Victor Ndunda · Production AI systems engineer
I help founders and product teams turn high-value workflows into reliable AI systems — RAG knowledge tools, agents, automation, forecasting, and fraud detection — with real evaluation, observability, and a clear path from pilot to production.
Every engagement begins with a diagnosis — never with a stack recommendation. Full packages with USD & KES pricing live on the services page.
Find the workflow worth automating — and what to leave alone.
Best for: teams that know a process is expensive but aren't sure AI is the answer.
Ship one reliable system, not a roadmap.
Best for: a defined high-value workflow with an internal owner.
What's breaking, why, and what to fix first.
Best for: an existing AI feature with quality, latency, cost, or trust problems.
Three of six shipped systems — each with the constraint that made it hard and the evidence that it works. All six live here →
A multilingual AI learning product that reached 11,000+ learners across 38 countries.
Africa-built climate-leadership platform. The AI Career Coach, the multilingual NLP pipeline (English · Français · Kiswahili), and personalized learning recommendations had to coach real learners in production.
A sales & ops AI agent that automates follow-ups, CRM logging, and approvals — with its reliability proven in a public 281-defect code review.
Multi-tenant SaaS with real OAuth (Google, HubSpot), encrypted tokens, Celery workers, and Langfuse observability. Every P0/P1 finding was remediated within one week of the review.
Real-time transaction scoring at 84% recall and a 1.37% false-positive rate — ~33 µs per decision.
Seven statistical detectors fused by a weighted-vote ensemble, calibrated by logistic regression, overlaid with a deterministic rules engine — every score explainable and auditable. Validated on 1,000 labelled transactions; methodology published with the code. Designed with controls relevant to PCI-DSS, PSD2, and GDPR automated-decision requirements — formal compliance is deployment-specific.
Pilot-to-production is where most AI dies. This engagement model is designed around getting past it.
Define the user, the workflow, the constraint, and the success metric. No stack choices before this — most "AI projects" fail here, not in the code.
Build the smallest useful slice and test it against real scenarios with a real evaluation set. Numbers or it didn't happen.
Harden, document, observe, and hand over a system your team can run — with the dashboards to prove it stays healthy.
I'm an AI engineer based in Nairobi, Kenya. I started in data science for fintech and financial services — forecasting, fraud detection, customer segmentation — and moved to production AI full-time when generative models matured: multi-agent platforms, RAG pipelines, on-device ML, and multilingual assistants.
What I actually do is take technically difficult AI ideas and make them work under real constraints — $80 phones, Swahili and French delivery, regulated finance, weak connectivity. The six systems below — and the client product above — are the evidence. I also write the engineering down: 10 long-form articles, 22,000+ words, because work you can't explain is work you can't trust.
I take on select client work — audits, sprints, and full builds. If you have a problem that needs real engineering, let's talk.
Currently building: Busara v7 pipeline scaling · KilimoPRO offline v2 · AgentReplay 1.x — write-ups in Insights.
From scoping audit to production deployment. Full-stack: Next.js, FastAPI, Flutter, PostgreSQL, Docker, Kubernetes.
Multi-agent DAG orchestration, RAG with pgvector/Pinecone, hybrid search, reranking. LangChain, LangGraph, CrewAI.
Holt-Winters forecasting, GARCH/VaR, K-Means++, OLS regression, Granger causality — implemented from scratch.
Offline-first, low-bandwidth, Swahili-capable. Systems that work on $80 phones and over SMS/USSD.
Every system below is designed, built, evaluated, and maintained end-to-end — live for real users, on real data, with real consequences.
A multi-agent data intelligence PWA running a multi-agent system across a 7-stage parallel DAG pipeline (v7.0). Statistical algorithms implemented from scratch. Live at busaraai.com with a Free → Pro → Team → Enterprise tier (99.9% SLA).
AI-powered agricultural intelligence for smallholder farmers across 8 IGAD countries — a 300M+ person market. On-device crop disease detection (TFLite) for 15 crops, market forecasting with GARCH/VaR, multi-LLM advisory council in Swahili + English, 9 free keyless API endpoints, SMS webhook for feature phones. $0/month to operate on free data sources.
A universal AI workflow agent for sales & ops teams. Automates post-meeting followups, lead assignment, support triage, approval routing, and weekly summaries across Gmail, Google Calendar, and HubSpot. Multi-tenant SaaS with real auth, OAuth, encrypted tokens, Celery workers, and Langfuse observability. Shipped with a transparent 281-defect code review and full P0/P1 remediation in one week — 83 backend tests passing.
Real-estate intelligence and trust infrastructure for African property — one ecosystem covering the full transaction lifecycle: Discover → Verify → Analyse → Finance → Invest → Transact → Manage. Verified titles & documents, investment scoring on every listing, market data, tokenized assets, and an "Ask Keja AI" assistant — with dedicated workspaces for buyers, sellers, landlords, tenants, investors, developers, banks, and institutions. Built for CHACADOM INVESTMENTS; live in public trial at keja.app.
Deterministic replay and counterfactual debugging for AI agents. Captures every LLM call, tool response, clock read, and RNG draw — then replays the run bit-exact with zero model calls. 0.67% overhead, 100% reproduction fidelity, 99 tests. LangGraph-first. Backed by a research whitepaper.
A real-time transaction fraud scoring engine in Go — 7 statistical detectors (Z-Score, IQR, Velocity, Geo-Distance, Device Fingerprint, Merchant Risk, Behavioral Anomaly) fused by a weighted-vote ensemble, calibrated by logistic regression, and overlaid with a deterministic rules engine. Scores in ~33 µs per transaction. Full case-management workflow, RBAC auth, Prometheus metrics, OTel tracing, and Slack alerts. 84% recall, 1.37% false-positive rate on 1,000 labelled transactions. Designed with controls relevant to PCI-DSS, PSD2, and GDPR Article 22 — formal compliance is deployment-specific.
33-agent self-evolving multi-agent data intelligence platform. 7-stage pipeline with trajectory-based learning inspired by the AReaL paper. Turborepo monorepo, Next.js 15.
GitHub →Production-grade time series forecasting toolkit. Holt-Winters, ARIMA/SARIMA, Prophet-style decomposition, ensemble forecasting with auto model selection. 87.3% mean accuracy across 12 financial metrics.
GitHub →RFM analysis + K-Means + Gaussian Mixture Models for actionable customer segmentation. 6 optimal clusters, 89% month-over-month stability, 23% lift in campaign response rate. Built for financial services.
GitHub →Celebrating climate leadership across Africa — an initiative of ACLA. Built with Next.js, Three.js, and AI. Live at ecoafricaawards.com.
GitHub →Click a category to filter. Every tool here is something I've shipped to production — from multi-agent DAGs to on-device ML.
Live from my GitHub — what I've been pushing lately. Updated in real-time via the GitHub API.
10 long-form articles on production AI — agent reliability, RAG, offline-first systems, statistical engineering. Three that buyers tend to start with:
AgentReplay reproduces agent failures bit-exact with zero model calls — the research basis behind the Reliability Audit offer.
Read the article → ForecastingHolt-Winters, GARCH, OLS — implemented from scratch. Why real math matters when forecasts drive decisions.
Read the article → Offline-firstOn-device disease detection, SMS fallbacks, and Swahili delivery — AI that works where connectivity doesn't.
Read the article →Direct channels for existing conversations — new ones usually start with the assessment call below.
Bring one workflow, bottleneck, or AI idea. You'll leave with an honest read on what's worth building, what's risky, and the smallest useful first step — even if that step is "not yet."
Prefer async? Email me directly — or find me on WhatsApp. Not ready to talk? Score your AI readiness in 2 minutes — free, private, no email needed.