Mohammad Nazrul Ansari

Distributed ML over 1.6M flights, built for decisions rather than scores

PySpark · LightGBM · MLflow · scikit-learn · FastAPI · Streamlit · LangGraph · Docker

• ▸Analysed 319,395 delayed US flights and found the headline result contradicts the assumption: weather causes 7% of delay minutes, while 39.8% come from an aircraft that was already late earlier in its rotation. Security, the thing passengers queue longest for, is 0.2%.
• ▸Engineered features on Apache Spark in a container — window functions, broadcast joins, partitioned Parquet — with every rolling aggregate shifted one day back and a strictly chronological split, so a flight's inputs can never contain its own outcome.
• ▸Trained and versioned a LightGBM classifier in MLflow, then optimised for decision quality rather than leaderboard position: isotonic calibration cut expected calibration error 2.5×, and the alert threshold is derived from the relative cost of a missed delay versus a false alarm rather than the 0.5 default, lifting recall on genuine delays to 77%.
• ▸Made the headline finding traceable rather than only citable. The download had trimmed away the aircraft identifier, so a knock-on delay could be counted and never followed; retaining it turned rotation analysis into a sort — partition by tail number, order by scheduled departure, and the buffer between one leg landing and the next departing is what absorbs a delay.
• ▸Simulating that on the busiest real rotation: ninety minutes injected into a Hawaiian inter-island aircraft flying fourteen legs on 28–39 minute turnarounds propagates to twelve of them and produces 1,187 minutes of cascade, an amplification of 13.2×. The simulation refuses to propagate across a discontinuity where a cancelled leg is missing from the data, because that gap is not a turnaround.
• ▸Kept the 0.672 AUC deliberately honest — the model predicts before pushback, so it never sees departure delay. Shipped PSI drift detection, SARIMA demand forecasting, and a two-agent LangGraph system: a router choosing between an analyst over aggregate history and a cascade specialist over rotations, with every figure returned from code and an explicit refusal when the data cannot support one.

One sentence to a complete storefront, in twenty seconds

Next.js 15 · NestJS 11 · PostgreSQL 16 · Prisma 7 · OpenAI structured outputs · Zod

• ▸Built an AI store builder that turns a one-sentence brief into a full storefront — brand, palette, typography, hero, categories, an eight-product catalogue in AED, and About/Contact pages — with live preview and inline editing.
• ▸Designed the pipeline so the model returns a schema-validated blueprint rather than code: prompt → LLM → StoreBlueprint (Zod) → assembler → deterministic React renderer. Model output is data, never executable, so prompt injection cannot yield XSS.
• ▸That boundary bought inline field-level editing, relational persistence, undo/redo as JSON snapshots, provider portability behind one adapter, and a test suite that runs with no API key — 83 tests passing.
• ▸Shipped as a monorepo with shared domain packages, Google sign-in, a REST surface, and automated screenshot capture driving headless Chrome so the docs never drift from the product.

AI blood report analyser — web & mobile

Next.js 16 · TypeScript · GPT-4o Vision · React Native (Expo) · Firebase

• ▸Built a health platform that analyses PDF and image blood reports with GPT-4o Vision in under 30 seconds, extracting biomarkers and flagging out-of-range values.
• ▸Generated personalised wellness protocols — meal plans, supplement stacks and lifestyle recommendations — from each user's blood chemistry, with health-score trend tracking across reports.
• ▸Added an AI meal analyser giving photo-based calorie, macro and micronutrient breakdowns with per-user daily logs.
• ▸Shipped a React Native companion app for scanning reports by camera, gallery or PDF, with marker-by-marker action plans and six health calculators.
• ▸Secured the platform with Firebase Auth and granular Firestore security rules.

Multi-tenant AI SaaS for real-estate agencies

Python · FastAPI · SQLAlchemy · PostgreSQL · Claude API · LangChain · Next.js

• ▸Built an agentic Claude backend with tool-calling: natural-language property search, lead-to-listing matching with reasoning, and ready-to-send pitches written in the client's own language.
• ▸Designed an async FastAPI + SQLAlchemy backend on PostgreSQL with true multi-tenancy, JWT auth and role-based access.
• ▸Developed a Next.js dashboard with a kanban pipeline, an interactive UAE market map and nine analytics chart types, fed by live listing ingestion and scraping pipelines.

VS Code extension

TypeScript · VS Code Extension API

• ▸Built a VS Code extension that copies a selection together with its file path and line range, formatted for pasting into an AI chat — removing a small friction repeated dozens of times a day.
• ▸Handles multi-cursor selections as ordered blocks, falls back to the whole file when nothing is selected, and can dump every open tab as a single context block with a size warning.
• ▸Zero runtime dependencies, an 11 KB package, and a tagged release workflow targeting both the VS Code Marketplace and Open VSX.

nazsats.com

This site — Next.js 16 with a Supabase-auth admin CMS, one-click AI draft generation, GitHub REST + GraphQL integrations and a hardened security layer (RLS, XSS sanitisation, CSP). Includes a public dashboard tracking my coding hours, languages and streaks from real editor telemetry, synced nightly.

CatCents

GameFi & DAO platform on Monad testnet (Founder, 2023–present). Solidity contracts for minting, burning and reward boosting; React/Next.js dashboard for wallet connection, quests and governance. 30,000+ member community.

Froggy Folios

Bitcoin Ordinals NFT collection with a competitive mini-game platform, leaderboards and whitelist checker.

Bao Bao · Skellies Lab · Test Tube · Smith NFT

NFT mint dashboards and wallet integrations across Aptos, Bera and Injective, including utility NFTs with staking rewards.

Nazsats

• ▸Delivered 20+ AI-powered applications, LLM solutions and full-stack platforms from architecture to deployment.
• ▸Built secure FastAPI, Flask and Node.js APIs with authentication, RBAC and scalable backend architectures.
• ▸Developed AI-powered data pipelines and intelligent search systems using real-estate datasets.
• ▸Delivered blockchain-based applications and smart contract solutions for Web3 clients.

Hudl, Bangalore

• ▸Analysed American football games, tagged in-game events and produced performance datasets used for team and player evaluation.

Strux Inc, Mumbai

• ▸Wrote 50+ SEO-optimised articles across a wide range of client industries, adapting tone and depth to each brand's audience.
• ▸Drove 20,000+ page views through clear, well-researched, beginner-friendly content.

SVKM's UPG College · CGPA 9.93

SVKM's UPG College · CGPA 7.7