Applied AI for healthcare decision systems.
Not another chatbot — systems people can trust.
I design workflow-integrated AI for regulated environments: decision support, data quality gates, and governance-aware automation.
My focus is deployable systems with clear decision boundaries, explainability, and real operational adoption.
Background
SectionI come from a multidisciplinary background spanning customer success, clinical support, and operational roles in international tech environments.
Working closely with healthcare professionals exposed me to the real friction points of clinical workflows — delays, ambiguity, manual workarounds, and systems that technically “work” but fail people.
My transition to AI started with hands-on problem solving.
I began prototyping decision support patterns and reliability-first data foundations. What started as curiosity became a mission: build MedTech AI that survives contact with reality.
AI Philosophy in Med-Tech
SectionAI in Med-Tech is not about adding a chatbot on top of existing systems. It’s an architectural shift that reshapes how data, decisions, and responsibility flow across the stack.
I focus on data-centric engineering, interpretable decision support, and workflow-integrated AI that clinicians can understand and trust — with governance, safety, and adoption treated as core design constraints.
Technology only matters when it works under real constraints.
Applied Stack & Skills
SectionLanguages
Python (applied), SQL, TypeScript
Applied AI & Decision Systems
Rule-based + model-assisted decision support, human-in-the-loop patterns, explainability, failure modes, governance-aware workflows
ML Foundations (working knowledge)
Baselines, model evaluation, gradient boosting concepts, lightweight forecasting patterns
Data Quality & Pipelines
Validation patterns, schema enforcement, trust gating, deterministic preprocessing, edge-case handling
APIs & Deployment (pragmatic)
API prototyping (FastAPI-style), container basics, CI/CD fundamentals, deploy patterns (serverless/VM), monitoring basics
Frontend for Decision UIs
Next.js, React, Tailwind — dashboards and guided UX for non-technical users
Systems & Architecture (conceptual)
System design thinking, observability concepts, policy gates, MLOps basics, reliability-first design
Framework-heavy deep learning is intentionally not the focus of this portfolio. The emphasis is on deployable systems, trust, and decision-making under real-world constraints.
Beyond Models: Product & Delivery Experience
SectionBefore focusing full-time on applied AI, I worked in international tech environments where systems are built, shipped, and held accountable.
This experience shapes how I design AI today: not as isolated models, but as products that must survive real users, workflows, and constraints.
I bring a product and delivery mindset to AI, with experience in cross-functional collaboration, workflow-first design, and stakeholder-ready communication — using tools like Jira, Power BI, and Salesforce in real delivery contexts.
This is why I care about governance, failure modes, and human-in-the-loop design as much as models and metrics.
Selected Projects
SectionThese projects are designed to demonstrate system thinking and real-world Med-Tech impact. The emphasis is on architecture, workflow leverage, and deployable design—not “more code for the sake of code.”
Clinical Data Quality Gate
Trust before metrics
In medical and dental clinics, data often arrives incomplete, inconsistent, or simply wrong. This project shows a reliability-first data quality layer that cleans and validates incoming records — and blocks downstream usage when trust cannot be established.
Input
Messy patient records (inconsistent fields and formats) • Missing or conflicting demographic / clinical values • Free-text notes (when present)
Transformation
Normalization (formats, field mapping) • Deterministic validation and plausibility checks • Explicit trust classification (OK / ATTENTION / BLOCKED) • Downstream gating with clear, human-readable reasons
Output
Trusted, structured patient records • Blocked records with clear reason messages • Suggested next-step actions (what to fix)
Technical choice
Hard validation rules and explicit trust gates; AI is presented only as an assistive extraction layer (not a decision-maker). In clinical workflows, bad data is more dangerous than missing automation. Trust must be established before analytics or decisions are allowed downstream.
What I deliberately didn’t do
No KPI dashboard, no black-box predictions, and no chatbot-driven decision-making.
Snapshot
Hook
Stop bad data before it contaminates clinical workflows.
Problem
Analytics and automation break when source data cannot be trusted.
Approach
Normalize and validate incoming data, then gate downstream usage with transparent rules.
Tech Stack
Next.js, TypeScript, Zod
Impact
Production-style data quality gate that protects downstream analytics and automation.
KPI Command Center
Cross-channel KPIs + explainable 7-day forecast
Most teams track many KPIs but still struggle to answer: What matters today? And what might break next? This dashboard turns customer, sales, and marketing signals into a clear health status, plain-English drivers, and concrete actions — using synthetic data only.
Input
Synthetic daily business signals (customer, sales, marketing) • Region + channel filters (EU/NA/APAC, Email/Chat/Phone) • Scenario sliders (ticket volume, marketing push)
Transformation
KPI aggregation per area (Customer / Sales / Marketing) • Overall health score (0–100) + 'Main signal' summary • Baseline vs model forecast (simulated gradient boosting behavior)
Output
Guided dashboard views (Overview → Area → Actions) • Human-readable 'why' explanations (drivers) • Recommended actions with impact level
Technical choice
Explainable forecasting + decision signals (not black-box predictions). Operational decisions need clarity and trust — the model supports judgment, it doesn't replace it.
What I deliberately didn’t do
No opaque end-to-end pipelines, no proprietary data, no 'AI magic' claims.
Snapshot
Hook
A business dashboard that stays readable, even for non-technical people.
Problem
Cross-team metrics are often tracked in silos, so early warning signals are missed.
Approach
Combine a small set of meaningful KPIs into one health view, then explain what's driving changes and what to do next.
Tech Stack
Next.js, TypeScript, Recharts
Impact
A portfolio-ready example of product-thinking + applied analytics with clear, explainable signals.
Multi-Agent Backoffice: Why Humans Still Matter
A comic-style demo of AI failure modes (and accountability).
A simulated MedTech backoffice where each AI agent makes a locally reasonable suggestion — and the system fails when those suggestions collide. The UI makes failure visible (as a comic), then forces a human sign-off step to show accountability.
Input
A situation snapshot (backlog, budget, constraints) • Short free-text context (what the team ‘hears’) • Simple policies (compliance strict, discount cap)
Transformation
Agents propose actions (structured outputs + confidence/evidence) • Conflict detector flags collisions (growth vs capacity, discounts vs cash) • Policy gate blocks unsafe actions (e.g., compliance shortcuts) • CEO sign-off step enforces human accountability
Output
Agent proposals (comic dialogues) • Visible collisions + severity • Rules: allowed / needs approval / blocked • Post-mortem: what failed and why
Technical choice
Deterministic agent outputs + readable governance layers (no LLM required). For a portfolio demo, reproducibility beats randomness. The message is about failure modes and governance, not model magic.
What I deliberately didn’t do
I intentionally did NOT build a fully autonomous agentic system — the goal is to show where AI stops and humans must decide.
Snapshot
Hook
A comic-style multi-agent office simulator that makes conflicts, rules, and accountability impossible to ignore.
Problem
Multi-agent systems can sound smart individually and still fail as a system: misalignment, overconfidence, policy blindness, brittleness.
Approach
Structured agent outputs → conflict detection → policy gate → CEO sign-off → post-mortem (all visible in UI).
Tech Stack
Next.js, TypeScript, Tailwind, deterministic simulation
Impact
Shows governance and human accountability patterns recruiters actually care about (beyond demos that ‘just work’).