Mohammad Ammar Mughees

• Designed and built Parallax, a ten-stage Python and TypeScript pipeline that converts open-access physics and astronomy papers into cinematic explainer videos using deterministic code-driven animation (Manim, Three.js, Remotion) rather than generative video, with every factual claim traced to a verified location in the source paper
• Implemented a deterministic, evidence-carrying scoring engine for automated paper selection in which each score component is attributed to a named signal with human-readable justification, weights are validated at import time, and the whole ranker is unit-testable offline (845 lines of logic against 579 lines of tests)
• Built a reproducible arXiv ingestion layer that stores raw API responses byte-for-byte with per-page SHA-256 manifests, rate limits politely, and can exactly replay any historical sweep from a date flag

Tech: Python 3.12 · Manim Community Edition 0.20 · Remotion 4.0 · React 19 · react-three-fiber · Three.js · TypeScript · Kokoro TTS

• Designed and built a safety-first automated trading and portfolio assistant for the Indian equity market on the Angel One SmartAPI, architected as four Model Context Protocol servers plus a headless decision agent
• Implemented layered approval gates — a deny-by-default stock universe with filters and a deterministic pre-trade rule gate covering affordability, shorting, per-order value, daily buy caps and maximum position fraction — with live-money order placement walled off behind an explicit phase gate
• Built an event-sourced SQLite paper-trading ledger priced off live market quotes, plus an autonomous scheduler and a token-free event stream for run observability

Tech: Python · FastMCP · Model Context Protocol · Angel One SmartAPI · SQLite event-sourced ledger · Anthropic API · Perplexity Sonar API · matplotlib

• Designed and built TrenLens, an AI-native content platform deployed on Cloudflare Workers, D1 and R2 (566 commits, ~98,000 lines of TypeScript across a 3-app, 5-package monorepo), covering a reader site, creator dashboard, admin panel and four-role role-based access control
• Implemented an in-product AI assistant exposing approximately 99 Model Context Protocol tools through a permissioned, tested tool catalog with a real tool-calling loop and persistent agent tasks stored in Cloudflare D1
• Built an evaluation-gated multi-model router that reduced AI inference cost by 30% on a 44-prompt golden-set benchmark while holding answer quality, task pass rate and tool-call success flat, with an automated gate that rejects any change causing a quality regression

Tech: TypeScript · React 19 · Vite · Cloudflare Workers · Hono · Cloudflare D1 · Cloudflare R2 · Cloudflare KV

• Co-developed EarthLabel, an open-source self-hosted satellite photo-interpretation platform (Node.js, Express, Leaflet) offering 10 land-cover classification schemas, point, pixel and grid assessment modes, multi-source high-resolution basemaps and server-side API key management, as a lighter alternative to Collect Earth Online
• Used the platform to run a 385-point blind human interpretation campaign that provided the independent validation reference for an MSc thesis accuracy assessment, including per-point time tracking and automatic imagery-source provenance logging
• Implemented GIS interoperability across CSV, GeoJSON, KML, KMZ and Shapefile inputs with field-aware CSV and GeoJSON exports and live Google Earth Pro camera synchronisation via KML NetworkLink

Tech: Node.js · Express · vanilla ES modules (no bundler) · Leaflet · Chart.js · Sentinel Hub · Google Earth Engine · Planet monthly mosaics

• Authored the MSc thesis 'Binary Cropland Classification from AlphaEarth Embeddings: A Geospatial Foundation Model Approach for Maine, USA' at Politecnico di Milano, showing that frozen 64-dimensional geospatial foundation-model embeddings classify cropland at 93.75% overall accuracy using lightweight classifiers with no fine-tuning
• Designed and ran a 385-point blind human photo-interpretation validation campaign; the model agreed with the human reference 95.3% of the time versus 91.7% for the USDA Cropland Data Layer labels it was trained on (McNemar exact p = 0.016)
• Demonstrated cross-year transfer from 2018 to 2023 across 36 train-evaluate pairs (92.5-94.3% accuracy without retraining) and a 140x reduction in labelled pixels for a 1.3 percentage-point accuracy cost

Tech: Python · Google Earth Engine · AlphaEarth Foundations embeddings · USDA Cropland Data Layer · scikit-learn · LightGBM · XGBoost · cuML (optional GPU)

• Designed and built a personal portfolio and blog platform in Nuxt 3 and Vue 3 over a Supabase/PostgreSQL content model with a custom admin dashboard, deployed in production on Vercel
• Implemented a bespoke LLM agent runtime for the platform: streaming chat with a typed tool loop, role-based access control and a consent gate on mutating operations, context budgeting with rolling conversation summarisation, and pgvector-backed semantic memory
• Developed a catalog of 40 typed content-management tools and exposed the application as a Model Context Protocol (MCP) server with scope-gated API keys, enabling external AI agents to read and update site content

Tech: Nuxt 3 · Vue 3 · TypeScript · Supabase (auth/storage/Postgres) · PostgreSQL · pgvector · postgres.js · Tailwind CSS

• Built a Claude Code agent that converts raw lecture transcripts into publication-quality LaTeX/PDF study notes, using a chunk-plus-state-file processing pattern so long recordings can be handled incrementally across sessions
• Packaged the agent as a reusable skill with a documented plugin template, separating subject-specific content rules from the generic transcript-to-notes pipeline

Tech: Claude Code agent skills · CLAUDE.md project instructions · LaTeX (custom preamble + templates) · Bash · Python (transcript chunking)

• Designed a cost-optimised multi-model LLM orchestrator that decomposes a prompt into a task DAG and routes each sub-task to the cheapest capable model, escalating only when a quality gate fails
• Implemented the routing layer as a FastAPI proxy with a local sentence-transformer complexity classifier and calibrated sigmoid thresholds, derived from a systematic review of the model-routing literature

Tech: Python · FastAPI · LangGraph · Streamlit · sentence-transformers/MiniLM · OpenHands · Open Interpreter · LLM routing

• Built and publicly released TrenLens Core, an open-source (Apache-2.0) local-first desktop AI assistant in Rust and Tauri v2 that connects any Model Context Protocol server at runtime into a single namespaced tool registry, distributed as signed Windows installers with an in-app verified auto-updater
• Engineered the application security boundary so the Rust host exclusively owns the SQLite database and all credentials: bring-your-own-key provider secrets are sealed with AES-256-GCM under an operating-system keychain master key and never cross the inter-process boundary into the web frontend
• Designed and implemented an end-to-end-encrypted remote-control protocol allowing a mobile progressive web app to drive the desktop engine through a zero-knowledge relay (Cloudflare Worker and Durable Object forwarding only opaque ciphertext) with QR-code pairing and live two-way conversation sync

Tech: Rust · Tauri v2 · rust-mcp-sdk · Next.js 15 · React 19 · TypeScript · Drizzle ORM over IPC · SQLite

• Built LandLens, a pan-India cadastral web platform unifying land records from 28+ separate state government portals into one mobile-first map (Next.js 14, PostgreSQL/PostGIS, Prisma, MapLibre GL), deployed to production and delivered solo across seven sprints
• Designed a pluggable scraper framework for government land-record portals with an abstract base class, per-state implementations, robots.txt compliance, politeness and backoff wrappers, response caching and structured run-result reporting
• Implemented a spatial data model on PostgreSQL/PostGIS with row-level security policies and audit logging, exposed through typed API routes with Zod-to-OpenAPI schemas, idempotency keys and Redis-backed rate limiting

Tech: Next.js 14 (App Router) · TypeScript · Prisma · PostgreSQL · PostGIS · Supabase · MapLibre GL · react-map-gl

• Built and maintained Whisper Transcriber Pro, an open-source cross-platform desktop transcription application in Python (Tkinter, faster-whisper/CTranslate2), released under MIT with tagged GitHub releases and SHA-256 checksums
• Migrated the transcription engine to faster-whisper with batched GPU inference, INT8 quantisation and voice-activity detection, measuring a ~50x real-time speedup (a one-hour recording transcribed in ~72 seconds on a 6 GB GPU)
• Implemented a streaming transcription UI that displays segments as they are produced with confidence indicators and time-based ETA, and exports plain text, detailed transcripts and SRT/VTT subtitles across 50+ languages

Tech: Python · faster-whisper · CTranslate2 · OpenAI Whisper · Silero VAD · CUDA batched inference · Tkinter · FFmpeg

• Built and released PRISMA Review Tool (open source, MIT, v1.5.1), an automated PRISMA 2020 systematic-review system exposing one state model through four interfaces: a command-line tool, a FastAPI service, a Next.js dashboard and a Model Context Protocol server for AI-assisted screening
• Integrated arXiv, OpenAlex, Semantic Scholar and Scopus search with DOI and fuzzy-title deduplication, two-pass screening and automatic generation of the PRISMA 2020 flow diagram, plus CSV and BibTeX export and PDF retrieval via Unpaywall and institutional access
• Used the tool to run a literature review from 3,070 retrieved records down to 17 included studies, producing the PRISMA flow diagram published in the MSc thesis, and a second review of 170 screened records on LLM model routing

Tech: Python 3.10+ · FastAPI · FastMCP (MCP server) · Next.js 16 · TanStack Query · arXiv API · OpenAlex · Semantic Scholar

• Co-developed an end-to-end computer-vision pipeline in Python and OpenCV for bowling video analysis: four-boundary lane detection with angle-constrained MSAC and RANSAC line fitting, ball tracking with homography rectification and RANSAC trajectory fitting, and toppled-pin counting by frame differencing and contour analysis
• Structured the system as three independently runnable packages of roughly 15,200 lines behind a phase orchestrator, with module-level algorithm design documents and a feature-branch and pull-request workflow

Tech: Python · OpenCV · NumPy · SciPy · scikit-learn · pandas · matplotlib · tqdm

• Built FlowCode, a visual data-pipeline builder in which users wire module blocks on a canvas and the system generates, validates and executes the corresponding Python in a Docker sandbox (TypeScript, React, React Flow, Fastify, Prisma)
• Implemented an agentic ReAct planning layer that assembles complete pipelines from plain-English prompts, using typed tools, machine-checkable run artifacts and a pluggable knowledge store that reuses patterns learned from previous runs
• Delivered approximately 39,500 lines of strict TypeScript across a seven-package Turborepo monorepo with 1,010 automated tests, 8 architecture decision records and CI enforcing linting and conventional commits

Tech: TypeScript 5.9 (strict) · React 18 · Vite · React Flow · Tailwind CSS 4 · Zustand + immer · Fastify 5 · Prisma 7

• Built and delivered CraftCart, a full-stack artisan marketplace (Nuxt 4, Vue 3, TypeScript, MongoDB) for an external client, covering dual seller/customer accounts with JWT authentication, a product catalogue across 8 craft categories, cart and checkout, and a five-state order workflow with automatic stock restoration
• Implemented a seller analytics dashboard (revenue excluding cancelled orders, top-selling products, order completion rate) over roughly 15 server API endpoints, plus a GitHub Actions pipeline mirroring the delivered code to the client's deployment repository

Tech: Nuxt 4 · Vue 3 Composition API · TypeScript · Nitro server routes · MongoDB · Mongoose · Pinia · Tailwind CSS

• Developed Yojna Sathi, an Android application (Flutter, Dart, Firebase) that matches Indian citizens to 41+ central and state government welfare schemes, scoring each user 0-100% against mandatory eligibility criteria such as age, social category, occupation, state, education and income
• Implemented the complete application lifecycle - multi-document upload, submission progress and four-state status tracking - plus push and local notifications, on a BLoC clean-architecture codebase backed by Firebase Auth, Firestore, Storage and Cloud Messaging

Tech: Flutter · Dart · BLoC · Clean Architecture · Firebase Auth · Cloud Firestore · Firebase Storage · Firebase Cloud Messaging

• Automated soil classification from Cone Penetration Test data by digitising the Robertson and Campanella (1983) qc-FR chart and fitting its boundary curves as degree-3 polynomials in log space, eliminating the endpoint artefacts produced by linear-space fitting on a logarithmic axis
• Built a Python and Jupyter pipeline that bulk-classifies CPT spreadsheets into five soil types with automatic header detection, interactive column selection, unit conversion and timestamped outputs including a classified workbook, fitted equation coefficients, a chart overlay and scatter plots
• Delivered as a documented open-source tool supporting a collaborator's geotechnical thesis and consulting work

Tech: Python · Jupyter · NumPy polyfit · pandas · matplotlib · SciPy interp1d · OpenCV · openpyxl

• Produced a 43-page technical report benchmarking the CropLand-CD change-detection dataset (Guangdong, China) against Google Earth Engine Dynamic World, including a downsampling validation at r = 0.998 and a reproducible Docker and Jupyter environment

Tech: Python · Google Earth Engine · Dynamic World · CropLand-CD benchmark dataset · rasterio · NumPy/pandas · Jupyter · Docker + docker-compose

• Co-developed and published planetscope-py, an open-source Python library for PlanetScope satellite imagery analysis (temporal and spatial density workflows, data export), released on PyPI with full packaging and documentation
• Delivered as top-graded university project (30 cum laude, Geoinformatics Project, Politecnico di Milano)

Tech: Python · Planet API · satellite imagery analysis · spatial/temporal density analysis · packaging (sdist+wheel) · GitHub wiki docs

• Built AI Screen Overlay, a cross-platform Electron/React/TypeScript desktop tool that captures any screen region via a global hotkey and queries multiple LLM providers (OpenAI, Anthropic, DeepSeek) with persistent SQLite chat history
• Implemented per-token cost accounting for 40+ models across provider pricing tiers, including correct cost attribution when the model is switched mid-conversation
• Added a canvas annotation layer with undo/redo over raw image data and an always-on-top, draggable and resizable overlay window with theming

Tech: Electron · React · TypeScript · SQLite · Tailwind CSS · Webpack · OpenAI API · Anthropic API

• Built StudySync, a Windows desktop application (Python, PyQt5) that infers study activity from active windows and processes and automatically pauses or resumes music across Spotify, VLC, Windows Media Player, iTunes and browser players
• Packaged the application for end users with PyInstaller and an NSIS installer plus a portable build, including system-tray integration, adjustable detection thresholds, activity logging and configuration export/import

Tech: Python · PyQt5 · psutil · pywin32 · winsdk · pycaw · Windows API · PyInstaller

• Designed and implemented a self-hosted multi-tenant Linux cloud platform whose Distribution Abstraction Layer exposes uniform package, service, path and resource APIs across six Linux distributions, systemd and SysV init systems, and cgroups v1 and v2
• Built PAM-integrated user provisioning with SQLite-backed records, per-user CPU, memory, disk-quota and GPU limits, and a token-authenticated Flask REST API behind an SSL-hardened Nginx portal, with shell test harnesses, CI and phase-gated design documentation (phases 0-2 of 6 complete)

Tech: Bash · Python · Flask · SQLite · Nginx · PAM · cgroups · JupyterHub

• Designed and built ui-detector, a packaged Python computer-vision library that detects and classifies user-interface elements in screenshots and extracts positioned text via OCR, exposing results through a structured JSON schema with bounding boxes and confidence scores
• Implemented three interchangeable detection backends — classical OpenCV, vision-language model, and a cost-optimised hybrid that uses computer vision for candidate proposal and calls the paid vision model only for classification

Tech: Python · OpenCV · NumPy · Pillow · pytesseract/Tesseract OCR · OpenAI Vision API (GPT-4o) · Anthropic vision API · PyAutoGUI

• Developed FreshFlow, a Flutter and Firebase mobile application for household grocery and expiry management, covering multi-location inventory, expiry alerts, a real-time shopping list and full email and Google authentication with verification and password reset
• Integrated Google ML Kit barcode scanning against the OpenFoodFacts API and on-device OCR for reading expiry dates from packaging, across roughly 17,300 lines of Dart organised into a layered architecture with a unit and widget test suite
• Delivered as the team coursework deliverable for Design and Implementation of Mobile Applications at Politecnico di Milano (grade 26/30)

Tech: Flutter · Dart · Firebase Auth · Cloud Firestore · Google Sign-In · Google ML Kit (barcode scanning + text recognition/OCR) · GetX · Riverpod

• Delivered an end-to-end data-quality assessment and remediation pipeline in Python over a ~319,000-record public crime dataset, measuring completeness, uniqueness, distinctness and constancy before and after cleaning
• Recovered roughly 11,000 missing geographic coordinate pairs through geocoding and imputed missing categorical fields with a Random Forest model, raising completeness to 99.97% across all attributes
• Resolved functional- and inclusion-dependency violations between linked crime and offense-code tables and standardised inconsistent text encodings, producing a documented, analysis-ready dataset

Tech: Python · pandas · sweetviz · Google Colab · Jupyter · PyDrive2 · geocoding · Random Forest imputation

• Co-authored histogram-manipulation, an open-source Python library for histogram stretching, matching and equalisation of multi-band geospatial rasters with GeoTIFF validation, metadata-preserving output and before/after visualisation, published on PyPI with wheel and sdist packaging
• Owned packaging and release engineering across four versions, maintaining semantic versioning and a changelog and reviewing and merging the co-author's module pull requests through a branch and pull-request workflow
• Delivered as the Geospatial Processing coursework at Politecnico di Milano (grade 29/30)

Tech: Python · rasterio · NumPy · scikit-image · OpenCV · matplotlib · setuptools packaging (wheel + sdist) · semantic versioning

• Built deep-learning competition entries in a four-person team for the Artificial Neural Networks and Deep Learning course, delivering an 8-class blood-cell image classifier that reached 72% accuracy using EfficientNetV2S transfer learning, SHA-256 duplicate removal and class rebalancing, against 62% and 30% baselines
• Designed a custom parallel dual-encoder U-Net with an ASPP bottleneck and a combined focal, dice and boundary loss for 5-class semantic segmentation of Mars terrain imagery, raising mean IoU from 39% to 52% over a plain U-Net baseline

Tech: TensorFlow 2.17 · Keras 3 · keras_cv · EfficientNetV2S · transfer learning · U-Net · ASPP · focal loss

• Implemented a complete GNSS positioning workflow in Python and MATLAB across nine laboratory exercises: reference-frame transformations, satellite orbit computation from broadcast ephemerides, ionospheric delay correction, and pseudorange least-squares position estimation
• Built location-based-service components including IMU dead reckoning, WiFi RSSI fingerprinting for indoor positioning, and Dijkstra shortest-path routing, and validated smartphone GNSS trajectories against a geodetic-grade receiver
• Achieved 30/30 in Positioning and Location Based Services (10 CFU), Politecnico di Milano

Tech: Python · NumPy · pandas · GeoPandas · Matplotlib · MATLAB · GNSS/GPS · least-squares adjustment

• Mapped the 2021 Montiferru wildfire (Sardinia, Italy) from pre- and post-fire Sentinel-2 Level-2A imagery in QGIS, deriving differenced Normalized Burn Ratio and NIR-ratio indices from multi-band virtual rasters
• Reduced cloud-shadow false positives by combining two spectral indices into a single thresholded decision rule, and classified fire severity against published dNBR severity bands
• Delivered the analysis as a 28-slide case study for Earth Observation Fundamentals, Politecnico di Milano

Tech: QGIS · Sentinel-2 L2A · Copernicus Browser · GDAL VRT · dNBR (differenced Normalized Burn Ratio) · NIR ratio index · index thresholding · false-colour composites

• Built and deployed the full technology deliverable for the Hypermedia Applications course: a statically generated Nuxt 2 website for a non-profit domestic-violence support centre, with 16 routes generated from Supabase tables, 18 Vue components, OpenGraph and SEO metadata, and a Netlify plus GitHub Actions deployment pipeline
• Implemented domain-specific safety and support features including a Quick Exit panic button and an in-page assistant calling the OpenAI chat-completions API under a constrained prompt that answers only from the centre's support document
• Delivered the implementation single-handed within a course team that contributed design and content

Tech: Nuxt 2.15 · Vue 2.7 · Vuetify 2 · Tailwind CSS · Supabase · OpenAI Chat Completions API · Google Maps JavaScript API · Netlify

• Developed AUPE RiskMonitor, a landslide-risk decision-support web application for the Lombardy region built on Italy's ISPRA IdroGEO open data, combining a Flask REST API over PostgreSQL/PostGIS with a Dash and Plotly dashboard featuring interactive Folium choropleth maps, filterable charts and drill-down tables
• Acted as repository owner and lead developer on a four-person team, contributing 41 of 59 commits, migrating the spatial database to AWS RDS and packaging the application for Windows with scripted install, launch and uninstall
• Delivered the project with full software-engineering documentation (requirements specification, design document and software release document) for the Software Engineering for Geoinformatics course at Politecnico di Milano

Tech: Python · Dash · Plotly · Flask · Folium · GeoPandas · Shapely · SQLAlchemy

• Built a multi-page WebGIS presenting a landslide-susceptibility study for a pre-Alpine area of Lombardy, consuming 15 WMS layers from a GeoServer instance through OpenLayers 9 with grouped layer switching, scale line, fullscreen and live coordinate readout
• Authored the methodology, results and roadmap content covering data collection, preprocessing, statistical modelling on a balanced 1,000-point sample and accuracy assessment, with Chart.js visualisations of area and exposed population per susceptibility class
• Contributed 30 of 32 commits as lead developer in a three-person group and published the site at no cost via a Vite multi-page build to GitHub Pages

Tech: OpenLayers 9 · ol-layerswitcher · WMS · GeoServer · Vite 5 · Chart.js 4 · Bootstrap 5 · SCSS

• Analysed Persistent Scatterer InSAR displacement time series (Sentinel-1 and COSMO-SkyMed) over a dam in MATLAB, correlating 25 scatterers across 161 epochs against reservoir level and water temperature
• Built and validated multivariate linear and autoregressive models with t-tested coefficient significance, reducing unexplained residual signal power by up to 51% and quantifying the structure's lagged response to environmental drivers
• Applied classical geodetic data-analysis methods end to end: spline interpolation, Fourier spectral analysis, and least-squares collocation with empirical covariance estimation

Tech: MATLAB · MATLAB Live Scripts (.mlx) · PS-InSAR · Sentinel-1 · COSMO-SkyMed · multiple linear regression · t-tests / hypothesis testing · autoregressive modelling

• Conducted a 3-D nonlinear finite-element analysis in Abaqus of a transmission-tower foundation on hill slopes, modelling soil-structure interaction across compression, uplift and seismic time-history load cases for flat and 45-degree slope configurations
• Ran a parametric and mesh-convergence study across three slope angles, two footing depths and three mesh densities to establish result stability
• Published the study as first author in AIP Conference Proceedings 3010, 030013 (2024), DOI 10.1063/5.0193872

Tech: Abaqus/CAE · Abaqus/Standard · finite element analysis · geotechnical soil-structure interaction · seismic time-history analysis · concrete damaged plasticity · ACIS solid modelling · AutoCAD