Projects

Work, described plainly

What we built, what it cost, what we'd do differently. Client names on request.

Agriculture · DR Congo1 month

Dispatching machines and crews across farm sites

A key player in Democratic Republic of Congo agriculture runs multiple farming sites, and needed one platform to manage them: dispatching machines and workers between farms, scheduling shifts, and tracking machine usage and maintenance. The hard part was the UX and the data flow between four roles — management, mechanics, workers and the dispatch team — and the notifications tying them together. We ran the discovery phase with them to reach a clear PRD, then built it with our in-house Agentic Development Method for autonomous development, deployment and simulated user testing. Fully operational one month from project start.

VercelSupabaseGitHubAnthropic LLMs
4 roles
coordinated in one system
Recruitment · early-stage startups1 week

A two-sided hiring platform, live in a week

A boutique agency hiring for operational roles at early-stage startups needed to match companies against a vetted database of candidates, and needed it running within a week of kickoff. Two-sided signup, a manual verification workflow covering both employers and candidates, personalised communication, and several routes for the customer to start using the product. Verification and matching were the features everything else hung off. We designed it to their brand book, decomposed it into a defined technical architecture and roadmap, and our agent team delivered it autonomously in five days, wired to an SMTP server for the email side.

Next.jsSupabaseVercel
1 week
autonomous delivery, design to live
Construction tech · Poland4 weeks

WycenAI: construction cost estimates from proprietary data

A Polish construction-tech startup building a platform where B2B users generate cost estimates for their investments — broken down to individual materials and services, priced accurately, and referenced against comparable recent investments. Week one was a deep analysis of their proprietary data using NLP, cleaning, classification and feature extraction, which produced concrete recommendations for data quality and a proposed design for the recommendation engine. We defined the technical KPIs that decide whether the product works, and built an evaluation harness to measure every feature we shipped against them. Then we designed and implemented the search and price recommendation engine — statistical analysis, a vector database, specialised vectorizers and LLM applications — generating accurate draft cost estimates for platform users.

NLPVector databaseLLMsEvals harness
4 weeks
data analysis to a live recommendation engine