Kanzu Agent

Inspiration Across Uganda and much of East Africa, the businesses that hold the economy together — micro-SMEs, savings groups, and SACCOs — are also the ones most locked out of modern financial tooling. They don't have reliable internet, they can't afford cloud subscriptions billed in dollars, and yet they're increasingly expected to keep clean books, meet basic AML/KYC hygiene, and produce records regulators and lenders trust.

When I saw the Africa Deep Tech Challenge's "Laptop LLM Challenge" brief — build a real, working AI system that runs entirely offline on the exact laptop already sitting in millions of homes and shops across the continent, an 8 GB, integrated-graphics machine costing $150–$500 — it clicked with work I was already doing on AML and fintech compliance systems. I wanted to take that same compliance discipline and shrink it down to something that runs, for free, forever, on hardware a shopkeeper or SACCO treasurer already owns, with zero cloud dependency and zero recurring cost. That became Kanzu Agent: a compliance-aware business copilot that lives entirely on-device.

What it does Kanzu Agent is an autonomous, offline AI agent that acts as a business copilot for micro-SMEs and savings groups. It runs completely locally — no API calls, no internet requirement, no data ever leaving the machine — and helps small business owners and SACCO officers with:

Day-to-day bookkeeping and transaction recording in plain language

Compliance-aware task automation — flagging transactions or member activity that look like they need AML/KYC attention before they become a real problem

Local orchestration of recurring business workflows (reminders, reconciliation, reporting) without any human needing to babysit the process

Privacy-first operation — sensitive financial and member data for a SACCO never touches a third-party server

(Fill in the concrete feature list you actually shipped for Gate 1 — e.g. which workflows are automated end-to-end, what the UI/CLI looks like, and whether African-language support made it into this build, since that earns a 15% bonus on the panel score.)

How I built it The core of Kanzu Agent is a quantized Qwen2.5-1.5B-Instruct model in GGUF format, chosen specifically because it's small enough to run comfortably inside the ADTC Standard Laptop's 7 GB memory ceiling while still being capable enough to hold structured business dialogue and reason about basic regulatory/compliance tasks.

Around that model, I built a local orchestration layer that:

Handles all inference on-device with no network calls

Structures free-form conversation into the AML/bookkeeping domain the agent is meant to operate in

Keeps every piece of business and member data local to the machine, matching the zero-cloud-dependency requirement of the challenge

I leaned on experience from compliance and financial-systems work I'd already been doing (my AML platform work in particular) to shape how the agent reasons about suspicious patterns and record-keeping, then adapted that logic to run inside a tightly constrained, fully offline environment instead of a distributed cloud backend.

(This is the section to tighten up with your real specifics: the inference runtime you used to run the GGUF model — e.g. llama.cpp — the orchestration language, how data is stored locally, and any RAG or retrieval component for SACCO-specific records.)

Challenges I ran into Fitting inside the memory ceiling. The 7 GB hard limit (exceeding it is an automatic disqualification, S_total = 0) meant every architectural decision had to account for the model's footprint plus the orchestration layer, with no discrete GPU to lean on.

Balancing capability with size. A 1.5B-parameter model is small enough to run on integrated graphics but light enough that keeping it reliably useful for compliance-sensitive tasks — where a wrong answer has real financial consequences — took careful prompt and workflow design rather than throwing more model at the problem.

Working solo. I built, tested, and iterated on the whole system — model selection, orchestration logic, and the compliance/business-domain framing — by myself, without a team to split the model work from the domain work.

(Add any specific technical blockers you hit — e.g. latency on the target CPU, quantization artifacts, or getting the model to stay reliably "on-domain" for AML/bookkeeping tasks.)

Accomplishments that I'm proud of Getting a genuinely useful, compliance-aware AI agent running end-to-end, entirely offline, on hardware that costs as little as $150 refurbished — with no cloud bill, no API key, and no internet dependency — while addressing a real gap I've seen directly in how African micro-SMEs and SACCOs currently handle bookkeeping and compliance.

What I learned Building for an 8 GB, GPU-less laptop forces a completely different design mindset than the distributed, cloud-native systems I usually build. Every architectural choice has to be justified against a hard memory budget, which sharpened how I think about trade-offs between model capability and real-world deployability — especially for the offline, privacy-constrained environments that a lot of African SMEs and SACCOs actually operate in.

Built With

• aml-kyc
• autonomous-agents
• fintech
• offline-ai
• offline-data-storage-(if-used)-ubuntu-?-reference-os-target-on-device-ai
• qwen2.5-?-the-on-device-model-(qwen2.5-1.5b-instruct-gguf)-[memory]-gguf-/-llama.cpp-?-quantized-on-device-inference-(confirm-which-runtime-you-used)-python-and/or-go-?-orchestration-layer-sqlite-?-local