

CURIOSITY

Ship before the playbook exists

On Prep Me there was no established pattern to copy, so I scoped features directly with the people using them and made calls on incomplete information.

INTEGRITY

Data quality is a product decision

Tests, schema standards and lineage aren't chores you get to after the feature ships. A model grounded on unvalidated records is confidently wrong.

IMPROVEMENT

Fix the gap you noticed

Nobody asked me to move the team onto dbt. I could see what our fragmented transformation logic was costing us, so I evaluated the options and led the migration.

Experience

Where the pipelines run

Asset management, payments and healthcare — the environments where a broken load is not a rounding error.

Apr 2025 — PresentCalifornia, USA

Capital Group

Senior Data Engineer

Ingest from third-party REST APIs, Salesforce, and vendor flat-file drops — CSV, XML, JSON and Parquet — landing everything raw in S3 with the source payload preserved, so a bad load can be replayed rather than reconstructed.
Promote raw files into standardised, queryable tables through a Medallion architecture: schema enforcement and type coercion on the way in, deduplication and conformed keys on the way up, incremental loads with full error handling throughout.
Design and run scalable AWS pipelines (Glue, Lambda, S3, Redshift, Athena) feeding analytics and AI workloads, orchestrated with Airflow.
Model the standardised layer with dbt on Databricks, and enforce data quality as a first-class product: test frameworks covering not-null, uniqueness, referential integrity and business rules, plus schema standards and lineage documentation.
Govern the platform through Databricks Unity Catalog — one catalog for access control, lineage and discovery across every layer, so who can read what is a property of the data rather than a property of the cluster.
Saw what the team's fragmented transformation logic was costing us, evaluated the options, and led the migration to dbt. Now maintain the standards the rest of the team builds against.
Ship through Harness for dev, SQA and production infrastructure, with Bamboo driving on-prem ETL jobs. Own monthly release cycles end to end, including the change and task records that go with them.

Apr 2024 — Apr 2025Chicago, USA

Northern Trust

Data Engineer

Built layered ELT architectures on Azure Databricks and AWS (Delta Lake, Redshift) over large-scale financial transaction data, with dimensional models tuned for Redshift and Snowflake workloads.
Modelled connected financial data in Neo4j and queried it with Cypher, where the questions were about relationships between entities rather than rows in a table — the shape of problem a join-heavy relational model handles badly.
Exposed that data through a federated GraphQL layer, building subgraphs that composed into a supergraph so consuming teams could query across domains without knowing which service owned what.
Owned governance and lineage through Unity Catalog; enforced schema validation and anomaly detection across every pipeline layer.
Ran Git-based workflows with code review and SQL quality gates (sqlfluff).

Oct 2023 — Apr 2024Phoenix, AZ, USA

American Express

Data Engineer

Designed and built ETL pipelines on Google Cloud, landing raw data in GCS and processing it through Spark jobs into BigQuery.
Handled streaming ingestion with Pub/Sub, so event-driven sources and scheduled batch loads fed the same warehouse rather than diverging into two versions of the truth.
Orchestrated the whole flow with Cloud Composer, loading into the target system and feeding the downstream consumers that depended on it.

Jun 2019 — Dec 2021Bangalore, India

Virtusa Consulting Services

Data Engineer

Delivered healthcare data platforms on AWS and GCP (Glue, BigQuery, Dataflow, Dataproc) with compliance-grade validation, audit trails, and lineage.
Orchestrated workloads via Airflow and Step Functions.

Education

Schools I've learnt from

Where the computer vision work and the engineering fundamentals came from.

Graduated 2023Texas, USA

Texas A&M University Corpus Christi

MS, Computer Science

Graduate work centred on applied computer vision: number plate recognition under rainfall, and a capstone system for classroom attendance by facial recognition.

Graduated 2019Hyderabad, India

Osmania University

Bachelor's, Information Technology

Undergraduate engineering foundation. Joined Virtusa as a data engineer straight out of the programme, working on healthcare data platforms within months of graduating.

Stack

What I reach for

Grouped the way I actually think about a system: the model layer, the platform under it, the data that moves through, and the wiring that keeps it repeatable.

ai / llm

RAG over enterprise dataPrompt & retrieval designDataset curation for groundingDaily AI-assisted development

cloud

AWS GlueLambdaS3RedshiftAthenaDynamoDBStep FunctionsEMRGCPBigQueryGCSPub/SubCloud ComposerAzure Databricks

data

PythonSQLPySparkSparkdbtDagsterAirflowKafkaSnowflakeNeo4jCypherGraphQL federationUnity CatalogDimensional modelingMedallion architectureParquet / CSV / XML / JSON

infra

TerraformKubernetesDockerHarnessBambooGitHub ActionsAzure DevOpsCI/CDRelease management

2023AWS Certified Cloud Practitioner

2024Neo4j Certified Professional

2023CS50 Python — Harvard

Support the work

Fuel the pipeline

Everything on this site was built on caffeine and stubbornness, roughly in that order.

Cups today: 0
click the cup

Send a coffee, or just send me something funny

If something here was useful — or you'd like to argue about whether dbt tests belong in CI — a Starbucks or Black Rock gift card is never turned away. Bad data jokes are accepted as an alternative currency.

Cup zero. The pipelines are green and I have no reason to believe otherwise.

☕ Starbucks eGift Black Rock eGift Send a bad data joke

Both gift flows will ask you for a recipient email — that's jyoshnagoshika@gmail.com. The joke costs nothing and, honestly, clears faster.

Contact

Got an ambiguous problem?

I'm comfortable owning one end to end without a playbook. Happy to talk about data platforms, retrieval systems, or anything that needs to survive contact with production.

jyoshnagoshika@gmail.com GitHub LinkedIn Instagram