FBS Technologies · Research & Engineering
Hard problems, chosen deliberately
The conditions global vendors treat as edge cases — intermittent networks, code-switched language, cash-adjacent payments — are our main cases. This is where the group does its original work.
- research aimed at shipping product
- Applied
- research aimed at shipping product
- models served inside tenant boundaries
- In-region
- models served inside tenant boundaries
- Swahili and English by default
- Bilingual
- Swahili and English by default
Research areas
Lines of work, each with a product destination
R-01
Swahili-English NLP
Language models tuned on code-switched East African usage, evaluated on customer-service outcomes rather than benchmarks.
R-02
Offline-first systems
Conflict resolution, receipt integrity, and fiscal compliance for clients that may be disconnected for days.
R-03
Document intelligence
Extraction models for invoices, IDs, and handwritten forms as they actually arrive — photographed, skewed, incomplete.
R-04
Payments risk
Fraud and failure-pattern models across mobile money rails, trained on reconciliation outcomes.
R-05
Low-bandwidth UX
Interface patterns that stay responsive on 2G fallbacks and entry-level Android devices.
R-06
In-region AI serving
Model deployment inside tenant and national boundaries, so intelligence doesn't require exporting data.
Where it lands
The products are the delivery vehicle
Research that survives contact with production ships inside the products themselves: assistants drafting customer replies in Swahili and English inside FlowCRM, document models reading invoices into FlowERP, forecasting inside FlowInsight. Customers never install “AI” — they get shorter queues and cleaner books.
Bilingual assistants in production
Models served at the edge, offline-tolerant
Tenant-bounded, in-region deployment
How we build
Three rules keep the lab honest
01
Ship inside products
Research earns its keep by landing in features that customers use, not in demos. Every research line has a product owner.
02
Measure on outcomes
A model is good when claim rejections fall or response times drop. We measure against operational outcomes, not benchmark scores.
03
Constraints are the brief
Bandwidth, devices, language, and regulation aren't obstacles to route around. They're the design brief that makes the work matter.
Research partnerships
Working on the same problems?
We are open to collaboration with universities and research groups on Swahili NLP, payments, and offline systems.