name:        Marutey Mani
role:        Full-Stack & ML Engineer · Multi-domain Operator
studying:    B.Tech Computer Science & AI @ Plaksha University  # Class of 2029
based_in:    Mohali, India
focus:       production-grade AI platforms · full-stack systems · MLOps
also:        designer · founder (x2 NGO) · published writer · MUN-er · guitarist
philosophy:  "Bring systems thinking to every context — then ship."

I'm a first-year CS & AI student who builds things that actually run in production — a deployed F1 stewardship intelligence engine over 1,085 FIA decision PDFs, a clinical-AI platform with FDA-aligned fairness gates (AUC 0.82), and a full campus-operations platform shipping to real Android devices. I move fluidly between engineering, operations, design, and research, and I've led a 20+ member cross-functional team and co-founded two NGOs along the way.

Whether I'm managing a Mars Rover project's media pipeline, enforcing demographic fairness on a clinical model, or writing crypto-governance policy for a national newspaper — the through-line is the same: turn messy, real-world problems into shipped systems.

The only searchable, evidence-linked FIA decision corpus in existence.

A production-grade F1 stewardship intelligence platform, built from scratch.

• 📄 Scraped, parsed & structured 1,085 FIA decision PDFs (2019–2025) via a Playwright + pdfplumber pipeline with SHA-256 dedup — linked to race-control messages, team radio (Faster-Whisper ASR + pyannote diarisation) & lap telemetry.
• 🔍 Hybrid BM25 + BGE-M3 semantic search with RRF over a pgvector HNSW index; a calibrated XGBoost 7-class penalty predictor with ECE/F1 validation gates.
• ⚡ Next.js 15 app (10+ pages) + FastAPI (9 routers, 20+ endpoints) with a WebSocket + Redis Pub/Sub live race mode serving top-5 precedents in < 5s — 132 passing tests, green CI, IaC via Terraform + Fly.io + Prefect, tiered monetisation.

Python · FastAPI · Next.js 15 · pgvector · XGBoost · BGE-M3 · Prefect · Terraform  →  github.com/racejudge-hq/racejudge-web

Calibrated ensemble inference with explainability & fairness enforcement.

• 🧠 A 5-model calibrated ensemble (RF, GBM, XGBoost, LogReg, CatBoost) stacked with an out-of-fold meta-learner; isotonic calibration + MAPIE conformal intervals → AUC 0.82, ECE 0.020, SHAP stability ≥ 0.60 Jaccard.
• ⚖️ Enforced demographic fairness (parity, equalized odds) against FDA SaMD / EU AI Act thresholds with automated PROMOTE / REJECT / HUMAN_REVIEW gates; FHIR R4, ICD-10 & version-traceable PDF export; Helm + Terraform + Kubeflow.

Python · scikit-learn · XGBoost · CatBoost · FastAPI · SHAP · MAPIE · Kubeflow  →  github.com/DisturbedSage5840C/NeuroSynth

A deployed, full-stack campus hygiene operations platform.

• 🏢 4 role-based dashboards (Admin, Supervisor, Staff, Resident) — photo-verified cleaning logs, 6-item per-washroom checklists, AI-assisted complaint prioritisation & supply inventory across 5 buildings.
• 📱 Shipped a companion Android APK via EAS Build; deployed on Render.com.

Node.js · Express · PostgreSQL (Neon) · Redis · Knex.js · React Native + Expo  →  github.com/DisturbedSage5840C/uhms

Quantifying the gap between trial-reported outcomes and real patient sentiment.

• 📊 Multi-source ingestion (OpenFDA, Reddit, ClinicalTrials.gov) with asynchronous NLP pipelines (FastAPI + Celery).

Next.js · FastAPI · PostgreSQL · Redis · NLP  →  github.com/DisturbedSage5840C/rwe-tracker

_{🔄 Live card — it re-scouts itself as my stats change. Built by @younesfdj · full scout report →}

I don't just write software — I run things, design things, and write about things.