Data Engineer — Phundit
Location: Accra, Ghana (Remote, GMT ± 0)
Data Engineer — Phundit
About Phundit
Phundit is a personal finance platform helping Ghanaians save, invest, and access credit. We're building a secured credit product backed by customers' own savings, fully deterministic, regulator-defensible, and designed to evolve from rules into fitted models as performance data accumulates.
The role
You're the first dedicated data hire. You'll build the data foundation that powers credit decisioning, produces regulatory deliverables for our partner bank, and creates the labelled datasets a future ML engineer will train against.
The event data exists. The pipeline from source to warehouse exists. What doesn't exist yet is the layer that makes it usable for credit: point-in-time reconstruction, feature computation with observation-point discipline, outcome labelling, and reproducible exports.
What you'll do
Immediately
- Design and implement a feature store supporting point-in-time queries across 26 defined characteristics
- Build transformation pipelines that produce clean, consolidated event streams with anti-gaming controls applied once at source
- Implement the daily position file and reconciliation statement required by our partner bank
- Set up data quality monitoring: freshness, completeness, reconciliation breaks, distribution drift
Within 6 months
- Implement outcome labelling with correctly defined performance windows and pre-netting status preservation
- Produce the first development sample: one row per matured credit cycle, features frozen at observation point, labelled after window close
- Build monitoring baselines: score distributions, gate failure rates, Population Stability Index
- Run leakage audits ensuring no post-decision information contaminates feature values
Within 12 months
- Compute univariate statistics (Information Value, Weight of Evidence) for each characteristic
- Build back-testing infrastructure: re-run decisions against historical state, compare, flag divergence
- Produce vintage performance tables and roll-rate matrices
- Prepare the handoff dataset for an ML engineer joining the team
What we're looking for
Required
- Strong SQL — window functions, precise numeric types, partitioned table design, point-in-time self-joins
- Python — transformation logic, data validation, scripting
- Google Cloud Platform — warehouse, storage, compute, scheduling (we're GCP-native, no multi-cloud)
- Power BI — building monitoring dashboards, regulatory reporting packs, and portfolio visualisations
- Understanding of event-sourced data — the difference between "current state" and "state as at a point in time" and why credit decisioning needs the latter
- Precision arithmetic — you know why floating point can't represent 0.55 exactly and why that matters at a score boundary
Strongly preferred
- Financial services or credit risk data — loan tapes, vintage analysis, WoE/IV, regulatory reporting
- Experience building ML-ready datasets — entity-level splits, observation/performance windows, leakage prevention
- Experience as an early data hire — comfortable with ambiguity, ownership, and building from scratch
- NoSQL-to-warehouse pipeline experience — denormalising document stores into analytical schemas
Nice to have
- dbt or Dataform
- Familiarity with Ghanaian financial regulation (Act 843, BoG reporting)
- TypeScript reading ability (our backend is TypeScript — you won't write it but you'll need to understand what events are emitted)
What you'll get
- Competitive compensation. You will be responsible for laying the foundations for Phundit ML credit models. And this will reflect in your salary
- Full ownership of the data layer from day one
- A precisely written specification (not a vague brief) defining exactly what the data must support
- Direct access to the CTO and CEO — this is a small team, not a hierarchy
- A GCP environment that's already running, not a greenfield cloud setup
- Power BI for all dashboarding and visualisation work, connected directly to the warehouse
- The opportunity to shape a credit system from its first labelled outcome through to its first fitted model