Turning uncertainty into advantage through data, models, and judgment.
Data is Everywhere.
Advantage is Rare.
I Build the Difference.
Too much data and big AI but no clarity? I design, deploy, and scale machine learning and GenAI systems that turn messy, real-world data into decisions people can act on.
Retrieval-augmented generation, tool-using agents and multimodal reasoning — including the harness and loop design that makes agents reliable, not just clever.
Eleven live, running visualisations of the techniques behind the projects above.
Sorting Algorithms
Bubble sort running live — watch it shuffle, compare and settle.
Neural Networks
A forward pass, visualised as signal pulses moving layer to layer.
Optimisation (TSP / VRP)
A real nearest-neighbour tour, redrawn on a fresh set of points every few seconds.
Time Series Forecasting
History drawn solid, forecast drawn dashed with a confidence band — regenerated on a loop.
Retrieve relevant context
Grounded answer
Retrieval-Augmented Generation
Documents get searched, relevant context retrieved, then grounded into the final answer.
Graph Embeddings
Connected nodes settle into clusters in embedding space — structure becomes geometry.
Reinforcement Learning
An agent explores a grid, step by step, chasing the reward — then resets and tries again.
|
Speech Recognition (ASR)
The engine behind Maya — a live waveform transcribed into text in real time.
Optical Character Recognition
A scanning pass over a document, extracting text line by line.
Perceive
Plan
Act (MCP tool call)
Observe
Agent Loops, Harnesses & MCP
The perceive → plan → act → observe cycle that keeps a tool-using agent reliable.
GPU-Accelerated Spatial Filtering
Filtering ~1M+ map features by decade and region entirely in the GPU shader — powers NYC Skyline Growth's live timelapse with zero per-frame recompute.
Featured Systems
Projects, grouped by the kind of problem they solve.
Shipped systems and research builds — each with the stack, the reasoning, and where to see it work.
Every project below is personal work, built and published in my own time. Source code for professional work — including everything built for Wema Bank Plc — is covered by NDA and isn't shown here, though I'm glad to speak to the impact and approach in conversation.
A documentation aid for primary health care in Nigeria. A nurse talks with a mother in English, Pidgin or Yorùbá, and the system drafts a structured antenatal record with WHO-based danger-sign flags. N-ATLaS 8B, fine-tuned with QLoRA, and N-ATLaS speech models draft the record; every finding must quote the words it came from, and polarity and support checks flag negated or unsupported findings for the nurse, who has the final word. Benchmarked on frozen test sets never used to choose settings, with every reported number published as JSON. Ships as a rate-limited FastAPI service on a GPU, an offline clinic hub that runs on an ordinary laptop with no graphics card, and a React phone app. Led with a team of nurses and a pharmacist and submitted to the National AI Innovation Challenge 2026 (NCAIR/NITDA).
A real-time disaster-intelligence system built for Africa. AIDER fuses live weather data, breaking disaster news and satellite imagery, then uses multimodal AI to generate actionable emergency briefs for any location on the continent.
FlaskGoogle Gemini Pro VisionOpenWeather APISerper.devNASA Earth API
An interactive 3D map of all ~1.08M NYC buildings, extruded and colour-coded by construction decade, with a GPU-driven timelapse animating the city's growth from 1800 to 2020. Built on a real join — NYC's official Building Footprints joined to the full PLUTO tax-lot dataset on BBL (99.85% match rate) — rather than a lossy lat/lon spatial join, with decade and borough filtering running entirely on the GPU via deck.gl's DataFilterExtension for smooth interaction at full dataset scale.
A cost/retry/timeout governance layer for LLM agents, built as an inheritable Python mixin rather than an external reverse proxy, so enforcement lives in the same codebase as the agent it governs and holds no matter what the model decides to do. Extracted into its own independent package, published on PyPI as nooa-governance, and hardened with a real test suite, a PEP 561 type marker, CI gating the release itself, and a wall-clock time budget across an agent's whole run.
An end-to-end retrieval-augmented generation pipeline on AWS, built as an AI Engineer Nanodegree capstone: PDFs are ingested into S3, indexed into Aurora Serverless Postgres with pgvector/HNSW, and served through a Bedrock Knowledge Base Agent that reasons over the corpus with Claude — with category-classification prompt validation keeping a heavy-machinery support bot on-topic.
A voice-first AI health companion on a single real-time speech pipeline. Maya holds natural conversation while calling out to a live toolbelt: weather forecasts for trip planning, reminders, health tracking (sleep, water, mood, medication), Spotify playback, messages to a doctor on Telegram, and an emergency SOS with repeating alerts. The public web demo runs in a hardened, metered mode with personal integrations off, and recovers from dropped speech connections mid-conversation.
A LightGBM-powered application built during the 3MTT Nigeria fellowship to predict Multi-dimensional Poverty Index (MPI) and quantify how tech-skills training moves the needle on poverty reduction — open-sourced as a knowledge showcase for fellow analysts.
A real-world variant of the Travelling Salesman Problem: ~3,957 geolocated points assigned across 41 agents, with routes optimised for minimum combined distance and balanced per-agent workload, respecting Earth-curvature distances.
OR-ToolsGeoPandasMapboxGoogle VisionTesseract OCR
09 / 12 · Analytics
NYC Green Taxi Operations & Demand Intelligence
A two-part capstone turning 1.73M raw NYC green-taxi trip records (2017–2020) into a dispatcher-ready analytics model: Python cleans and profiles the data, while a SQL Server build/analysis pipeline answers demand, revenue and zone-performance questions for operations decision-making.
PythonPandasSQL Server (T-SQL)Data CleaningEDA
10 / 12 · ML & Data Science
Nepal Climate Risk & Resilience — "Before the Water Rises"
risk threshold
A data-storytelling app tracing Nepal's climate and disaster risk from 1990 to the 2026 Bhote Koshi flood — from hazard evidence, through exposure and vulnerability, to regression-backed evidence supporting a $100M resilience-fund allocation. Built on NASA POWER, HydroSHEDS, national census, DesInventar/UNDRR and UNDP RAPIDA data.
A purpose-built scraping pipeline that collected fellow-level training data from a specialised public database, feeding the dataset later used to train ESAAM's MPI model.
A proxy-backed scraping pipeline over the Oxylabs API that collects and parses live Indeed job postings, then rolls them up into a cleaned dataset with market insights on roles, locations and demand.
A documentation aid for primary health care in Nigeria. A nurse talks with a mother in English, Pidgin or Yorùbá, and the system drafts a structured antenatal record with WHO-based danger-sign flags. N-ATLaS 8B, fine-tuned with QLoRA, and N-ATLaS speech models draft the record; every finding must quote the words it came from, and polarity and support checks flag negated or unsupported findings for the nurse, who has the final word. Benchmarked on frozen test sets never used to choose settings, with every reported number published as JSON. Ships as a rate-limited FastAPI service on a GPU, an offline clinic hub that runs on an ordinary laptop with no graphics card, and a React phone app. Led with a team of nurses and a pharmacist and submitted to the National AI Innovation Challenge 2026 (NCAIR/NITDA).
A real-time disaster-intelligence system built for Africa. AIDER fuses live weather data, breaking disaster news and satellite imagery, then uses multimodal AI to generate actionable emergency briefs for any location on the continent.
FlaskGoogle Gemini Pro VisionOpenWeather APISerper.devNASA Earth API
An interactive 3D map of all ~1.08M NYC buildings, extruded and colour-coded by construction decade, with a GPU-driven timelapse animating the city's growth from 1800 to 2020. Built on a real join — NYC's official Building Footprints joined to the full PLUTO tax-lot dataset on BBL (99.85% match rate) — rather than a lossy lat/lon spatial join, with decade and borough filtering running entirely on the GPU via deck.gl's DataFilterExtension for smooth interaction at full dataset scale.
A cost/retry/timeout governance layer for LLM agents, built as an inheritable Python mixin rather than an external reverse proxy, so enforcement lives in the same codebase as the agent it governs and holds no matter what the model decides to do. Extracted into its own independent package, published on PyPI as nooa-governance, and hardened with a real test suite, a PEP 561 type marker, CI gating the release itself, and a wall-clock time budget across an agent's whole run.
An end-to-end retrieval-augmented generation pipeline on AWS, built as an AI Engineer Nanodegree capstone: PDFs are ingested into S3, indexed into Aurora Serverless Postgres with pgvector/HNSW, and served through a Bedrock Knowledge Base Agent that reasons over the corpus with Claude — with category-classification prompt validation keeping a heavy-machinery support bot on-topic.
A voice-first AI health companion on a single real-time speech pipeline. Maya holds natural conversation while calling out to a live toolbelt: weather forecasts for trip planning, reminders, health tracking (sleep, water, mood, medication), Spotify playback, messages to a doctor on Telegram, and an emergency SOS with repeating alerts. The public web demo runs in a hardened, metered mode with personal integrations off, and recovers from dropped speech connections mid-conversation.
A LightGBM-powered application built during the 3MTT Nigeria fellowship to predict Multi-dimensional Poverty Index (MPI) and quantify how tech-skills training moves the needle on poverty reduction — open-sourced as a knowledge showcase for fellow analysts.
A real-world variant of the Travelling Salesman Problem: ~3,957 geolocated points assigned across 41 agents, with routes optimised for minimum combined distance and balanced per-agent workload, respecting Earth-curvature distances.
OR-ToolsGeoPandasMapboxGoogle VisionTesseract OCR
Analytics · 2024
NYC Green Taxi Operations & Demand Intelligence
A two-part capstone turning 1.73M raw NYC green-taxi trip records (2017–2020) into a dispatcher-ready analytics model: Python cleans and profiles the data, while a SQL Server build/analysis pipeline answers demand, revenue and zone-performance questions for operations decision-making.
PythonPandasSQL Server (T-SQL)Data CleaningEDA
ML & Data Science · 2026
Nepal Climate Risk & Resilience — "Before the Water Rises"
risk threshold
A data-storytelling app tracing Nepal's climate and disaster risk from 1990 to the 2026 Bhote Koshi flood — from hazard evidence, through exposure and vulnerability, to regression-backed evidence supporting a $100M resilience-fund allocation. Built on NASA POWER, HydroSHEDS, national census, DesInventar/UNDRR and UNDP RAPIDA data.
A purpose-built scraping pipeline that collected fellow-level training data from a specialised public database, feeding the dataset later used to train ESAAM's MPI model.
A proxy-backed scraping pipeline over the Oxylabs API that collects and parses live Indeed job postings, then rolls them up into a cleaned dataset with market insights on roles, locations and demand.
PythonOxylabs APIPandas
Research & Scholarship
Published work, and the discipline behind the models.
Peer-reviewed research and conference work in learning science and social research — the same rigor applied to every model shipped.
Data Analytics & Business Intelligence Tutor · Babskenky
Teaching analytics engineering, dbt, SQL, Python, Power BI, and Git/GitHub workflows to aspiring data professionals — translating production data practice into a structured, mentor-led curriculum.
Analytics EngineeringdbtSQLGit & GitHubPower BIPython for Analysts
Graph & Network Models for African Trade and Economic Systems
How relational, network-based machine learning — beyond flat tabular models — can capture the way African commodity and trade systems actually move together, especially under climate and macroeconomic shocks. Early-stage, ongoing work.
Machine Learning for Financial Risk & Decision-Making
Applying predictive modelling and econometrics to banking and fintech risk problems — bridging his Financial Engineering training with production ML practice in a live banking environment.
Learning Science & the Protégé Effect
Peer-reviewed research on interactivity and the Protégé Effect (teaching-to-learn) as retention factors in biology education — the original research thread his academic career started from.
AI for Public Sector & Social Impact
Applying geospatial analysis, forecasting and NLP to public-interest problems — election-integrity detection, flood prediction, and poverty-index modelling among them.
Open Datasets
Public datasets published for open use are on the way — check back soon, or follow on Kaggle for the first release.
Open Source Contributions
Contributions to open-source projects outside his own repos — coming soon.
Career Trajectory
Experience & Impact
From biology first-class honours to fintech AI — a path built on statistics, research rigor and production engineering.
Kaggle badges: Bronze Medal — Notebook: AIDER · Completed 5-Day Gen AI Intensive (Kaggle × Google) · Python Coder · Kaggle Community Member · 2 Years on Kaggle
Communities
Registered Data Scientist
Data Science Nigeria (DSN/AIPlus/2025/44220)
Registered Data Analyst
Nigerian Society of Data Analysts & BI Experts (NSDABIE)
Registered Teacher
Teachers' Registration Council of Nigeria (TRCN; OS/R/07203)
This is the actual map Time — the AI trained on part of me, below — answers from: experience, projects, research, skills, hobbies, philosophy, wired together as a living graph instead of a list. It's a fraction of what's actually in my head, but it's real, and it's what Time draws on when it talks to you.
Loading the brain…
Query the Brain
Talk to Time.
Time is trained on a slice of my brain — my work, my thinking, my way of answering — and replies as me, in first person. It's not the full me, honestly probably not even 10% of me, but it tries hard with what it's got. Find it again anytime via the floating icon.
time · a partial replica of me, doing its best
Hey — I'm Time, a replica of part of Malik's brain: his work, his projects, his way of thinking, rebuilt to talk to. I don't know everything he does — honestly, probably not even 10% of it — but I answer as him, first person, doing my best with what I've got. Ask me anything.
Let's Talk
Have a problem worth solving with data?
Whether it's a production ML system, a GenAI pipeline, or a research collaboration — reach out directly.