Saahil MehtaAssociate Director, Data & Insights

Showreel

Two systems I built at South Pole: Orca, the company's AI assistant, and the governed data platform every answer runs on.

Figures as of 26 Oct 2026.

The film. Every figure in it is on this page.Download

Orca

South Pole's enterprise AI assistant. I designed and built it, and it opened to the whole company in June 2026.

The product

A reconstruction of the Orca interface. The accounts and figures are invented.

How it works

  1. ContextData, documents, mail and owners
  2. ReasoningPlans, queries and research
  3. GuardrailsRead-only checks, run as you, logged
  4. OutputAnswers, artefacts and drafts in your tools
  5. LearningMemory, feedback and certified meaning
Learning feeds back into context. Orca never writes to the warehouse and never uses a shared identity for data.

What it does

Ask
Live answers from the marts as you, with charts and tables.
Research
Cited deep research in the background, up to 40 minutes.
Schedules
Any prompt daily, weekly or monthly, emailed when done.
Artefacts
Documents, sheets and decks, exported to Google Workspace.
Dashboards
Account, market and buyer boards, plus a personal board.
Memory
Shared chats and artefacts, personal memory and instructions.

Adoption

Since the first release on 11 June, 335 people have signed up and 272 have asked a question. 188 used Orca in the last 30 days.

It has run more than 15,000 governed queries as the person asking and produced over 1,600 artefacts, nearly 250 of them exported to Docs, Sheets and Slides.

A release ships about once a month, each gated by an answer-quality evaluation. Release 2.0 carried more than 150 improvements.

Running cost and footprint

The data platform

The governed data platform every team and Orca read from. My team and I rebuilt it from 2023 to 2026 while it stayed in daily use.

2023 to 2026

  1. 2023InheritedInfrastructure changed by hand, half of pipeline changes skipped review, and nothing tracked whether runs succeeded.
  2. 2024StabilisedWe wrote the standards for marts, pull requests and review, and repaired the failing runs.
  3. 2025CodifiedThe core of the estate moved into Terraform in about four months, secrets were encrypted in code, and review became the default.
  4. 2026GovernedSensitive columns are masked at source, Dagster runs the marts, and the Orca family runs on the platform.
Platform measures by year, 2023, 2024, 2025, 2026
Measure2023202420252026
dbt runs that succeededNot tracked47.7%99.6%99.8%
Changes reviewed before merging48%85%Over 99%Required
Data tests in the marts166243260389
Lines of code we run59,90065,80076,500251,000
Terraform lines for the infrastructureBy handBy hand8,00011,600

Counts at year end; 2026 to 19 October. From the git history of the platform repositories, dbt run artefacts and the orchestration run history.

Scale

Savings

  1. Banked, September 2026

    dbt Cloud

    Dagster on GKE

    Renewal quoted at over 50k EUR a year on a three-year lock-in

  2. Banked

    Power BI Pro

    Orca Visuals on Lightdash

    About CHF 25,000 a year for 220 seats

  3. Avoided

    A paid data catalogue

    Orca Semantics on OpenMetadata

    40k to 60k EUR a year

  4. Avoided

    Paid observability tooling

    GCP-native logging and monitoring

    About 11k USD a year

  5. Avoided

    BI copilot and automation tools

    Orca

    About 20k EUR a year

About 150k EUR a year banked or avoided. GCP logging and monitoring cost about 710 EUR over the last twelve months.

How it fits together

The systems in production, what each one does and what flows between them.

How it fits togetherSource systems lands in Raw layer; Raw layer transformed by Data Marts Core; Data Marts Core builds Data marts; Data marts defined in Orca Semantics; Orca Semantics gives meaning to Orca; Data marts queried as the signed-in person Orca; Orca Semantics certified metrics for Orca Visuals; Orca answers The company; Orca Visuals dashboards for The company; Data Foundation Plan sets the roadmap for everything below Data Marts Core.SOURCESINGEST AND LANDTRANSFORM AND OWNSERVE AND ANALYSEPEOPLELANDS INTRANSFORMED BYBUILDSDEFINED INGIVES MEANING TOQUERIED AS THE SIGNED-IN PERSONCERTIFIED METRICS FORANSWERSDASHBOARDS FORSETS THE ROADMAP FOR EVERYTHING BELOWSource systemsoperational systemsRaw layerlanded by pipelinesData Marts Coredbt on DagsterData martstables in BigQueryOrca Semanticsdefinitions and ownersOrcathe AI assistantThe companyeveryone at South PoleOrca Visualsdashboards as codeData Foundation Planthe strategy the executive committee adoptedUNDER EVERY LAYEREnterprise data platformone Terraform estate for the companyData Access Governancegrants with a reason and an audit trailCost Deep Divespend and anomalies on every projectData quality at scaletests and alerts routed to ownersService levelsfreshness and reliability targets

Data Foundation Plan

the strategy the executive committee adopted

sets the roadmap for everything below

  1. Sources

    • Source systems

      operational systems

      lands in Raw layer

  2. Ingest and land

    • Raw layer

      landed by pipelines

      transformed by Data Marts Core

  3. Transform and own

    • Data Marts Core

      dbt on Dagster

      builds Data marts

    • Data marts

      tables in BigQuery

      defined in Orca Semantics

      queried as the signed-in person Orca

    • Orca Semantics

      definitions and owners

      gives meaning to Orca

      certified metrics for Orca Visuals

  4. Serve and analyse

    • Orca

      the AI assistant

      answers The company

    • Orca Visuals

      dashboards as code

      dashboards for The company

  5. People

    • The company

      everyone at South Pole

Under every layer

  • Enterprise data platform

    one Terraform estate for the company

  • Data Access Governance

    grants with a reason and an audit trail

  • Cost Deep Dive

    spend and anomalies on every project

  • Data quality at scale

    tests and alerts routed to owners

  • Service levels

    freshness and reliability targets

Owners are named once, in the semantic layer, and access approvals, quality routing and Orca's context all reuse them.

Let's talk.

About building AI systems people actually use, leading data teams, or your data and cloud estate.

LinkedIn Book a call

Also available asresume.pdfresume.jsoncv.mdllms.txtstatus.json