Victor Ndunda

• Tapi Learn — client engagement for ACLA (2026): built the 24/7 AI Career Coach (CV reviews, interview prep, career advice), the multilingual NLP pipeline (English · Français · Kiswahili), personalized learning recommendations, and the opportunities board — serving 11,000+ learners across 38 countries with a 4.8/5 average rating (client-reported).
• Keja AI — client engagement for CHACADOM INVESTMENTS (2026): built the "Ask Keja AI" assistant, per-listing investment scoring, and title & document verification tooling for a 7-stage real-estate platform — live in public trial with 87+ listings across 12 Kenyan counties in 3 languages.
• FirstLight (Oct 2026): designed and shipped a self-initiated Geo-AI disaster-response demo — H3 need-score zoning, Dijkstra routing on post-event road graphs, hospital isochrones, human-in-the-loop review, and hash-verified situation reports.
• Deliver end-to-end engagements: needs assessment → architecture → build → evaluation → production deployment → team training. Clients span fintech, agriculture, logistics, real estate, and SMBs — from $2,500 assessments to $50,000+ multi-agent programs.
• Built the IntelliFlow platform — a 33-agent TypeScript system with self-evolving trajectory-based learning that became the foundation of Busara AI.

• Designed and shipped AgentReplay — an open-source deterministic replay and counterfactual debugging layer for AI agents. Records every LLM call, tool response, clock read, and RNG draw, then replays the run bit-exact with zero model calls.
• Measured 0.67% recording overhead (vs. 5% target; incumbents range 0–15%), 100% reproduction fidelity on the synthetic validation set, and 99 tests all passing.
• First-class LangGraph integration via bind_graph context manager; adapters for OpenAI SDK, Anthropic SDK, CrewAI, AutoGen (v0.2 + v0.4+), and raw agent loops.
• Counterfactual mutation engine — "would the agent still have done X if Y had been different?" becomes a 5-second, zero-cost experiment.
• CI regression runner: captured failures become permanent, zero-cost tests on every PR.
• Published a 22-minute research whitepaper covering the METR time-horizon research, the observability gap, and the full architecture.
• Stack: Python, LangGraph, OpenAI SDK, Anthropic SDK, Click CLI, mkdocs-material docs, GitHub Actions CI.

• Designed and shipped a multi-agent platform with a 6-stage parallel DAG pipeline (Ingest → Engineer → Detect → Forecast → Infer → Report) with circuit breakers and smart caching — serving live users at busaraai.com.
• Implemented real statistical algorithms from scratch: Holt-Winters forecasting, Z-score/IQR/EWMA anomaly ensembles, OLS regression, Granger causality, K-Means++ clustering.
• Integrated GLM-4.6, GPT, Claude, and Gemini into a multi-LLM council with streaming responses and citation-grounded RAG.
• Cut average analysis latency by ~40% via smart caching and parallel DAG execution.
• Stack: TypeScript, Next.js 15, React 19, PostgreSQL, Redis, Docker, Kubernetes, PWA.

• Built an offline-first Flutter app for African farmers with on-device crop disease detection via TensorFlow Lite — runs on $80 Android phones.
• Engineered market price forecasting using GARCH/EWMA/VaR, and a multi-LLM advisory council covering agronomy, markets, climate, and finance.
• Integrated 22+ data sources: KilimoSTAT, FAOSTAT, Google Earth Engine, KALRO, open weather APIs.
• Aligned with Kenya's NASIP 2026-2030 agricultural strategy; supports SMS/USSD channels for low-end phone users.
• Open-sourced on GitHub; designed for scale across East African smallholder farmers.

• Conducted in-depth research on retirement-industry trends and regulatory frameworks to support evidence-based policy recommendations.
• Collected, cleaned, and analyzed large datasets for senior management and stakeholders — detailed reports and visualizations.
• Research underpins the publication The Transformative Role of AI and Machine Learning in Modernizing the Retirement Industry.

• Executed ETL (extraction, transformation, loading) processes for financial datasets across pension products.
• Created interactive dashboards and BI reports tracking key metrics for financial advisors and management.
• Conducted statistical trend analysis and translated data insights into actionable recommendations.

• Delivered financial forecasting systems (ARIMA, Holt-Winters, exponential smoothing, backtesting) for retirement benefits and fintech clients — 87.3% mean accuracy across 12 financial metrics.
• Built a real-time fraud-detection microservice in Go — Z-Score + IQR + Velocity detectors in a weighted-vote ensemble, scoring transactions in ~50–300µs for 50,000+ TPS.
• Shipped customer-segmentation engines (RFM analysis, K-Means, Gaussian Mixture Models, DBSCAN) for financial services — 6 optimal clusters, 89% month-over-month stability, 23% increase in campaign response rate.
• Built data pipelines, dashboards, and analytics tooling using Python, pandas, scikit-learn, statsmodels, and Plotly.

• Processed and analyzed benefits applications in a regulated pension environment, ensuring compliance with statutory requirements.
• Used data analysis to identify patterns and optimize benefits-processing workflows.
• Managed sensitive member data with strict adherence to security and confidentiality protocols.

• Coursework: Data Structures & Algorithms, Database Management Systems, Statistical Methods, Software Engineering, Network Security, Machine Learning.
• Final-year project: an early crop advisory system that seeded the KilimoPRO concept.
• Achievement: Top-quartile standing in the School of Computing.

Real-estate intelligence & trust infrastructure for CHACADOM INVESTMENTS — a 7-stage transaction lifecycle (Discover → Manage), per-listing investment scoring, verified titles, tokenized assets, and the "Ask Keja AI" assistant across 8 stakeholder workspaces. 87+ listings, 12 Kenyan counties, EN·SW·FR.

Multilingual AI learning platform for ACLA — 24/7 AI Career Coach, EN·FR·SW NLP pipeline, personalized recommendations, and opportunities board. 11,000+ learners across 38 countries, 4.8/5 average rating (client-reported).

Geo-AI decision layer for disaster response — H3 need-score zoning, Dijkstra routing on post-event road graphs, hospital isochrones, human-in-the-loop review, and FNV-1a hash-verified sitreps. Fully client-side; every artifact simulated and provenance-labeled.

Deterministic replay and counterfactual debugging layer for AI agents. Records every LLM call, tool response, clock read, and RNG draw — replays bit-exact with zero model calls. 0.67% overhead, 100% reproduction fidelity, 99 tests. LangGraph-first with adapters for OpenAI, Anthropic, CrewAI, AutoGen. MIT-licensed. Backed by a research whitepaper.

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: FastAPI + Celery + Next.js 14 + SQLAlchemy 2 + Alembic, with real auth (JWT + Clerk webhooks), OAuth (Google + HubSpot), encrypted tokens (Fernet), Langfuse observability. Shipped with a transparent 281-defect code review and full P0/P1 remediation — 83 backend tests passing.

Enterprise-grade 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, overlaid with a deterministic rules engine. Scores in ~33 µs/transaction. Full case-management workflow, RBAC auth (API key + JWT), Prometheus metrics, OpenTelemetry tracing, Slack alerts. 84% recall, 1.37% FPR on 1,000 labelled transactions. Built for PCI-DSS / PSD2 / GDPR Article 22 compliance. Multi-stage distroless Docker image (~15 MB), Kubernetes deployment with HPA.

Production multi-agent data intelligence platform with a 6-stage parallel DAG pipeline, real statistical math, RAG with citations, and streaming responses. GLM-4.6-powered.

Flutter + TensorFlow Lite app for African farmers. On-device crop disease detection, GARCH market forecasting, multi-LLM advisory council, 22+ data sources, SMS/USSD fallback.

33-agent self-evolving multi-agent data intelligence platform. 7-stage pipeline (Ingest → Engineer → Detect → Forecast → Infer → Cluster → Report) with trajectory-based learning inspired by the AReaL paper. Turborepo monorepo, Next.js 15, TypeScript, 18 passing tests.

Production-grade time series forecasting toolkit for financial metrics. Holt-Winters, ARIMA/SARIMA, Prophet-style decomposition, ensemble forecasting with automatic model selection. 87.3% mean accuracy (1-MAPE) across 12 financial metrics with 12-24 month forecast horizons.

RFM analysis + K-Means + Gaussian Mixture Models for actionable customer segmentation. 6 optimal clusters (silhouette 0.42), 89% month-over-month stability, 23% increase in targeted campaign response rate. Built for financial services with 3D PCA/t-SNE visualizations.