Physician (MBChB, MCS) · AI & Software Engineer · Bioinformatics

Samuel Mbote

Physician and AI engineer turning biomedical data into decisions you can trust.

I unite frontline clinical medicine, biomedical research and production engineering — building data-driven, rigorously tested AI across genomics, healthcare and the full stack, with the rigour to make its reasoning verifiable.

  • MBChB · MCS (COSECSA)
  • IBM-Certified AI Engineer
  • Certified Full-Stack SWE
  • Data Science & Annotation
  • Biomedical Research
View my workGitHubLinkedInEmail
SM
Open to work
Clinician · Researcher · Engineer
3-in-1
Clinician · Researcher · Engineer
Tested open-source projects
9
Tested open-source projects
Automated tests written
1,600+
Automated tests written
TP53 pathogenic hotspots correct vs ClinVar/IARC
7/7
TP53 pathogenic hotspots correct vs ClinVar/IARC

About

I'm a physician and general surgery registrar (MBChB; MCS, COSECSA), an IBM-certified AI engineer and an IBM-certified full-stack software engineer — a rare combination of frontline clinical judgement, biomedical research training and the engineering depth to ship production systems end to end. I work fluently across the whole pipeline: framing the research question, designing the study, wrangling and annotating the data, modelling it, and delivering it as a tested, deployable product.

That range is my edge. I bring native command of clinical and molecular biology to data problems most engineers can't read, and I bring rigorous software engineering to research that most clinicians can't build. My work makes AI's scientific reasoning verifiable — hallucination guardrails, ground-truth benchmarks against ClinVar and IARC, and rigorously tested, offline-first pipelines (1,600+ automated tests across nine open-source projects) — so the results hold up to clinical and scientific scrutiny.

Data & Research

I own the whole chain — from the research question and study design, to the data itself, to the model and the shipped product. These are the areas I go deep.

Data Science & Machine Learning

End-to-end data science: exploratory analysis, statistical and ML modelling, feature engineering, and honest evaluation. I turn messy, high-dimensional biomedical data into models whose performance I can defend — measured against ground truth, not vibes.

  • EDA
  • Statistical modelling
  • ML / classification
  • Feature engineering
  • Model evaluation (F1 · ROC · calibration)
  • pandas · NumPy · scikit-learn

Data Analysis & Analytics

I interrogate data to answer real questions: hypothesis testing, biostatistics, cohort and outcome analysis, and clear, decision-ready visualisation. Comfortable from raw SQL and spreadsheets to reproducible analytical pipelines.

  • Biostatistics
  • Hypothesis testing
  • Cohort & outcome analysis
  • Data visualisation
  • SQL
  • Reproducible pipelines

Data Annotation & Labeling

Multi-domain annotation with a clinician's eye for quality. I design labelling schemas and rubrics, run annotation across audio, image and clinical-text modalities, and enforce quality with inter-annotator agreement and model-in-the-loop review.

  • Audio annotation & transcription QA
  • Medical image annotation (radiology · pathology · surgical)
  • Clinical & genomic text labeling
  • Rubric & schema design
  • Inter-annotator agreement
  • Model-in-the-loop QA

Healthcare & Biomedical Research

Trained to do research properly. I scope questions, choose the right study design, and reason from evidence — grounding every AI system in verifiable biomedical fact rather than plausible-sounding output.

  • Study design (RCT · cohort · case-control · cross-sectional)
  • Systematic review & evidence synthesis
  • Experimental design & hypothesis-driven research
  • Clinical & molecular biology
  • Ground-truth benchmarking (ClinVar · IARC)
  • Scientific writing

Science & Methods

The mathematics and physics under the software. Each discipline below is one I use in my own code, with the equation I reach for most often.

Computational Neuroscience

C dV/dt = −ḡNa m³h (V − ENa) − ḡK n⁴ (V − EK) − gL (V − EL) + I(t)

I treat neurons as dynamical systems and spike trains as point processes. I integrate Hodgkin–Huxley conductance models and leaky integrate-and-fire neurons numerically, read their behaviour off phase planes and bifurcations, and model spike counts as Poisson processes. For neural coding I fit spike-triggered averages and linear–nonlinear–Poisson encoders, decode stimuli with Bayesian estimators, and measure what a neuron tells us about the world with entropy and mutual information. Network dynamics, Hebbian and STDP plasticity, and temporal-difference reinforcement learning complete the picture.

  • Hodgkin–Huxley & LIF models
  • Phase-plane analysis
  • Poisson spike statistics
  • STA & LNP encoding models
  • Bayesian decoding
  • Entropy & mutual information
  • Hebbian & STDP plasticity
  • TD reinforcement learning
  • MATLAB · Octave

Biomedical Imaging & Medical Physics

s(k) = ∬ m(r) e^(−i2π k·r) dr ⇒ m(r) = F⁻¹{ s(k) }

Imaging is applied physics written in linear algebra. In MRI I think in k-space: the scanner samples spatial frequencies, and a 2D inverse FFT returns the image, so aliasing, field of view, Nyquist sampling and partial-Fourier acquisition all follow from the maths. In CT the sinogram is the Radon transform of tissue attenuation, and filtered back-projection (ramp filter, then back-project) inverts it. I reason from Beer–Lambert attenuation, Bloch-equation T1/T2 relaxation and the point-spread function to explain contrast, resolution and noise before I segment and quantify anything.

  • k-space & 2D FFT reconstruction
  • Radon transform & filtered back-projection
  • Bloch equations (T1 · T2)
  • Beer–Lambert X-ray attenuation
  • Ultrasound & nuclear imaging physics
  • PSF · MTF · SNR
  • Image segmentation & quantification
  • Medical image annotation

Signal Processing

H(s) = Y(s) / X(s), stable ⇔ Re(pᵢ) < 0 ∀ poles pᵢ

I move between the time, frequency and complex-frequency domains as the problem demands. FFT and inverse FFT for spectral analysis and frequency-domain filtering, Welch power spectral density, the convolution theorem, zero-phase IIR filter cascades in second-order sections for physiological signals such as ECG, and wavelets when a signal refuses to be stationary. Laplace and z-transforms let me write the transfer function, place the poles and zeros, and know whether a system is stable before I simulate a single sample.

  • FFT · inverse FFT
  • Welch PSD
  • Convolution theorem
  • Butterworth / SOS filtering
  • Wavelet transforms
  • Laplace & z-transforms
  • Nyquist–Shannon sampling
  • ECG · R-peak · HRV analysis

Scientific Computing & Numerical Methods

ρc ∂T/∂t = ∇·(k∇T) + ωb ρb cb (Ta − T) + Qmet + Qext

I write simulation code that is verified against the physics, not just run. Explicit and stiff ODE integrators (RK45, BDF, LSODA) for pharmacokinetic and glucose–insulin models; finite differences with Crank–Nicolson and the Thomas algorithm, and small sparse finite-element solvers, for PDEs such as the Pennes bioheat equation; Levenberg–Marquardt and SLSQP for fitting and constrained optimisation. Every solver earns trust the same way: method of manufactured solutions, observed convergence order, Richardson extrapolation, von Neumann and CFL stability, and unit-checked quantities throughout.

  • ODE solvers (RK45 · BDF · LSODA)
  • Finite differences · Crank–Nicolson
  • Finite elements (sparse CSR)
  • Nonlinear least squares
  • Constrained optimisation
  • Convergence & stability analysis
  • Linear algebra & eigen-analysis
  • NumPy · SciPy
  • GPU compute (CUDA · ROCm)

Statistical Programming & Probabilistic Modelling

p(θ | x) = p(x | θ) p(θ) / ∫ p(x | θ′) p(θ′) dθ′

Statistics is how I decide what a result is worth. I write reproducible analyses in Python and SQL: maximum-likelihood and Bayesian estimation, generalised linear models and logistic regression, survival analysis for clinical outcomes, bootstrap and Wilson intervals instead of bare point estimates, and false-discovery-rate control when many hypotheses are tested at once. Monte Carlo simulation propagates uncertainty; Markov chains and hidden Markov models describe sequences; Fisher information and Sobol indices tell me which parameters the data can actually identify.

  • Maximum likelihood & Bayesian inference
  • GLMs · logistic regression
  • Survival analysis
  • Bootstrap & interval estimation
  • Hypothesis testing & FDR
  • Monte Carlo simulation
  • Markov chains & HMMs
  • Identifiability & sensitivity analysis
  • pandas · statsmodels · SciPy · scikit-learn

Spatial Data Analysis & Mapping

∇²f = ∂²f/∂x² + ∂²f/∂y², I = (N/W) Σᵢⱼ wᵢⱼ zᵢ zⱼ / Σᵢ zᵢ²

Medicine is spatial, from a tumour margin on a scan to a disease cluster on a map. I treat images and geographic data alike as fields over space: spatial autocorrelation with Moran's I, kernel density estimation and interpolation (inverse distance weighting, kriging) for disease mapping, and point-pattern analysis for clustering. On image fields I use gradient and Laplacian operators, convolution kernels and Fourier-domain filtering for edges and texture. Coordinate reference systems and choropleth maps turn the result into something a clinician or a health ministry can act on.

  • Spatial autocorrelation (Moran's I)
  • Kernel density estimation
  • IDW & kriging interpolation
  • Point-pattern analysis
  • Laplacian & gradient operators
  • Coordinate reference systems
  • Choropleth & epidemiological mapping

Systems Engineering

ẋ = Ax + Bu, y = Cx + Du, x̂ₖ = x̂ₖ⁻ + Kₖ (yₖ − H x̂ₖ⁻)

I build health technology the way safety-critical systems are built. Requirements are written in EARS “shall” form with IDs and a traceability matrix; the V-model gives every requirement a verification method; FMEA and ISO 14971-style hazard analysis come before code; interface control documents and architecture decision records keep the design honest. Control theory sits underneath: state-space models, PID loops, Kalman state estimation and stability margins. At the organisational scale I model care as queues, flows and feedback loops (Foundations of Healthcare Systems Engineering, Johns Hopkins).

  • Requirements engineering (EARS)
  • V-model & traceability
  • FMEA & risk analysis
  • ADRs · ICDs · C4 / UML
  • State-space & PID control
  • Kalman filtering
  • Queueing & process modelling
  • System dynamics · causal-loop diagrams

Computational Biology & Bioinformatics Algorithms

F(i,j) = max{ F(i−1,j−1) + s(xᵢ,yⱼ), F(i−1,j) − d, F(i,j−1) − d }

Underneath every bioinformatics tool is an algorithm I can reason about. Dynamic programming for Needleman–Wunsch and Smith–Waterman alignment; de Bruijn graphs and Eulerian paths for genome assembly; the Burrows–Wheeler transform and FM-index behind fast read mapping; profile HMMs and position-specific scoring matrices for motifs and protein families; distance and likelihood methods for phylogenetics; and protein-structure reasoning with AlphaFold and ESM-2. At scale I run them as Nextflow and Snakemake workflows, with variant calls benchmarked against GIAB truth sets.

  • Dynamic-programming alignment
  • de Bruijn graph assembly
  • BWT & FM-index
  • Profile HMMs & PSSMs
  • Phylogenetics (NJ · ML)
  • Protein structure (AlphaFold · ESM-2)
  • Variant calling & GIAB benchmarking
  • Nextflow · Snakemake

Skills

Tools I use in shipped, tested work — plus what I'm actively growing into. Nothing here is aspirational unless it's labelled so.

Data science & analytics

  • Exploratory data analysis (EDA)
  • Statistical modelling
  • Biostatistics & hypothesis testing
  • Machine learning / classification
  • Feature engineering
  • Model evaluation (F1 · ROC · calibration)
  • Data visualisation
  • pandas · NumPy · scikit-learn
  • SQL

Data annotation & labeling

  • Audio annotation & transcription QA
  • Medical image annotation
  • Clinical & genomic text labeling
  • Rubric & schema design
  • Inter-annotator agreement
  • Model-in-the-loop QA
  • Quality control

Biomedical & clinical research

  • Study design (RCT · cohort · case-control)
  • Systematic review & evidence synthesis
  • Experimental design
  • Clinical & molecular biology
  • Ground-truth benchmarking
  • Scientific writing
  • Clinical data management
  • Drug safety & pharmacovigilance analytics

AI / ML engineering

  • LLM application design
  • Retrieval-augmented generation (RAG)
  • Multi-agent orchestration
  • Anti-hallucination guardrails
  • Model evaluation & benchmarking
  • Prompt engineering
  • Gemini & OpenAI APIs
  • LLM & AI-agent evaluation
  • Deep learning (Keras)
  • Generative AI applications

Bioinformatics & genomics

  • Variant effect analysis
  • BioPython
  • NCBI Entrez
  • InterPro / domain annotation
  • Phylogenetics
  • ClinVar / IARC benchmarking
  • NGS variant calling (germline · somatic · long-read)
  • Nextflow · Snakemake
  • Mutect2 · Clair3 · Sniffles2 · minimap2
  • GIAB benchmarking · MultiQC
  • AlphaFold & ESM-2
  • AutoDock Vina
  • HL7 FHIR R4
  • CDISC SDTM & ADaM
  • SAS XPT · define metadata

Full-stack engineering

  • Python
  • TypeScript / JavaScript
  • React
  • Next.js
  • FastAPI
  • Django
  • Node.js · Express
  • Streamlit
  • REST APIs
  • Tailwind CSS
  • Docker · Kubernetes
  • Microservices & serverless
  • Git · GitHub Actions (CI/CD)
  • Google Cloud · IBM Cloud
  • GPU compute (CUDA · AMD ROCm)
  • MATLAB · Octave
  • pytest / TDD

Mathematics & physics

  • Linear algebra & vector spaces
  • Fourier analysis (FFT · iFFT)
  • Laplace & z-transforms
  • ODEs & PDEs
  • Probability & Bayesian inference
  • Information theory
  • Numerical analysis
  • Optimisation
  • Electromagnetism & medical physics

Systems & scientific engineering

  • Requirements & V-model
  • FMEA & risk analysis
  • Control theory (PID · state-space)
  • Kalman filtering
  • Verification & validation
  • Reproducible workflows

Currently building & learningIn progress

  • Rust
  • AWS
  • Azure

Projects

Selected work across clinical genomics, structural biology and applied AI — each shipped, tested, and grounded in verifiable results.

Precision Onco Africa

FlagshipLive

26 specialist AI agents, including a 6-agent tumour board, with a built-in reasoning-audit layer

A research platform that reasons over cancer biochemistry — protein structure and folding, variant effect, and drug–target interaction — with an AI-audit layer that scores every agent's output for accuracy, hallucination and citation quality before release. Runs fully offline on 8GB RAM.

  • Structural biology: AlphaFold p53 structures (per-residue pLDDT), ESM-2 variant-effect scoring, ΔΔG stability, cavity/druggability
  • Drug–target modelling: AutoDock Vina docking, MDM2 / APR-246 inhibitor scoring, PARP synthetic-lethality mapping
  • AI-reasoning audit: dual guardrail + per-agent harness scoring accuracy, hallucination rate and citation quality
  • Benchmarked vs ClinVar/IARC: 7/7 known pathogenic TP53 hotspots correctly classified; 566 tests; PHI de-identification; HL7 FHIR R4 output
PythonAlphaFoldESM-2AutoDock VinaDockerLLM agents
Live appGitHub

NeuroSignal

Open source · tested

Drop in an EEG recording, get a validated report whose every number is traced to a computed value

A validation-first EEG pipeline: channel quality, a spectrum split into aperiodic (1/f) and oscillatory parts, phase coupling that ignores volume conduction, candidate seizure-like events, and a plain-language explanation that is mechanically checked against the computed metrics. Every method was tested on synthetic data with known answers, and the hypotheses were pre-registered before the code existed.

  • Aperiodic exponent within 0.014 median error on 200 synthetic spectra; agrees with specparam in 200/200
  • Debiased wPLI connectivity: PLV reported coupling in 92% of uncoupled pairs, dwPLI stays near zero
  • Claim checker caught 2,152/2,152 corrupted explanations; interactive HTML report with hoverable citation chips
  • Pre-registered hypotheses, held-out CHB-MIT test committed before download, failures reported; 107 tests
PythonMNESciPyNumPyClaude APIHTML reports
GitHubPreprintDOI

BioEng-TermBench

Open source · tested

Biomedical engineering tasks with physics-aware graders for training and evaluating AI agents, with the graders validated like instruments

Five terminal tasks in the Harbor format: ECG denoising and HRV, two-compartment pharmacokinetic fitting and dosing, a closed-loop insulin controller tested in the loop, a verified Pennes bioheat solver for thermal ablation planning, and debugging a PK package. Each grader checks engineering correctness (units, conservation, order of accuracy, safety at the true parameters, robustness under uncertainty), not just that the script ran, and is itself validated by mutation testing, independent references in Python, Octave and Julia, and a red team against the sandbox.

  • Grader validation: Wilson / rule-of-three false-reject bounds, realistic wrong solutions as negative controls, grader mutation testing, 7 red-team attacks on the sandbox
  • Models: McSharry dynamical ECG, two-compartment PK with likelihood-ratio regions, Bergman glucose-insulin with PID and exercise, Pennes bioheat with Arrhenius damage
  • Numerics: Crank-Nicolson finite volumes and hand-written finite elements, manufactured solutions and measured order 2.0; GPU benchmark with a fused CUDA kernel (112x vs NumPy in float64, 231x in float32 on a T4)
  • 289 tests; Snakemake validation campaign with SHA-256 provenance; byte-identical verdicts checked in CI; uv-locked, hash-pinned Docker images
PythonNumPy/SciPyCuPy/CUDAGNU OctaveJuliaSnakemakeDockerHarbor
GitHubPreprintDOI

NGS Variant Pipeline

Open source · tested

Raw sequencing reads to an evidence-cited clinical variant report: germline, somatic and long-read, laptop to cloud

A reproducible, containerised Nextflow pipeline: reads, QC, alignment, variant calling, annotation, prioritisation and an AI-assisted, evidence-cited report, with a single MultiQC across every stage. The same pipeline runs on a laptop or scales to cloud/HPC.

  • Short-read germline, somatic (Mutect2) and cohort joint genotyping; long-read via minimap2 + Clair3 + Sniffles2
  • GIAB accuracy benchmarking (precision / recall / F1 split by SNP and INDEL)
  • Evidence-cited clinical report plus HL7 FHIR output
  • 164 tests + CI; conda or Docker/Singularity, laptop to cloud
NextflowPythonNGSVariant callingDockerLLM reporting
GitHub

Healthcare AI Radar

Live

Scientific-literature intelligence + a research-publication scout

A resilient pipeline that tracks healthcare-AI literature across arXiv, PubMed, Nature Medicine and the press, ranks it with a transparent Scoop Score, and produces newsletter-ready digests. Its Publication Scout mines the same literature for concrete, venue-matched opportunities to publish.

  • Multi-source ingestion with resilient fetching (browser-TLS fallback for bot-blocked feeds)
  • Five-layer anti-hallucination LLM guardrail — live-verified to produce zero fabricated content
  • Transparent 0–100 Scoop Score; rule-based classification and de-duplication
  • Publication Scout: clusters literature into ranked, venue-matched publishing briefs; 90 tests
Pythonfeedparsercurl_cffiGeminiGitHub Actions
Live RadarGitHub

NeuroSonix

Open source · tested

Locale-agnostic audio annotation + agentic evaluation platform

An audio annotation and agentic-evaluation platform that scores voice-agent replies across five quality dimensions in four role-play domains, with Whisper/Gemini model comparison. Built FastAPI + React, it runs offline with graceful degradation when providers are unavailable.

  • Five-dimension scoring of voice-agent replies across four role-play domains
  • Whisper / Gemini model comparison with physics-based audio quality metrics
  • FastAPI + React architecture; 133 tests, CI/CD and Docker
  • Offline-first with graceful degradation when cloud providers are unreachable
FastAPIReactWhisperGeminiPythonpytest
GitHub

TP53 Bioinformatics Pipeline

Live

Nucleic-acid & protein workflows for TP53, from sequence to phylogenetics

A live bioinformatics pipeline for TP53: NCBI sequence retrieval, DNA translation and 6-frame ORF discovery, pairwise alignment, multi-species phylogenetics, and InterPro domain annotation — plus a cancer-mutation biochemistry heatmap across hotspot residues and cancer types.

  • NCBI Entrez retrieval, DNA translation, 6-frame ORF discovery, pairwise alignment
  • Multi-species phylogenetics (Neighbor-Joining: human, mouse, rat, chimp, zebrafish)
  • EMBL-EBI InterPro domain annotation (25 confirmed TP53 domain hits)
  • Mutation heatmap: 12 hotspot residues × 11 cancer types; 63 tests + CI/CD
PythonBioPythonNCBI EntrezInterProStreamlit
Live appGitHub

CDISC Clinical Trial Safety Dashboard

Open source · tested

Raw EDC data to SDTM, ADaM and regulatory-style safety tables, with a fault-tolerant pipeline and live dashboard

An end-to-end clinical data pipeline: messy EDC-style exports are cleaned and quarantined under 20+ data-quality rules, mapped to CDISC SDTM (DM, AE, LB) and ADaM (ADSL, ADAE, ADLB), checked against Pinnacle 21-style conformance rules, and summarised into the safety outputs a medical monitor or biostatistician reviews.

  • SDTM DM/AE/LB and ADaM ADSL/ADAE/ADLB: study days, baseline flags, treatment-emergent (30-day window) and first-occurrence flags
  • TEAE overview, SOC/PT incidence, risk differences vs placebo with Newcombe hybrid-score 95% CIs, lab shift tables
  • eDISH plot and potential Hy's law screen; SAS XPT v5 export with define-style metadata and a SHA-256 audit manifest
  • 70 tests at 98% coverage; CI on Python 3.11-3.13, bandit, pip-audit, container health check; Docker image on GHCR
PythonpandasCDISC SDTM/ADaMSAS XPTStreamlitAltairDocker
GitHub

BioEvidence Forge

Open source · tested

Self-hosted biomedical evidence monitor: PubMed to a traceable archive and citation-complete evidence briefs

A service that continuously collects PubMed literature for configured research topics, stores every record with full provenance in a local SQLite archive, scores relevance transparently and writes Markdown evidence briefs for human review. Built to respect public API policies and fail safely.

  • Rate-paced PubMed E-utilities client with bounded retries, exponential backoff and Retry-After handling
  • Hardened parsing (defusedxml) and validated records keeping PMID, DOI, PMC ID, source URL and retrieval time
  • SQLite archive in WAL mode with automatic schema migration; FastAPI review service and scheduled collector
  • 146 tests with a 95% coverage gate; unprivileged, read-only container; CI on Python 3.11-3.13
PythonFastAPISQLitehttpxPydanticDocker
GitHub

SurgiLogic-QA

Coming soon

Clinical-reasoning support for post-operative surgical monitoring

A clinical-reasoning tool for post-operative surgical monitoring, built in Python and Streamlit. It draws on real surgical domain expertise to structure post-op assessment and surface reasoning that supports safer monitoring decisions.

  • Structured clinical reasoning for post-operative monitoring
  • Built from first-hand surgical domain knowledge
  • Lightweight Python + Streamlit delivery
PythonStreamlitClinical reasoning
GitHub

Publications, education & certifications

Publications

Medical degrees

  • MBChB (Bachelor of Medicine and Bachelor of Surgery)Kampala International University
  • MCS (Membership of the College of Surgeons), General SurgeryCollege of Surgeons of East, Central and Southern Africa (COSECSA)

Certifications

  • AI Engineering Professional CertificateIBM
  • Full Stack Software Developer Professional CertificateIBM
  • BioinformaticsTechnical University of Denmark (DTU), Coursera
  • BioinformaticsUC San Diego, Coursera
  • Computational NeuroscienceUniversity of Washington, Coursera
  • Foundations of Healthcare Systems EngineeringJohns Hopkins University, Coursera
  • Biomedical Imaging FundamentalsEPFL, Coursera
  • Annotate and Analyze Objects for VisionCoursera
  • AI AgentsMicrosoft
  • AI in Healthcare (Azure)Microsoft