Experiments

The work is the proof.

u.lab is built around real projects. Every experiment starts with a question.

Experiment library

The panels below are teaching examples written by u.lab to show the shape of an experiment. They are not client work.

Question / Problem: Can an internal operations team automate document classification?
Hypothesis: Fine-tuning an open-source model on regional data will reduce sorting overhead by 40%.
Experiment: Ran an isolated 48-hour pilot testing 5,000 legacy records against a sandboxed server.
Evidence: Sorting accuracy hit 88%, but local data pipeline infrastructure requires optimization before live deployment.
Lesson: A clear understanding of where AI can help, what works and what it would take to implement it.
Next Step: Refine localized pipeline caching mechanisms before scaling test databases.

Question / Problem: Is there actual consumer demand for a niche digital trade service?
Hypothesis: Targeted landing pages with high-intent deposit options will yield a >5% conversion rate.
Experiment: Deployed 3 distinct functional landing pages to segmented audiences over 14 days.
Evidence: Conversion tracking stayed below 1.5%; evidence suggests alternative distribution channels or fundamental product adjustment are required before engineering.
Lesson: Evidence to decide whether to build, change or stop.
Next Step: Initiate structured user interviews to pinpoint workflow drop-offs.

Question / Problem: Can a critical document verification process be shortened from weeks to days?
Hypothesis: Digital indexing of incoming physical receipts will clear the initial bottleneck.
Experiment: Shadowed workflow and introduced a basic digital tracking queue for one provincial intake desk.
Evidence: Processing velocity increased by 300% within the testing group; the approach is structurally viable for scale.
Lesson: A tested improvement that can be measured and potentially scaled.
Next Step: Present structural throughput data to regional technical leadership.

Question / Problem: Can off-grid verification hardware function reliably under erratic connectivity environments?
Hypothesis: An asynchronous local-first database synchronization protocol can prevent transactional dropouts.
Experiment: Maintained 3 field test units in low-signal corridors running continuous simulated requests.
Evidence: Local caching prevented all data loss; network syncing must handle larger packet compression.
Lesson: Evidence about whether the technology is worth pursuing.
Next Step: Finalize data compression configurations for lower bandwidth parameters.

Question / Problem: What is the optimal market entry vector for a new supply-chain service?
Hypothesis: A lightweight WhatsApp-based coordination tool will see higher adoption than a native mobile application.
Experiment: Managed 50 delivery flows manually via an automated chat interface across a 7-day period.
Evidence: Users showed 92% retention on the chat layer; building a native app is deemed an unnecessary initial expense.
Lesson: Better information before a larger investment is made.
Next Step: Codify the automated conversational flow scripts into stable internal systems.

Over time, these experiments become a public record of what we've learned.

Have a problem worth testing?

Tell us what is not working. We will tell you how we would investigate it — and what a first experiment could look like.