Proof, not promises

Plenty of people know that showing up on LinkedIn works, but the blank page and the time it takes stop them from posting. Suflate set out to remove that friction: speak or type a rough thought and get back a finished post that actually sounds like you. It needed to be a complete product, not a demo, spanning capture, AI writing, design, scheduling, and publishing.

☞ Numbers, not adjectives.

We built Suflate end to end as a full-stack AI product. That meant a capture pipeline for voice and text, an AI writing engine tuned to learn each user's tone and vocabulary (their 'content DNA'), a library of 20+ writing styles plus custom styles, one-click carousel generation, a month-at-a-time posting planner, multi-workspace collaboration with roles, scheduling, and direct publishing to LinkedIn. Content and recordings are encrypted and never used to train models. Guardrails and prompt design keep outputs on-voice and reliable rather than generic.

“I capture ideas the moment they hit, on a walk, between calls, and by the time I sit down the post is ready. It killed the blank page for me completely.”

— Collins, Founder at Basua Pty Ltd

A Tier-1 South African bank ran a critical regression cycle by hand before every release. It took three people a full week clicking through the same flows, which meant releases were slow, expensive, and stressful, and defects still slipped through because manual coverage was inconsistent.

☞ Numbers, not adjectives.

We audited the existing test coverage and release process to find the highest-value, most-repeated flows. We designed an automation framework fit for the bank's stack, then built an end-to-end regression suite covering those critical journeys and wired it into CI/CD so it runs automatically. Flaky tests were triaged and hardened so the suite stayed trustworthy. Finally, we handed over documentation and trained the team to own and extend it.

“The regression run went from something we dreaded to something we barely think about.”

— QA Lead (details withheld for confidentiality)

Gray Mount Projects ran their project intake and tracking the way most project-led businesses do: out of an inbox and a set of spreadsheets. Work arrived by email, someone read it, logged it, chased it, and recorded time by hand. It worked, but it was manual, easy to drop, and gave no single view of what was in flight. They also wanted a brand identity that matched the quality of their work.

☞ Numbers, not adjectives.

We delivered a full rebrand for Gray Mount Projects, then automated the operational core. We built an email-driven workflow where inbound mail is ingested, classified, and turned into structured project items that move through a pipeline automatically, so a new email becomes a tracked project step rather than a message someone has to action by hand. The same workflow captures timesheet data from the work as it happens, removing manual time entry. The principle is simple: let the inbox drive the system, with structured data and an audit trail falling out of it automatically, while human review stays in the loop where judgement matters. We also rebuilt their website on modern, well-structured, fast foundations, which is what makes it legible to AI systems as well as to people.

Payment platforms in banking are unforgiving. Card processing, cross-border settlement, and SWIFT statement messaging all have to be exactly right, and the combinations of currencies, fees, and failure modes are enormous. Validating them by hand couldn't scale or keep pace with delivery.

☞ Numbers, not adjectives.

We designed a scalable automation framework in Java with Selenium, Cucumber, TestNG, and RestAssured, and built over 1,000 automated test cases across web and mobile banking journeys. We automated validation of SWIFT messages (MT940 and MT942) for compliance and statement accuracy, and covered card flows, multi-currency settlement, exchange-rate and fee logic, and negative and edge cases. The suite was integrated into Jenkins CI/CD so every build was gated, and test data was managed cleanly across SIT, UAT, and NFT environments.

Payment reconciliation across multiple currencies is unforgiving. Small discrepancies compound into real financial and regulatory risk. The reconciliation workflows were complex, and testing them manually left gaps that only surfaced once they reached production.

☞ Numbers, not adjectives.

We built automated API-level and end-to-end coverage for the reconciliation flows, focusing on the currency and edge-case combinations most likely to break. The suite was integrated as a release gate in CI/CD so no build could ship without passing reconciliation checks. Reporting was made readable so failures pointed straight to the cause.

The Nedbank Machine Learning & Data Science Challenge 2026 asked entrants to predict how many transactions each customer would make over the next three months, the Nov 2015 to Jan 2016 holiday window, from historical transactions, demographics, and financial data. It matters commercially because transaction forecasts drive resource planning, cash management, and customer engagement. It is hard because the target is heavily skewed (most customers transact rarely, a few hundreds of times) and the prediction window is exactly when behaviour diverges from the norm. The competition is judged purely on measurable model performance, which is the evidence-led standard Nullius is built on.

☞ Numbers, not adjectives.

We built a reproducible end-to-end pipeline. A temporal panel was constructed over nine historical cutoff dates so the model learned from many points in time, with strict date filtering to prevent any future-information leakage. From that we engineered 266 features: temporal lags, rolling statistics, year-on-year seasonal comparisons, trend slopes, demographics, financials, and calendar seasonality. The breakthrough was leave-one-cutoff-out target encoding on high-cardinality categories (such as residential city, 550 levels) with smoothing to prevent overfitting. We trained a diverse ensemble, LightGBM, XGBoost, and CatBoost plus a Tweedie-objective model for diversity, each bagged across seeds, tuned with Optuna. The final model used segmented blending: customers were split by recent activity (low, mid, high) and per-segment blend weights were optimised with SLSQP against a held-out month, because sparse and heavy users need different model mixes. Everything runs from one command in about 30 minutes using Polars lazy evaluation.

“Judged on numbers against 251 teams, the model finished third. That is the only kind of proof that counts.”

— Nullius Technologies

Getting a small business registered and funding-ready is slow, confusing, and full of paperwork. A startup wanted to remove that friction with software that could guide founders end to end, and prove the idea worked with real users.

☞ Numbers, not adjectives.

We co-led development of an AI-powered platform that walks users through CIPC business registration, dynamically generates investor-ready business plans, and recommends tailored funding products. We integrated the OpenAI API for business-plan generation and Cloudinary for file storage, built secure token-based authentication with role-based access and encrypted credentials, and shipped a fast, responsive front end across a full-stack build.