Reference architectures

We deliver under NDA. Here is what we build.

Annotated diagrams of real systems we have shipped. Client names withheld; architecture and deployment shape preserved.

Named references available on the strategy call.

50K+docs / dayTier-1 East African bank · Doc Intelligence
99.7%extraction accuracy200K-page gold-set eval
8 wkbuild timeInitial production deploy
0bytes egressedAcross 100% of deployments
12NDA referencesAvailable on strategy call

BFSI document intelligence pipeline

On-premise · Tier-1 East African bank

Scanned loan-application forms ingested at branch level, OCR'd, structured-extracted with a fine-tuned VLM, and pushed into the core banking system. Reduces processing time from days to minutes; runs entirely inside the bank's perimeter.


   Branch scanner ──▶ Pre-processing  ──▶  Vision LLM  ──▶  Schema validator  ──▶  Core banking
                            │                  │                                      │
                            └──── Audit log ───┴──────  Eval pipeline (daily)  ───────┘

Telco contact-center stack

Customer cloud · African MNO

Live call audio streamed to STT, semantic search over knowledge base, agent-assist UI with next-best-action and compliance prompts. CRM written back in real time. Deployed in customer's AWS Cape Town VPC.


   Live call ──▶ STT ──▶ Semantic search ──▶ Agent-assist UI  ──▶  CRM write-back
                          │                       │
                          └───── Knowledge base   └──── Compliance prompts

Government on-prem private LLM

Air-gapped · National public-sector agency

Llama 3.x 70B running on on-premise H100 cluster, with vLLM serving and a retrieval-augmented stack over the agency's internal knowledge base. Fully air-gapped: zero internet egress.


   Internal apps  ──▶  Agency auth  ──▶  vLLM (Llama 70B)  ──▶  Internal vector index
                                                  │                       │
                                                  └─── Audit log + eval ──┘
                                                  ✕  no internet egress

Conversational BI on a customer warehouse

Customer cloud · Indian fintech

Natural-language interface that compiles questions into SQL against a customer's Snowflake warehouse. Semantic cache for common queries. Deployed in customer's AWS Mumbai region with DPDP-compliant data handling.


   User question ──▶ NL → SQL  ──▶  Semantic cache  ──▶  Snowflake warehouse
                          │                                   │
                          └──── Eval set + governance ────────┘

The closer

Build the AI you'd be proud to own.

Thirty minutes to talk through your stack, your data, and the AI opportunity you care about most. No pitch deck. No sales theatre.

Book a 30-min strategy call →or email kevin@ubuntuonline.co.ke
Ubuntu Online · Nairobi · 2026
Book a strategy call