A three-week build, three technologies, and one rule we refuse to break
By Kiprono Ngetich
The before
Before building Daraja Workspace, our project managers spent an average of 6-8 hours per week on status reporting.
That's not an exaggeration we tracked it.
The breakdown looked like this:
2 hours chasing updates from team members
2 hours reconciling conflicting information
2 hours formatting reports for different stakeholders
2 hours in status meetings discussing what was already outdated
According to the Project Management Institute, organizations waste an average of 11.4% of their investment due to poor project performance — much of which stems from inadequate communication and status visibility.
We were living that statistic.
The insight
One day, I looked at the raw material our teams were already producing: weekly summary messages, task comments, deadline discussions, Slack threads.
Every single one contained structured project intelligence project names, responsible parties, deadlines, blockers, next steps. It was just trapped in unstructured text, invisible to anyone who needed to report on it.
The insight was simple: the best status report is one that writes itself.
We stopped asking "how do we make people update status sheets?" and started asking "how do we extract status from work people already do?"
The build
We built the first version in three weeks, using three core technologies:
Week Focus Technology
1 Real-time collaboration layer WebSockets + optimistic UI updates
2 AI extraction pipeline Cohere Command-A + prompt engineering
3 UI + fallback handling Conditional persistence (POST-once model)
The POST-once model was the key architectural insight: the system calls the AI at most once per week, regardless of how many people view the status sheet. Storage grows by one record per week per project. No redundant API calls. No spiraling costs.
The result
After six weeks of internal use:
Status reporting time: 6-8 hours per week → effectively zero (the sheet maintains itself)
Information freshness: Updated daily via AI extraction + real-time human edits
Team adoption: 100% of active projects migrated
Server cost: One API call per project per week (regardless of view count)
But the numbers don't capture the real change.
The Thursday afternoon panic is gone. Project managers don't chase updates anymore — they read them. They clarify. They unblock. The time they used to spend formatting tables now goes to actual project management.
The rule we refuse to break
Throughout the build, we held one line:
AI writes the first draft. Humans own the truth.
Every AI-generated status row carries a flag — isManuallyEdited — that distinguishes model output from human input. When a project manager corrects a deadline or adds a comment, that change propagates in real time and never gets overwritten.
We don't let the AI learn from corrections in a way that overrides human judgment. The model assists. It does not decide.
What's next
The current system is one-way: AI → humans. Human corrections don't yet flow back to improve the AI's understanding.
We're closing that gap. The roadmap includes:
Two-way feedback loop — AI learns from human edits over time
Export pipeline — one-click PDF and XLSX exports
Cross-week analytics — trend visualization across multiple weeks
Cell audit trails — per-change history with username and timestamp
The broader takeaway
We built this because our Thursdays were broken.
If yours are too, you don't need better templates or stricter deadlines. You need a different architecture. One where status visibility is a side effect of working — not another job.
The information you need is already there. It's in your Slack, your email, your weekly updates, your task comments.
You just need to extract it.
Kiprono Ngetich builds AI-assisted collaboration tools.
Top comments (0)