THE LAB / AI SYSTEMS / FULL-STACK PRODUCTS
I turn messy problems into intelligent systems.
Available for projects · Remote-friendly · CET timezone
Specialising in AI automation, medical ML, and full-stack products. MSc Data Science · First Class Honours · available for freelance projects worldwide.
lab.output
> experiment: ddi_predictor status: complete accuracy: 81% (benchmark) model: FNN + PSL hybrid > experiment: chess_robot status: deployed detection: 87% piece accuracy stack: OpenCV · C++ · Python > experiment: ocr_pipeline status: production speedup: 120% API response client: GWS Technologies > availability: open timezone: CET (UTC+1) response: < 24h
Work / Case records
Experiments
Selected builds across AI automation, medical AI, full-stack platforms, and computer vision.
ExperimentEXP-01
Professional Build
Case recordOCR + LLM Bank Statement Processing
A custom AI pipeline that converts messy bank statements into structured, business-ready data — built with EasyOCR and LLM post-processing.
Problem
Financial documents are slow to process manually and often arrive in inconsistent formats.
Method
EasyOCR handles the document layer. LLM workflows normalize fields, explain ambiguous rows, and prepare structured exports — deterministic extraction kept separate from language-model reasoning so the pipeline stays auditable.
Outcome
Messy financial documents become structured, exportable business data.
ExperimentEXP-02
Computer Vision Prototype
Case recordChess Playing Robot
An autonomous chess-playing robotic arm that sees the board, infers state, and executes moves — 87% piece detection accuracy.
Problem
Robotic chess requires accurate visual detection, board mapping, and movement control.
Method
OpenCV detects pieces and infers board state. Vision output maps into robotic movement via Python/C++ logic — perception connected directly to actuation.
Outcome
87% piece detection accuracy and fully autonomous end-to-end gameplay.
ExperimentEXP-03
Thesis / Medical AI
Case recordDrug-Drug Interaction Predictor
A hybrid FNN + Probabilistic Soft Logic model that predicts drug-drug interactions with 81% accuracy on benchmark datasets.
Problem
Drug interactions are complex, high-risk, and difficult to model using only one reasoning approach.
Method
A neural predictor models interaction patterns. Probabilistic soft logic adds symbolic reasoning pressure — keeping predictions interpretable, not just accurate.
Outcome
81% accuracy on benchmark datasets — blending neural prediction with relational reasoning.
ExperimentEXP-04
Hackathon Build
Case recordBuddy
A conversational therapy platform built around fine-tuned LLMs — designed so the interface, prompting, and emotional cadence work together.
Problem
Students and young people often need accessible first-line emotional support.
Method
Fine-tuned LLM behavior, Python services, and a Vue interface were designed as a unified system around softer conversational engagement — not assembled as separate parts.
Outcome
Emotional engagement up 24% in hackathon testing — measured, not estimated.
ExperimentEXP-05
Web Platform Prototype
Case recordOasisConnect
A full community platform prototype — authentication, posts, cloud image storage, fuzzy search, and social graph mechanics shipped in one build.
Problem
Online communities need simple tools for posting, discovering content, and creating group spaces.
Method
Built authentication, post management, fuzzy search, cloud image storage, and early messaging/follow features.
Outcome
Created a functional full-stack social platform prototype.
ExperimentEXP-06
Deployed Game Experiment
Case recordExploratium
A deployed 2D side-scrolling game with full authentication, state persistence, and a live public URL.
Problem
Interactive game projects require state management, UI flow, player progress, and responsive gameplay.
Method
Built a browser-based side-scroller using JavaScript, HTML, Tailwind, sessions, and cookies.
Outcome
Created a deployed interactive game experiment with login/signup and gameplay.
ExperimentEXP-07
Design System / Community Website
Case recordIEEE Website Wireframe Proposal
A full website proposal for IEEE Middlesex University Mauritius — structured to convert visitors into members, not just list events.
Problem
Technical communities need more than event listings; they need clear membership pathways and credible positioning.
Method
Designed a structured wireframe system that connects community identity, technical content, and membership conversion.
Outcome
Created a structured community website design with a strong membership CTA and technical community positioning.
Work with me
Have a messy process, product idea, or AI problem?
I can help you turn it into a working prototype, automation workflow, or full-stack product.
Services / What I can build
Commercial work, without the noise.
I help teams turn unclear technical ideas into usable systems: models, automations, dashboards, APIs, web apps, and MVPs that can be tested with real users.
Currently accepting new projects. Typical turnaround: 2–4 weeks for prototypes, 4–8 weeks for full builds.Build protocol 01
Intelligence layer
AI & Automation Systems
ForFor teams with a prediction problem, a document workflow, or a process that shouldn't require a human.
DescriptionEnd-to-end AI system design and deployment — from model architecture to production pipelines. Specialising in document intelligence, biomedical ML, and LLM-powered automation.
What I can build
Discuss an AI idea- OCR & document extraction pipelines
- Drug interaction and classification models
- LLM extraction and automation workflows
- ML dashboards and model evaluation
- Deployment-ready prototypes
Build protocol 02
Product layer
Full-Stack Products & MVPs
ForFor founders and startups who need a working product, not a template.
DescriptionFull-stack web applications and MVPs built from scratch — React, Next.js, Node, databases, auth, and deployment. From wireframe to shipped product.
What I can build
Start a project- React / Next.js / Vue.js applications
- REST APIs and backend systems
- Database design and optimisation
- Authentication and user management
- MVP architecture and rapid prototyping
Build protocol 03
Clarity layer
Technical Consulting
ForFor students, founders, and small teams who need architecture clarity before they build.
DescriptionArchitecture reviews, AI feasibility assessments, stack selection, and project planning. A focused session that saves weeks of wrong decisions.
What I can build
Book a consultation- Architecture and system design reviews
- AI feasibility and scoping
- Stack and tool selection
- Project planning and roadmapping
Open collaboration
Bring the problem. I'll map the simplest working system.
About / Principal Investigator
Mikaia Razafintsalama
I'm Mikaia Razafintsalama - Mikey to most people - an AI engineer and full-stack developer currently pursuing an MSc in Data Science at the University of Debrecen.
My work sits between intelligence and usability: models that predict, interfaces people understand, and products that solve real problems. I've worked on medical AI research, OCR + LLM automation, computer vision robotics, full-stack platforms, and community products.
The common thread is simple: I like turning complex systems into tools people can actually use.
PIMikaia Razafintsalama
AI Engineer / Full-Stack Developer / MSc Data Science
MSc Data Science, University of Debrecen, expected Aug 2027
BSc Computer Science, Systems Engineering, First Class with Honors, Middlesex University Mauritius
IEEE leadership and AIESEC/event coordination across technical and student communities
Professional experience with OCR, WordPress plugins, backend optimization, Angular/.NET, Docker, and CI/CD
Reagents / Skill compounds
The stack, catalogued.
I treat skills like reagents: small named tools that become useful when I combine them into an experiment. Pick a category, then hover or tap a tile to see where it appears in the lab records.
Process / Protocol
From hypothesis to deployment.
A compact build protocol for moving from unclear idea to usable system.
- 01
Diagnose
Clarify the problem, users, constraints, and success signal.
- 02
Map
Turn the workflow into architecture, data paths, and product flow.
- 03
Prototype
Build the smallest useful system that proves the approach.
- 04
Build
Expand the prototype into stable product and system pieces.
- 05
Deploy
Ship the system where people can use it.
- 06
Improve
Use feedback and metrics to refine the product after launch.
Contact / Open Collaboration
Have a problem worth solving?
Let's turn it into something usable.