BI Strategist & Systems Developer
I build systems that make data answer back.
Inventory, finance and forecasting software for healthcare, a finance platform for a six-division manufacturing group, and the dashboards that sit on top. Designed, modelled and shipped end to end.
See the workSelected work
Index
Each project keeps its own color. Scroll through and the page takes it on.
Regnum Pecuniae
Clinical operations software for clinics and pharmacies. Inventory, finance, people and forecasting in one multi-tenant platform, built so each clinic sees only its own data and every forecast shows its working.
- Problem
- Stock, expiry, spend and demand sit in separate spreadsheets. Nobody has time to reconcile them, so expiries and stock-outs are found too late.
- Stack
- React 19 and Chart.js, Flask on gunicorn, Azure SQL, Redis for sessions, a PostgreSQL standby mirror. Deployed on Railway with push-to-deploy, and also packaged as a Windows desktop app.
- Forecasting
- Two ensembles. One blends eight base forecasters with ant colony optimisation. The other only admits tree and deep-learning models once there is enough history for them to add signal, and weights every model by its rolling accuracy.
- Isolation
- SQL Server row-level security keyed on the clinic. A connection with no clinic stamped sees nothing, so failures fail closed. The app's own database user cannot switch the policies off.
- Security
- JWT sessions with refresh, TOTP multi-factor, single-use hashed invitation links, role capping across six tiers, idempotent writes and a full audit log.
- Status
- Running in a pilot facility. Results will be published once measured, not before.
Swarm ensemble · how the colony blends forecasts
Illustrative seriesSix base models forecast the last six months they were not trained on. Ants favour the models that score well, pheromone builds on them, and the final forecast is the pheromone-weighted blend.
Performance work · measured on the live deployment
Response time in seconds · from a client in Botswana
- Finding
- The queries were not slow. The app was 165 ms from its database and paid that cost many times per request.
- Fixes
- Connection pooling, schema checks moved from every request to deploy time, and a live-poll bug fixed so one idle tab went from hundreds of requests a minute to two.
- Result
- 30 endpoints dropped from 50.9 s to 39.0 s in total, and the intermittent "site can't be reached" errors stopped.
- Testing
- 63 endpoint checks plus 7 regression guards, 16 least-privilege checks, 17 invitation and escalation checks, 43 demo-data checks and 42 frontend tests.
01Inventory and expiry
Batches, expiry dates, reorder levels, movement history and refill suggestions.
02Finance
Invoices, budgets, cash-flow forecast, payroll, fixed assets with depreciation, and accounting exports.
03AI insights
Demand forecasts, anomaly detection, expiry predictions and smart alerts.
04Ask your records
Plain-language questions answered from the clinic's own data.
05Data upload
CSV and Excel import with auto-mapping and row-level error reports.
06Guided demo
A seeded demo clinic, rebuilt daily, with a tour that walks the real pages.
Signature Ateliers
An event planning platform for a planner in Dar es Salaam. Clients build their package on the public site; the team runs the whole event from a studio behind it, from the first quote to the last RSVP.
- Client
- Corporate and social event planner in Dar es Salaam, with rentals to follow. "From sassy to classy."
- Public site
- Pick an occasion, see what comes as standard, add anything from the full catalogue, and send it with the enquiry. Prices never reach the browser: they are left out at the database query, not just hidden.
- Studio
- Behind each enquiry, an event workspace: quotes, guest lists, tokenised RSVPs, invitations, a contribution (michango) ledger, vendor bookings, rental stock, and a floor planner that already stores its layout in 3D.
- Rules
- The catalogue is data, so a new event type is a seed change with no code edits. Money is whole shillings, never floats. Sent quotes snapshot their prices.
- Stack
- Next.js 16, React 19, Prisma 7 and PostgreSQL, Tailwind 4, Zod, React Three Fiber, Vitest. Deployed on Railway.
The spine · one guest count drives everything
Illustrative eventWhen a guest accepts their RSVP, the confirmed headcount moves. Per-person quote lines, table count, print run and vendor quantities all follow from that one number.
Finapp
A personal money app for people whose money sits in a bank, a mobile money wallet and cash, in more than one currency. It tracks, budgets and advises. It can never move money.
- Principle
- Read-only by design. No bank credentials, no bank APIs, and an allowlist that blocks the server from calling any host it was not built to call.
- Stack
- A pnpm monorepo. Expo and Expo Router for native iOS and an installable web app; Hono, Drizzle and PostgreSQL for the API; one shared package of Zod schemas. Deployed on Railway.
- Planning
- Payday allocation templates, category budgets, and subscriptions that reserve their money when due and wait for you to confirm the real charge, then roll forward by calendar month.
- Local fit
- Categories adapted for Botswana and Tanzania, down to mobile money fees, airtime and medical aid. Account pickers know which currency each provider uses in each country.
- Design
- Liquid Glass navigation on iOS. Account cards use the real ID-1 card ratio, and their text colour is computed from the card's luminance at the WCAG crossover, never picked by eye.
Payday plan · split it before you spend it
IllustrativeAccount cards · ink computed from luminance
Financial Hub
The financial intelligence platform for a Tanzanian manufacturing group, built as its junior systems developer. Costing, pricing, break-even, production planning and cash flow for six divisions, from textiles to a gypsum mine.
- Client
- Manufacturing group in Tanzania: production, retail, mining, fibre, teak and logistics. Name withheld.
- Evolution
- Started as a Streamlit prototype on spreadsheets, moved to PostgreSQL, then rebuilt as a Next.js and FastAPI platform with full feature parity.
- Stack
- Next.js 14, TypeScript, Tailwind, TanStack Query. FastAPI, PostgreSQL 16, Alembic. JWT with bcrypt, Docker, Railway. Claude for narratives and Ask AI.
- Costing
- Product cost is computed from the bill of materials, never typed in. Contribution margin, break-even and a weighted mix break-even follow from it for each price tier.
- Governance
- Prices and BOMs are versioned, never overwritten. Items are soft-deleted, every change is audited, and BOM categories can be locked with their own passwords.
- Import
- Excel uploads are matched to fields by fuzzy rules, with GPT-4o reviewing low-confidence columns. Every batch can be undone. The mine's daily site workbooks are parsed by section, so layout shifts do not break them.
- Ask AI
- Staff ask questions in plain language. A rule parser answers what it can and Claude handles the rest, but every query passes a guard that allows only read-only SELECTs on approved tables. Questions can be scheduled to arrive hourly, daily or weekly.
- Monitoring
- Six loops check mining anomalies, plan against actual, cash-flow reconciliation, missing data and unusual expenses, then notify the right department. They never change business data.
- Forecasting
- A lightweight bake-off fits each method on an 80/20 holdout, scores it by error, and refits the winner on the full series. It runs in the request, with no heavy ML dependencies.
Smart reallocation · fit the plan into 26 days
Illustrative productsReport Engine
Board decks that build themselves. On the first of every month the Hub assembles last month's figures into a formal PowerPoint deck, writes the commentary, and emails it to the directors and CFO.
- Decks
- A company month-end summary, a mining report under the mine's own brand, a production report, and a department budget-versus-actual pack.
- How
- A small slide library in python-pptx (title, KPI, table, native chart, narrative) with a theme per department. The first decks were hand-tuned in pptxgenjs, and their design became the library's.
- Commentary
- Claude writes the narrative from the computed figures only. With no AI key it falls back to a plain factual sentence, so the deck always ships.
- Data first
- Figures come from the same repository code as the app, with guards against double-counting, so the deck and the dashboard always agree. Source inconsistencies are flagged, never smoothed over.
One slide library, a theme per department
Sample slides · switching department swaps tokens, not layouts
In-store Feedback
A QR code review flow for a retail chain. Four questions at the till, a follow-up when the customer comes back, and an admin dashboard that turns the answers into retention and complaint tracking.
- Captures
- How the customer found the store, price perception, main reason to buy, and a staff rating with a follow-up for low scores. One first review per phone number.
- Stack
- Next.js 16, React 19, Supabase Postgres, Zod validation, Excel exports with exceljs, deployed on Railway.
- Security
- Row-level security lets the public only insert, never read. Admin sign-up is limited to the company's email domain by a database trigger, not just the form.
- Dashboard
- Retention and churn from return visits, retention by discovery channel, sentiment split, at-risk customers, and complaint reports. Three stars or fewer counts as a complaint.
Staff rating distribution
Sample responses
How I work
Rules first.
AI where it pays.
Most operational problems are solved by clean data and clear rules. Models come in where the rules run out, and they always show their reasoning.
Cheap, auditable, always on
Reorder points, expiry windows, margin floors and alert thresholds. They cost nothing to run and anyone can check them.
- Data models designed for the decision, not the report
- Validation at import, so bad rows never reach a dashboard
- Role-based access with a full audit trail
Forecasts that explain themselves
Demand forecasting, anomaly detection and plain-language queries, used where patterns are too many for fixed rules.
- Every suggestion shows the records behind it
- Measured against a simple baseline before it ships
- Runs locally when data must stay on site
Contact
Your data already knows. Let's ask it.
Open to remote BI, analytics and systems roles, and to focused builds for businesses that run on spreadsheets today.