Real AI & Data Systems,
Built Under African Constraints.
Most technical content assumes stable power, cheap GPUs, and fast internet. I document what actually works in production. Systems engineered in Kenya, tested against the constraints most tutorials never mention.
Read by engineers building real systems.
Signal Feed
Live signals from X, Hacker News, GitHub, and Reddit
Data Engineering
Pipelines & ETL Systems
Most failures are not the model. They are the pipeline.
Artificial Intelligence
LLMs & Deep Learning
Agents are overhyped. Reliability is the real bottleneck.
Machine Learning Systems
Model Training & Deployment
Training metrics lie. Production metrics tell the truth.
AI Automation
Agents, MLOps & Workflows
The hardest part is not building the agent. It is knowing when not to.
From the Field
Systems analysis, architecture deep dives, production postmortems, and applied research. Written from experience, not speculation.
The Hidden Cost of Switching Embedding Models
We changed our embedding model mid-project. Here's what broke, what we didn't anticipate, and what re-indexing at scale actually costs.
12 published · 1 in progress
Ongoing
Signature Series
A 3-part production series. Read in order. Each part builds on the last.
PART 01
Signature Series
How It Actually Runs
Real systems, explained without the marketing diagram.
Every system that runs at scale has a gap between the architecture diagram and what's actually deployed. This series tears down real systems — payments infrastructure, data platforms, ML pipelines — and explains exactly how they work in production, constraints included.
In this series
How M-Pesa Handles 10 Million Transactions a Day
Inside a Real-Time Fraud Detection Pipeline
How Recommendation Engines Work at Low Data Volume
PART 02
Signature Series
If I Built This Today
Retrospective system design. What experience actually changes.
Looking back at systems I actually built in production. Not a highlight reel — the real decisions I made, why I made them given the constraints at the time, and what I would do completely differently with what I know now.
In this series
The Data Pipeline I Regret Building
Our First ML System: Every Wrong Assumption, Documented
Rebuilding a Feature Store After It Failed at Scale
PART 03
Signature Series
What Breaks in Production
Honest postmortems. No blame, no hand-waving.
Production failures documented in full: what the symptoms were, what we initially diagnosed, what was actually wrong, and what changed so it would not happen again. Written as engineering records, not incident theater.
In this series
The Silent Data Loss We Found Three Months Late
When Our Model Passed All Tests and Failed in Production
The Latency Regression We Shipped to 100k Users
Start with Part 01 and read in sequence for the full picture.
Applied System Design
Research
Working papers on AI reliability, uncertainty quantification, and trustworthy systems. Built from production experience, not theoretical defaults.
A Reliability-First Framework Combining Adaptive AI with Hallucination Safeguards
Autonomous AI agents in enterprise create a specific tension: the system needs to adapt, but it also needs to produce outputs you can stake a business decision on. This framework combines Reinforcement Learning, RAG, and uncertainty quantification. The core contribution is the H-Score, a metric for measuring hallucination risk as a system property across IT, healthcare, and logistics.
H-Score
Hallucination risk metric for enterprise AI agents
CWAS
Confidence-Weighted Action Selection for reliable RL
REAC
Enterprise Agent Corpus for benchmarking
CRISIS-LLM 2.0: Trustworthy AI for Disaster Response
Existing disaster AI fails on three measurable fronts: false alerts from unverified social data, no real-time multimodal verification, and black-box decisions responders cannot interrogate. CRISIS-LLM 2.0 addresses all three through Cascading Trust Propagation, fusing satellite imagery, drone feeds, IoT sensors, and LLMs into one uncertainty-aware pipeline targeting a 30% false alert reduction.
Cascading Trust
30% false alert reduction vs text-only baselines
Multimodal Fusion
Satellite, drone, IoT, and social data unified
Uncertainty UI
Confidence scoring for emergency operators
Adaptive Multimodal AI for Real-Time Disaster Response Optimization
Kenya loses an estimated KES 300 billion annually to climate-related disasters. The AI tools used in response still rely on text-based social media analysis, static GIS maps, and delayed satellite feeds. This framework fuses IoT sensors, satellite imagery, and social media into a Hybrid Transformer and RL architecture that generates probabilistic risk maps and optimizes resource allocation in real time, validated against 2024 Kenyan flood and landslide events.
Hybrid Transformer-RL
Multimodal fusion feeding a Deep Q-Network for real-time resource optimization
30% Response Reduction
Target improvement validated against real Kenyan disaster events
Edge-First Design
Lightweight inference for field devices in low-connectivity East Africa
Dynamic Bias Mitigation in Social Media Recommender Systems
Algorithmic bias in recommender systems is not a bug you fix once. It drifts continuously as user behavior shifts, social events unfold, and platforms update their models. Static fairness audits catch yesterday's bias. This framework uses bias-aware Gated Recurrent Units to detect fairness drift in real time and online constrained optimization to recalibrate recommendation weights continuously, targeting 40% disparity reduction without full model retraining.
Temporal Bias GRU
Real-time drift detection across gender, language, and geographic protected attributes using rolling recommendation outcome sequences
Online Fairness Optimizer
Lagrangian recalibration via gradient descent maintaining equal opportunity difference below 0.1 continuously
Bias Drift Benchmark
First empirical dataset of fairness drift trajectories across election cycles, viral events, and algorithm updates
Explainable AI for Carbon Accountability in Decentralized Energy Systems across Sub-Saharan Africa
Sub-Saharan Africa contributes 3% of global emissions but bears the worst climate consequences. The continent's 3,000+ minigrids displace diesel with solar but have no carbon accounting layer. This framework combines LSTM, SHAP, and Multi-Objective RL to optimize battery health, energy dispatch, and carbon attribution simultaneously, with explainability designed for community cooperative operators in Swahili, not just data scientists.
LSTM-SHAP Battery AI
Hot-climate battery degradation prediction calibrated for East African PAYG solar conditions
MORL Carbon Optimizer
Multi-objective RL balancing cost and diesel displacement across P2P minigrid nodes
Community-Trusted XAI
SHAP explanations legible to cooperative operators and EPRA regulators, not only analysts
Privacy-Preserving Generative AI for Secure Healthcare Data Sharing in Kenya
Kenya's health AI research is stuck in a standoff: researchers need patient records to train clinical models, but hospitals cannot legally share them under the Data Protection Act 2019. This framework combines Wasserstein GANs with differential privacy to generate synthetic EHRs that are statistically realistic, clinically useful for malaria and maternal health tasks, and provably private, designed for Kenya's 47-county devolved health infrastructure.
DP-GAN Framework
Wasserstein GAN with differential privacy for synthetic EHR generation under Kenya's DPA 2019
Privacy-Utility Balance
Jensen-Shannon divergence metrics preserving clinical signal for malaria and maternal health AI
County-Ready Deployment
Lightweight Docker container integrating with KHIS/DHIS2 across 47 county health systems
Experiments · Prototypes · Real Implementations
System Lab
Real systems built under real constraints. Each entry documents the problem, the design decisions, and what actually happened — including what failed. Not a portfolio. An engineering record.
About the Author
Sam Odongo
Nairobi, KenyaSenior Data Engineer · Applied Data Scientist
I build AI and data systems where most architectures fail. Low bandwidth, unstable infrastructure, real users, real constraints.
Most technical content assumes stable power, cheap GPUs, and fast internet. I work in Nairobi, Kenya, where none of those are guaranteed. That gap between theory and production is where I live and where I write from.
I hold an MSc in Artificial Intelligence and have spent years building production data systems across fraud detection, real-time streaming, RAG pipelines, and MLOps infrastructure. Every system I document here has been tested against constraints most tutorials never mention.
If you are building systems that have to work in the real world, not just on a benchmark, this is where you should be.
Join engineers learning how to build AI systems that don't break in production.
Real systems. Real constraints. Written from Nairobi, tested against the infrastructure most tutorials never mention. No spam. No theory. Unsubscribe anytime.
Written by a practitioner, not a marketer.