Benan D. Allan Technical Portfolio · Data & AI

Benan D. Allan

AI Systems Architect · PhD Cognitive & Computer Scientist

A decade of designing predictive intelligence systems for health and behavioural domains — interrogating what each data point truly represents, establishing evidential weight through statistical testing, and constructing variables when the signal doesn't exist in the raw data. Background spanning cognitive science and brain & computer science, applied to the foundational layer that sits above the modelling layer.

10+ yrs
Predictive systems experience
6M
Users in production · Stardust
9 dom
On-device temporal CNN · Nar
PhD
Padova / Edinburgh
Contact

benan@cogdev.ai

+33 7 88 51 52 08

Perpignan, France

Online

benandemir.com

cogdev.ai · CID consultancy

Languages: EN · TR · IT · ES · FR

Selected work · 2014 — 2026 Page 1 / Cover
Technical Portfolio · Data & AI Systems · 2014 — 2026

Predictive intelligence, built from the data up.

A technical portfolio for engineering and data teams. A decade of designing predictive systems for health, behavioural, and financial domains — schemas, feature engineering, model architecture, ML pipelines, serving infrastructure, and production deployment. The architecture decisions that determine whether a model works in production are usually made before training starts. That's the layer this work lives in.

Available for select consulting · CID · Perpignan / Remote
See the projects → Predictive systems
Scroll
Predictive Architecture
Feature Engineering
Multimodal Fusion
Statistical Testing
Behavioural Modelling
EEG · fMRI · GSR
ML Pipelines
Deployment & MLOps
Predictive Architecture
Feature Engineering
Multimodal Fusion
Statistical Testing
Behavioural Modelling
EEG · fMRI · GSR
ML Pipelines
Deployment & MLOps

Approach

What does this data actually mean?

Most teams skip the foundational layer — what each data point truly represents, how much evidential weight it carries, whether the signal is even there to be modelled. I start there.

My work sits above the modelling layer: representational data design, evidential weighting through statistical testing, and constructing variables when the signal doesn't exist in the raw stream. The architecture decisions made before a model is trained — those are usually the ones that determine whether it works in production.

A decade across cognitive neuroscience, women's health, behavioural coaching, immersive tech, and quantitative finance. Different domains, same method.

Capabilities

Six layers of an
intelligent system.

From interrogating raw data to shipping the model behind a production endpoint.

01 / Architecture

Predictive system architecture

End-to-end design of the system that turns signal into prediction — data flow, feature stores, model registry, serving layer, feedback loops.

System designFeature storesServing
02 / Representation

Representational data design

Deciding what each data point means, what it can carry, and how it should be encoded before it ever reaches a model. The decisions made here determine ceiling.

SchemaEncodingDomain modelling
03 / Features

Feature engineering

Constructing variables when the raw signal doesn't carry them — from temperature-derived ovulation indicators to absorption ratios in orderbook microstructure.

BBT modelsMicrostructureDerived signals
04 / Evidence

Evidential weighting

Statistical testing applied at the data layer — what should this point count for, how confident is the signal, when does the model defer instead of guess.

Hypothesis testingCalibrationUncertainty
05 / Behaviour

Cognitive & behavioural modelling

Modelling humans at the individual rather than aggregate level — personality, intention, trust, attention, decision-making. Built on a cognitive-science framework.

PersonalisationProfilingCognitive traits
06 / Deployment

Pipelines & deployment

From notebook to production endpoint — pipeline orchestration, VPS / serverless deploy, monitoring, and the boring infra that keeps inference running at 3 AM.

MLOpsFirebaseVPS / systemd

Featured projects

Production systems I've shipped.

Each build below shows the full stack — data layer, models, training methodology, serving infrastructure, and operational scale.

Stardust · Head of Science & AI · 2024 — 2026 Production · 6M users

Stardust Predictions

BBT-based cycle, period & ovulation prediction at 6M-user scale

Built the AI function from the ground up — predictive models, data architecture, and ML infrastructure where none previously existed. BBT-derived models for cycle length, period length, and ovulation onset; spec-first feature design before training. Led a science team of 3 and guided an engineering team of 5 across research, architecture, and deployment.

ModelsBBT trend detection · cycle-length regression · ovulation classifier
Training dataBBT timeseries · cycle history · self-reported symptom & behavioural events
Eval80%+ real-time accuracy · 90%+ with optimal use, measured on next-cycle prediction
ArchitectureData layer + feature store + ML infra + model registry — built ground-up
PythonTime seriesBBT modellingFeature engineeringProduction MLFemTech
System note Inherited a product with no AI function. Designed and shipped: representational schema for the cycle-event data, BBT-derived feature constructors, a calibrated regression+classifier ensemble, a model registry, and the serving path back to the app — supporting 6M users in production.
6M
Users in production
80%+
Real-time accuracy
90%+
Optimal-use accuracy
3 + 5
Scientists / engineers led
CID · Women's health · 100% on-device ML In development

Nar

Hormonal life-stage intelligence — Core ML, on device

iOS app covering 9 health domains across 6 hormonal life stages. Full science spec written before the model layer. All inference runs on-device — no server-side ML, no biometric data leaving the phone.

ModelTemporal CNN — 9 domain scores from 14-day biomarker windows (Float16, Core ML)
Data inputsHRV · resting HR · sleep stages (deep/REM) · SpO2 · respiratory · skin-temp Δ · steps · recovery — via HealthKit & Oura
ArchitectureNARKit internal SPM module · BaselineNormalizer · BiomarkerVector · LifeStageLens · TrendDetector
StackSwiftUI · Core ML (.mlpackage) · HealthKit · Oura SDK · NARKit · NarInsightsTestKit (XCTest)
SwiftCore MLTemporal CNNHealthKitOuraOn-deviceFemTech
CID · Infant nutrition Live · App Store

TinyTaste

Personalised infant nutrition tracking

iOS app for tracking solids introduction in 0–12 month old infants. Behind it: a nutrient profile database, DRI-based deficiency reasoning, and a recommendation engine that personalises against feeding history rather than aggregate norms.

Data layerPer-food NutrientProfile against 0–12mo DRI
ReasoningDeficiency detection · stage-aware suggestions
StackSwiftUI · Firebase Auth / Firestore · Apple Sign-In
PricingFreemium · $4.99/mo · $39.99/yr
SwiftFirestoreNutrient modellingiOS
CID · Behaviour change Beta

AIM

AI-powered goal achievement

An iOS goal-achievement app built on a cognitive-behavioural data model — goal decomposition, motivational state tracking, and AI-driven check-ins that adapt to the user's profile rather than running a fixed script.

Data layerGoal · subgoal · check-in · motivational state
ModellingProfile-conditioned LLM coaching · adaptive check-ins
StackSwiftUI · Firebase · Claude / GPT API
ApproachBehaviour-change theory grounded
SwiftLLMBehaviour changeiOS
CID · Quant / Polymarket Live · VPS

Sniperbot

Multi-asset prediction-market trading

An autonomous trading system for Polymarket's price-touch markets across BTC, ETH, SOL, XRP. Five generations of architecture — from naive sniping to a modular family-state engine with shared blockers, regime ladders, and evidence-based promotion gates.

Data layerCLOB orderbook · whale wallets · structural microstructure
FeaturesTouch probability · absorb ratio · net minting · sweep vs stealth
ModellingFamily-state machine · per-family smart-sell · regime ladder
DeployVPS · systemd · Telegram alerts · auto-halt
PythonPolymarket CLOBMicrostructureSystemd

Predictive systems

Production AI from the CV.

Roles where I built or led the data and AI function — from architecture to deployment.

Stardust App
Head of Science & AI
Jan 2024 — Feb 2026

Built the AI function from the ground up — designed and implemented all predictive models, data architecture, and ML infrastructure where none previously existed.

Developed BBT-based prediction models for cycle length, period length, and ovulation. Achieved 80%+ real-time accuracy and 90%+ with optimal use, serving 6 million users.

Led a science team of 3 and guided an engineering team of 5 across research, architecture, and deployment.

6M
Users served
80%+
Real-time accuracy
90%+
Optimal-use accuracy
CoachHub
Senior Data & Behavioural Scientist
2022 — 2023

Designed the end-to-end personalisation framework from scratch — mapping behavioural and personality data points across the full user journey to build dynamic user profiling.

Outcome: improved coach-coachee match quality by 86% and lesson completion by 56% within two weeks of deployment.

+86%
Match quality
+56%
Lesson completion
2 wks
To impact
Genoemote
Interim CTO
Dec 2021 — 2022

Led research and technical strategy on measuring and analysing brain activity as a physiological signal for emotion recognition.

Designed the architecture of a trustworthiness rating system operating within VR environments — combining behavioural and neurological signals into a real-time assessment tool.

Humaine AI
Lead Cognitive & Computer Scientist · Co-founder
Nov 2020 — 2023

Co-founded and led AI development of five proprietary models — including cognitive trait deduction from online behaviour, personality-driven product recommendation, and UX optimisation based on real-time user profiling.

Guided all model architecture and data design across client implementations, ensuring each system modelled human behaviour at an individual rather than aggregate level.

H-Farm
Advanced Tech Lead Researcher & PM
2018 — 2019

Led immersive technology development for Bulgari, Gucci, Fendi, Adidas, and Ferrari — VR shoe trial, virtual flagship stores, 3D product interaction systems.

Conducted behavioural data analysis of recorded in-VR human sessions to enhance system design, interaction patterns, and UX across all deployments.

CID
Founder & CEO
March 2025 — Present

AI consultancy delivering end-to-end solutions from architecture design to production deployment. Specialise in context-aware AI systems for companies that need to understand human behaviour at a contextual level.

Design custom ML pipelines including data selection, model architecture, and fine-tuning strategies. Parent company for the projects above.

Pipeline anatomy

From raw signal to served prediction.

A canonical pipeline I'll adapt to a domain — what changes is the signal source and the feature constructor; the spine stays the same.

01 · Source
Ingest
Raw events, sensor streams, orderbook, biometrics. Validate at the boundary.
02 · Mean
Representation
What does each point mean? Schema, encoding, missing-data semantics.
03 · Build
Features
Construct variables that don't exist in the raw stream — derived signals, aggregations.
04 · Weight
Evidence
Statistical testing — what should this point count for, with what confidence.
05 · Model
Predict
Model architecture, training, calibration, fine-tuning where it earns its keep.
06 · Serve
Deploy
Endpoint, monitoring, drift detection, feedback loop back to step 02.

Feature engineering

Variables that weren't there yet.

Derived signals built when the raw data didn't carry the answer — across domains.

BBT-derived ovulation
Stardust
Basal body temperature isn't a label — it's a noisy waveform. Built features that surface the thermal shift defining ovulation onset across irregular cycles.
Touch probability
Sniperbot
Polymarket reach / dip markets resolve on path-dependent touches, not endpoints. 2× endpoint-probability fix replaces the naive mispricing estimate.
Absorb ratio & net minting
Sniperbot v3
Structural microstructure signals built from anonymised orderbook flow — sweep vs stealth, absorption against the move, when whales are dark.
Cognitive trait deduction
Humaine AI
Latent personality features inferred from clickstream and dwell behaviour — built to drive personalisation at the individual rather than segment level.
Match-quality features
CoachHub
Mapped behavioural and personality data points across the user journey to build a dynamic profile feeding match-ranking — +86% match quality.
Nutrient deficiency reasoning
TinyTaste
Per-food nutrient profile × 0–12mo DRI windowed against feeding history — features that drive next-food suggestion and gap detection.
VR-session behavioural fingerprint
H-Farm
Recorded in-VR sessions reduced to behavioural features that drove iterative system, interaction, and UX improvements across luxury client deployments.
EEG / fMRI emotion features
PhD · Genoemote
Multimodal physiological-signal fusion (EEG, fMRI, GSR, eye-tracking) into emotion-recognition features for VR-based real-time assessment.

Deployment

Where the systems actually run.

The unglamorous infra layer that keeps inference live.

VPS · systemd
Sniperbot lives here — patch script → scp → ssh → systemctl restart. Telegram alerts via shared bot. Auto-halt on drawdown.
Firebase
Auth, Firestore, Functions for TinyTaste & AIM. Apple Sign-In, Keychain persistence, server-side merge writes.
App Store
iOS distribution for TinyTaste & AIM. xcodegen-managed projects, version source-of-truth in project.yml.
GitHub Pages
Marketing surfaces — cogdev.ai, benandemir.com, this portfolio. Static, fast, custom domain via CNAME.

Methods & tools

From brainwaves to backends.

The kit underneath the work — research instruments, modelling stacks, and shipping infrastructure.

Languages

  • Python data, ML, automation
  • Swift / SwiftUI iOS production
  • JavaScript / Node web, tooling
  • R statistical research

ML / AI

  • scikit-learn · XGBoost
  • PyTorch · TensorFlow
  • Claude / GPT APIs
  • Fine-tuning · prompt engineering

Data

  • pandas · numpy · scipy
  • Firestore · BigQuery · Postgres
  • Multimodal fusion EEG · fMRI · GSR · eye-track
  • Orderbook / market microstructure

Deploy / Ops

  • Firebase · Cloud Functions
  • VPS · systemd · cron
  • GitHub · GitHub Pages · CI
  • Telegram alerts · monitoring

Research instruments

  • EEG physiological signal
  • fMRI neural correlates
  • GSR autonomic arousal
  • Eye-tracking attention & intent

Statistical

  • Hypothesis testing
  • Mixed-effects models
  • Bayesian inference
  • Calibration · uncertainty quantification

Domains

  • Women's health · FemTech
  • Behavioural science · coaching
  • Cognitive neuroscience · VR
  • Quantitative trading · prediction markets

Frameworks I use

  • Representational data design
  • Evidential weighting
  • Spec-first ML
  • Individual-level modelling

Contact

Have data that needs interpreting?

Open for select consulting — predictive system architecture, data & AI function build-out, or a thoughtful conversation about what your data actually means.

benan@cogdev.ai benandemir.com cogdev.ai