Hi, I'm Irfana Aslam

Who I Am & What I Do

My Tech Stack

Core specialisations first. Everything below is a tool I have shipped something real with, listed in the order it matters to my work.

BitWithBite AVS, AI Virtual Production Studio

One prompt in. A finished, subtitled, scored, multi-format film out. Built entirely in Python on top of Blender, running offline on a CPU-only machine with no API keys and no subscriptions.

🎥 Watch it work, a film AVS produced by itself

“The Last Lighthouse”, generated from a short written brief. Script, casting, character design, camera work, lighting, voice acting, subtitles, music and the final cut were all produced by the pipeline with no manual editing.

📷 Real output from the pipeline

Every frame below is a genuine, unretouched render straight out of AVS, no concept art, no mockups. Click any still to enlarge.

Dialogue scene & auto-framingCharacter casting, facial expressions, speaker-aware camera, burned-in subtitles

Procedural environmentGenerated set dressing, CC0 textures, narration timing fitted to the spoken audio

Animal character castingScript-driven avatar selection with quadruped rig handling

Mood-driven lightingSky HDRI & colour grade chosen automatically from narrative tone

Preview tier + upscalingRaytraced studio set with practical lights, run through the upscale pass

Character Designer previewPlain-English design → single preview frame in ~30–60s

Auto multi-format exportSame film re-framed to 9:16 vertical for Shorts / Reels

AI/ML Projects

Six flagship systems first, each with a full technical write-up. The complete catalogue of 43 projects follows underneath.

No projects match that search. Show all

★ Flagship Project

🎬

Featured AI Blender Desktop Python

BitWithBite AVS, AI Virtual Production Studio

One prompt becomes a finished film: script, cast, scene, camera, voice, render, edit.

Actual rendered output — cast, set, lighting, dialogue and subtitles, all generated

A complete, offline AI production house built on Blender + Python. One prompt becomes a finished film: the system writes the script, casts and designs characters, builds a 3D scene graph, places cameras and lights, records neural voiceover, renders in headless Blender, then edits, subtitles, scores and exports the result to every social format, with zero API keys and zero running cost. Includes a director-style review loop (samples → notes → approval → final render) so nothing expensive renders before you approve it.

Script generation3D scene assemblyNeural voiceoverHeadless renderingAutomated editing

39Engine modules

~9,900Lines of Python

29Films produced

37Rigged avatars

$0API cost

BlenderPython (bpy) Ollama / Local LLMStreamlit FFmpegPiper Neural TTS NumPyPillow JSON Scene Graph

🎥 Watch Demo & Screenshots 🧠 Full Case Study

● Shipped · 29 films produced

🏗️

AIResearch

AI-CAD, Sketch-to-Energy Building Design

A hand-drawn massing sketch goes in; a measured, energy-simulated building comes out.

Actual pipeline output — voxel massing reconstructed from one sketch

A seven-stage research pipeline that takes a hand-drawn building sketch and returns a simulated building. It retrieves similar CAD drawings, estimates depth, reconstructs the form in 3D, voxelises it into a storey model, writes an EnergyPlus thermal model, then ranks the options on a Pareto front rather than collapsing them into a single score. Every stage has both a research backend and a dependency-free fallback, so the whole pipeline runs on any machine.

0.934Retrieval Precision@1

2.6sPer candidate, CPU

54Modules

28/28Tests passing

Sketch retrievalDepth estimation 3D reconstructionEnergy simulation Multi-objective optimisation

PythonPySide6trimesh NumPy/SciPyEnergyPlus IDFNSGA-II 3D Reconstruction

📖 Case Study

● Client project · delivered & tested

🔐

AIResearch

SmartFaceGuard, Face Recognition Pipeline

Deep-learning facial embeddings paired with vector similarity search for identity matching.

A face recognition system built as a three-stage vision pipeline. MTCNN locates every face in a frame, a FaceNet network converts each one into a 512-dimension embedding, and a FAISS index matches that embedding against known identities by nearest-neighbour search. An image-quality gate screens frames before they reach recognition. Built for scenarios where the identity database keeps growing and linear comparison stops scaling.

Face detectionFacial embeddings Similarity searchIdentity matching Image-quality filter

PythonPyTorchMTCNN FaceNetFAISSOpenCV

📖 Case Study

● Research prototype · local, camera required

🧱

Deep LearningComputer Vision

Fine Crack Detection, U-Net Segmentation

Pixel-level segmentation of hairline cracks for structural inspection.

Actual model output — held-out test set, input vs ground truth vs prediction

A U-Net trained from scratch to segment cracks one or two pixels wide against noisy concrete — the fine-structure case where ordinary detectors fail. Trained and evaluated end to end on a CPU with no GPU. Recall exceeding precision is the right direction to err for inspection work, since a missed crack costs far more than a false alarm.

0.785Dice

0.658IoU

0.858Recall

0.763Precision

36 minCPU training

Semantic segmentationCustom U-Net Dice/IoU evaluationCPU-only training

PythonPyTorch U-NetOpenCVNumPy

📊 Full Results

● Trained & evaluated · measured on held-out split

📈

AIData Science

AI Stock Predictor

Predictive market analytics across 20 interactive dashboards, with demo and live data modes.

Actual application output — labelled demo snapshot, not live market data

Twenty tabs of market analysis in one Streamlit application: technical indicators, VADER news sentiment, fundamentals, backtesting, Monte Carlo simulation, options chains, crypto and portfolio tooling — on a data source needing no API key. Credentials were moved out of the source tree into environment variables, and an explicitly labelled demo/live data split means a recorded snapshot can never be mistaken for a live quote.

31Modules

20Dashboards

7,888Lines of Python

30/30Checks passing

Time-series analysisML forecasting Sentiment analysisBacktesting Monte CarloInteractive dashboards

PythonStreamlityfinance scikit-learnPlotlypandas

📖 Case Study

● Local prototype · research tool, not financial advice

✍️

AILive Product

Smart CV Editor

Upload a CV, edit it structurally, tailor it to a job, export a clean document.

A working résumé-editing product rather than a demo. It reads an existing CV, breaks it into editable structured sections instead of a wall of text, and supports AI-assisted rewriting and job tailoring, with saved versions and document export. Built to be usable by someone who has never opened a developer tool.

Document parsingStructured editing AI-assisted tailoringDOCX export Version history

JavaScriptPHP HTML5CSS3DOCX

🚀 Try the Tool

● Shipped · live on site

🧵

MS ThesisDeep Learning

Fabric Defect Detection, MS Thesis Research

Solving a data problem instead of an architecture problem in industrial visual inspection.

A deep-learning system for automated textile quality control, built around the observation that the binding constraint is data rather than architecture — fabric defects are rare by definition, so real datasets are small and badly imbalanced. The thesis inverts the problem with a generator that synthesises realistic defects with controlled variation in type, size, orientation and lighting, producing balanced labelled data at a scale real collection cannot reach.

>98%Detection accuracy

100,000Images generated

10Defect classes

3Baselines outperformed

Synthetic data generationCNN classification Class imbalanceIndustrial inspection

PythonTensorFlow KerasOpenCVCNN

📖 Case Study

● Research · manuscript in preparation

🤖

AI Live

BitWithBite AI Assistant

Retrieval-augmented support agent that answers from your own documents, with citations.

Production-ready RAG-based AI customer support agent. Auto-crawls the website nightly, processes PDFs and markdown, retrieves with FAISS vector search, and generates answers with source citations. 100% free, zero API costs.

Document ingestionVector retrievalCited answersNightly re-crawl

PythonFastAPILangChain FAISSOllamaStreamlit DockerRAG

🚀 Live Demo 📖 Case Study

● Live on this site

📄

AINLP

AI Resume Analyzer & Career Coach

Parses a CV, scores it against ATS rules, and generates a tailored improvement plan.

Parses PDF/DOCX resumes, extracts 300+ tech skills, runs ATS compatibility checks, semantic scoring, and uses local LLMs to generate cover letters, 8-week learning roadmaps, interview prep, and career plans. $0 API cost.

CV parsingSkill extractionATS scoringCareer planning

PythonFastAPIspaCy sentence-transformersOllamaGemma3

🚀 Live Tool 📖 Case Study

● Live on this site

💻

AIPython

Coding Challenges, Judge & Interview Prep

Sandboxed code judge with adaptive difficulty and an AI debugger.

A gamified Python practice platform: a real server-side judge runs submissions against test cases with a hard-enforced execution timeout and a code-safety pre-check, a step-through debugger traces real interpreter execution line by line, and an AI reviewer critiques submitted code. Also ships a skill tree, career roadmaps, mock interviews, a teacher mode for authoring custom challenges, and a downloadable portfolio of solved work.

Code executionAutomated judgingAdaptive difficultyAI debugging

PHPPython (proc_open)CodeMirror sys.settraceGemini

🚀 Live Tool 📖 Case Study

● Live on this site

🚗

Computer VisionTracking

Vehicle Detection & Multi-Object Tracking

Background-subtraction detection with persistent per-vehicle identity across frames.

Actual run output — detection with persistent tracking IDs

Detects moving vehicles in traffic footage using MOG2 background subtraction inside a defined region of interest, then assigns each one an identity that survives across frames via a Euclidean-distance tracker. A demonstration of classical CV holding its own: no training data, no GPU, and it runs in real time.

Motion detectionMulti-object tracking ID persistenceRegion of interest

PythonOpenCV MOG2NumPy

🎥 Watch the Run 📖 Case Study

● Verified · video output produced

✍️

AINLP

AI Storytelling Studio for Screen Industries

Turns a premise into a structured, scene-by-scene storyboard for screen production.

Actual output, generated storyboard, 4 scenes with camera, lighting and narration

A writing tool for screenwriters and directors, running on local LLMs. The Streamlit dashboard handles single and batch generation, keeps a saved story library, and offers continuation and variation modes plus analytics on story metrics and word frequency. spaCy picks out characters and locations and highlights them inline. Genre, tone, length and creativity are all adjustable, and you can feed it a CSV for batch runs or export what it produces.

Story generationScene structuringEntity extractionBatch generation

PythonStreamlitLocal LLMs spaCyTransformersPrompt engineering

🎬 Related: AVS 📖 Case Study

● Verified · real storyboard exported

🛡️

AIBackend

Aegis, Containerised Content Intelligence Platform

Containerised content-intelligence platform with a operator dashboard.

Dashboard, dispatch, live pipeline monitor and job history (shown idle)

Six services orchestrated with Docker Compose. A FastAPI export gateway takes the request, an orchestrator manages the flow, a worker handles scraping and file generation, and a content-intelligence service does the AI analysis. PostgreSQL stores results and Redis carries live progress and shared state, all behind nginx on a bridged network. The point of the split is that long content jobs run in the background instead of holding a request open.

Content analysisDashboard UIContainerised deployExport pipeline

PythonFastAPIDocker Compose PostgreSQLRedisnginx Microservices

📖 Case Study

● Prototype · dashboard running

👁️

ResearchComputer Vision

Micro-Behaviour Analysis

Tracks 543 body, face and hand landmarks per frame from live video.

A computer-vision system for body-language measurement. MediaPipe Holistic runs over live camera or video and recovers up to 543 landmarks per frame — 33 pose points, a 468-point face mesh and 21 points per hand — tracked continuously and rendered as an overlay. It measures geometry only: the system reports coordinates and makes no inference about emotion, intent or state of mind.

Pose trackingFace meshHand trackingLive video processing

PythonMediaPipeOpenCVNumPy

📖 Case Study

● Research prototype · extraction stage

Research Prototype · extraction stage only

🧠

Local LLMPrototype

NeuroSlide, Local AI Presentation Studio

A topic or a spreadsheet goes in; a real PowerPoint deck comes out, generated offline.

Actual generated charts — 9 produced in one pass, 4 shown

Type a topic, get a real .pptx. A local model writes the deck structure; 13 hand-built layout templates render it. The model returns a slide type and the builder picks the layout, so it never has to reason about geometry. Verified producing a valid 16:9 deck on llama3.2 with no API key. One of the two prototypes where the zero-API-cost pattern behind the current BitWithBite AI tools was worked out.

Deck generation13 slide layoutsChart generationPDF speaker notes

PythonOllamaStreamlit python-pptxreportlab

📖 Case Study

● Local prototype · runs offline

Local Prototype · 5 versions · requires Ollama

∞

Local LLMEdTech

MathGenius, Local AI Math Solver

SymPy computes the exact answer; a local model writes the teaching around it.

Actual solver output — roots marked at x = 3 and x = 4

An offline maths tutor that runs two engines side by side. SymPy computes the exact symbolic answer; a local language model writes the explanation, alternative methods, hints and common mistakes around it. Keeping them separate means a small model that misreads a problem cannot corrupt the answer — the symbolic result stands beside it. Covers algebra, calculus, trigonometry, matrices and statistics, with graphs generated per problem.

Symbolic solvingStep explanationsAuto graphingQuiz generation

PythonSymPyOllama StreamlitPlotly

📖 Case Study

● Local prototype · runs offline

Local Prototype · 5 versions · requires Ollama

🧊

BlenderPrecursor

Blender Scene Automation

JSON scene description in, rendered animation out — the origin of the AVS architecture.

Actual render output — scene built entirely from the JSON manifest

A scene-description language for 3D animation: characters, positions, colours, motion targets and duration are declared in JSON, and the script drives Blender to build, light, animate and render the result headlessly. No manual modelling and no GUI step. This is the earliest working expression of the idea BitWithBite AVS is built on — a scene is data, and a program turns that data into video.

Procedural scene buildKeyframe animation Headless renderingData-driven pipeline

PythonBlender (bpy) EEVEEFFmpeg

📖 Case Study 🎬 See what it became

● Verified · renders 48/48 frames

Historical Prototype · repaired, renders

🩺

ResearchRebuilt

Clinical Decision Support System

Six published clinical scoring rules, each shown with its working and its source.

A decision-support calculator covering six published clinical scoring rules — Wells DVT and PE, CHA₂DS₂-VASc, CURB-65, qSOFA and Centor. Every score is returned with the criteria that produced it, the arithmetic, the source citation and the rule's stated limits, so a clinician can check it rather than trust it. Unanswered criteria are reported by name instead of being silently counted as absent. Educational prototype — not a medical device.

Clinical scoringCited sourcesShown workingInput validation

PythonStreamlit Clinical scoring rulesZero-dependency engine

📖 Case Study

● Educational prototype · not a medical device

Local Prototype · rebuilt 2026-08-18 · educational only

🦠

PHPAI

COVID-19 AI Dashboard

Live epidemiological dashboard with forecasting and news sentiment.

Live COVID-19 tracking dashboard with Chart.js visualisations, 14-day case forecast, top-5 countries ranking, and keyword-based VADER-style sentiment analysis. Runs on any PHP host, no Python, no API required.

Time-series forecastingSentiment analysisLive dashboards

PHPChart.js Keyword NLPdisease.shcPanel

🚀 Live Demo

● Live on this site

💰

PHPAI

Competitor Pricing Scraper

Scrapes competitor pricing on a schedule and surfaces movement.

Web scraper that extracts competitor prices from any URL using PHP cURL + regex, visualises price distributions with Chart.js bar and doughnut charts, and generates rule-based market insights (price spread, cheapest/most expensive site). Fully deployable on cPanel.

Web scrapingPrice trackingChange detection

PHPcURL Chart.jsKeyword NLPcPanel

🚀 Live Demo

● Live on this site

🔍

Deep LearningUpwork

Deep Learning-Based Defect Detection Using Advanced Computer Vision

Convolutional defect classification for a live textile production line.

A production defect-detection system for textile manufacturing, delivered as a freelance engagement. Convolutional models classify fabric defects from line imagery, with preprocessing and augmentation tuned for the lighting and motion conditions of a live production line, and transfer learning used to reach usable accuracy without a large in-house dataset. Packaged for automated inspection rather than offline analysis.

Defect classificationImage preprocessingData augmentationTransfer learning

PythonTensorFlow OpenCVCNNTransfer Learning

● Client project · delivered

May 2024 · Upwork Project

🛡️

Computer VisionAI

AI-Based Surveillance System for Threat Detection

Computer-vision threat detection over surveillance video.

AI-driven security monitoring system for real-time threat detection. Integrates YOLOv5 for object detection, enhancing security operations with Flask-based backend.

Object detectionVideo analysisAlerting

YOLOv5OpenCV Deep LearningPythonFlask

● Client project

Oct 2023 – Dec 2024

🤟

Deep Learning

Real-Time Sign Language Recognition

Recognises sign-language gestures from a live camera feed.

Deep learning model to recognise and translate sign language gestures into text in real-time. Increased accessibility for the hearing-impaired community using CNN-based gesture classification.

Gesture recognitionReal-time inferenceHand tracking

PythonOpenCV TensorFlowCNN

● Client project

Sep 2021 – Mar 2022

📋

NLPAI

Automated Resume Screening System using NLP

Ranks candidate CVs against a role using natural-language matching.

AI-driven resume screening system to rank candidates based on job fit. Automated HR recruitment, reducing manual effort by 60% using NLP-powered semantic matching and TensorFlow models.

NLP matchingCandidate rankingBulk processing

PythonNLP SpaCyTensorFlow

● Client project

Oct 2022 – Jun 2023

🧬

BSc ThesisPublished

Multi-Epitope Vaccine Design (Computational Biology)

Computational immunology pipeline for designing a multi-epitope vaccine construct.

Designed a novel multi-epitope vaccine candidate for coronaviruses using in silico approaches. Research published in Microorganisms (IF 4.5), a high-impact peer-reviewed journal.

Epitope predictionMolecular dockingImmune simulation

PythonBioPython Molecular DockingImmunoinformatics

● Published research

Sep 2020 – Jun 2021

🏥

MLHealthcare

Machine Learning Model for Disease Prediction

Predictive model for disease risk from structured clinical variables.

Predictive model for diagnosing diseases based on patient symptoms and test results. Achieved high accuracy in disease classification using real-world medical datasets.

Feature engineeringClassificationModel evaluation

Scikit-learnPandas NumPyMachine Learning

● Client project

Jan 2019 – May 2020

📊

UpworkLive Demo

Competitor Pricing Scraper Dashboard

Scrapes competitor pricing on a schedule and surfaces movement.

Built a full-featured Pricing Intelligence Dashboard using Python, Streamlit, and BeautifulSoup. Scrapes live product pricing data from e-commerce sites, processes it with Pandas, and visualises insights through 6 interactive Plotly charts. Features a category scraper for 29 genres, keyword search, multi-filter sidebar, Value Score ranking system, auto-generated insights, and CSV/Excel export.

Web scrapingPrice trackingChange detection

PythonStreamlit BeautifulSoupPandasPlotly

● Live · deployed on Streamlit Cloud

🚀 Live Demo ⌥ GitHub

Mar 2026 · Upwork Project

🦠

UpworkAI Forecasting

COVID-19 Dashboard, Python, Streamlit, AI Forecasting & News Sentiment

Streamlit dashboard combining case forecasting with news sentiment.

Developed a portfolio-ready COVID-19 analytics dashboard using Python and Streamlit. Implemented dynamic KPIs, AI-driven forecasting (Prophet), vaccination tracking, and RSS-based news sentiment analysis. Features dark mode, country comparison, interactive Plotly charts, and CSV export, demonstrating data visualisation, machine learning, and practical Python development.

ForecastingSentiment analysisInteractive charts

PythonStreamlit ProphetVADER SentimentPlotlyFeedparser

● Local prototype

🚀 Live Demo ⌥ GitHub

Mar 2026 · Upwork Project

🎯

UpworkDeep Learning

AI-Based Object Detection Using Python & TensorFlow

Object detection over images and video streams.

Developed a custom AI-based object detection system using Python, TensorFlow, and Keras. Handled data cleaning, annotation, and augmentation. Trained CNN models with transfer learning and evaluated performance using precision, recall, F1 score, and IoU. Deployed the model using Flask for real-time use.

Object detectionModel servingVideo inference

PythonTensorFlow KerasFlaskCNN

● Client project

Sep 2024 · Upwork Project

🏭

UpworkOptimization

Optimized Production Planning Using Python & Machine Learning

Machine-learning driven production scheduling and capacity planning.

Developed Python algorithms to optimise inventory management and production planning by minimising stock days and improving order fulfilment. Automated Excel data processing and reporting using openpyxl, enabling dynamic calculations for stock levels and production schedules. Created demand and stock simulation scripts for testing with robust error handling for large datasets.

Demand forecastingSchedule optimisationCapacity planning

PythonPandas openpyxlAlgorithm DesignData Modelling

● Client project

Sep 2024 · Upwork Project

🏆

AIWeb

BitWithBite Coding Challenges v2

Coding practice platform with a visual debugger, skill tree and interview mode.

Full AI-powered coding challenge platform with 20 features: Visual Step-by-Step Debugger, Adaptive Difficulty, Skill Tree, Interview Mode, Portfolio Generator, Career Roadmap, and Teacher Dashboard. Embeddable anywhere.

Visual debuggerAdaptive difficultySkill treeInterview mode

PythonFastAPIOllama XP SystemAdaptive AI

● Client project · delivered

🗺️

WebLive

AI Learning Roadmap Generator

Turns a career goal and current skills into a dependency-ordered learning plan.

Tell it your goal and current skills, get a personalised, dependency-ordered learning roadmap with real market data, project suggestions, and certification recommendations.

Skill-gap analysisDependency orderingProject suggestions

HTMLJavaScriptPHP

🚀 Live Tool

● Live on this site

📰

WebLive

AI Paper Summarizer

Structured summaries of academic papers, with findings and methodology separated.

Academic paper summarization tool. Paste an abstract or full paper and receive a structured summary, key findings, methodology, and plain-language explanation, ideal for research workflows.

Structured summaryKey findingsPlain-language rewrite

HTMLJavaScriptPHP

🚀 Live Tool

● Live on this site

💡

AILive

AI Concept Explainer

Plain-English explanations of programming, maths and science concepts.

Get instant plain-English explanations for any concept in programming, math, or science. Breaks down complex topics into clear, beginner-friendly language with examples. Built for BitWithBite students.

Concept breakdownWorked examplesBeginner-first phrasing

HTMLJavaScriptPHPLLM API

🚀 Live Tool

● Live on this site

📝

AILive

AI Assessment Checker

Scores a submitted answer and explains what was right and what to fix.

Submit any answer, coding, math, or science, and get instant AI-powered feedback with a score, what you got right, what needs improvement, and tips to improve. Free tool on BitWithBite.

Automatic scoringTargeted feedbackImprovement tips

HTMLJavaScriptPHPLLM API

🚀 Live Tool

● Live on this site

🧮

AILive

AI Math & Physics Solver

Step-by-step worked solutions for algebra, calculus and mechanics.

Solve math and physics problems step-by-step with detailed worked solutions and explanations. Supports algebra, calculus, mechanics, and more, ideal for students who want to understand the method, not just the answer.

Step-by-step workingAlgebra & calculusMechanics

HTMLJavaScriptPHPLLM API

🚀 Live Tool

● Live on this site

🧠

AILive

AI Adaptive Quiz Engine

Adaptive assessment with a learner model, spaced repetition and concept diagnosis.

Advanced adaptive quiz platform with 8 question types, Bloom's Taxonomy levels, personalised learner brain model, spaced repetition, survival mode, and real-time concept diagnosis. Learn any topic, programming, science, history.

8 question typesBloom's taxonomy levelsSpaced repetitionConcept diagnosis

HTMLJavaScriptPHPAdaptive AISpaced Repetition

🚀 Live Tool

● Live on this site

🎓

AILive

AI Tutor Pro

On-demand tutoring with guided problem-solving rather than plain answers.

Personalised AI tutoring for programming, math, and science. Ask anything and receive expert-level explanations, guided problem-solving, and curriculum-aligned support, available 24/7 for free on BitWithBite.

Guided problem-solvingCurriculum-alignedMulti-subject

HTMLJavaScriptPHPLLM API

🚀 Live Tool

● Live on this site

📖

AILive

AI Homework Helper Pro

Explains each step of a solution so the method transfers, not just the answer.

Step-by-step AI homework assistance for programming, math, and science. Explains each step of the solution so students actually learn rather than just copy. Integrated into the BitWithBite platform.

Step explanationsMulti-subjectLearning-first design

HTMLJavaScriptPHPLLM API

🚀 Live Tool

● Live on this site

⚡

AIWeb

BitWithBite Coding Challenges v1

Auto-judged coding challenges with progressive hints and an AI mentor.

AI-powered coding challenges platform with auto-judge, XP system, badges, streaks, 3-level progressive hints, AI Mentor, AI Code Review, and leaderboard. Embeddable via iframe or popup.

Auto-judge3-level hintsAI mentorLeaderboard

PythonFastAPI OllamaXP/Badges

● Client project · delivered

📚

AIPHP

AI Study Assistant

PHP study assistant with accounts, saved preferences and rate limiting.

PHP-based AI study assistant with MySQL backend, user preferences, rate limiting, and authentication. Helps students get instant explanations, practice questions, and study guidance.

User accountsSaved preferencesRate limitingMySQL backend

PHPMySQL AuthRate Limiting

● Client project · delivered

💬

AIPHPChatbot

AI FAQ Bot

Answers common questions instantly and escalates when confidence is low.

Smart FAQ chatbot with keyword matching, confidence scoring, and LLM fallback. Answers common questions instantly, shows a colour-coded confidence bar, and escalates to email when uncertain. Deployable on any PHP hosting.

Keyword matchingConfidence scoringEmail escalation

PHPJavaScriptKeyword NLP Keyword MatchingEmail Escalation

🚀 Live Demo

● Live on this site

🤖

AIPHPChatbot

AI Website Chatbot

Site assistant that detects intent, captures leads and recommends services.

Full-featured AI chatbot assistant for websites: intent detection, lead capture (saves name + email to JSON), product & service recommendations, Markdown rendering, and embeddable widget mode. Built for businesses and freelancers.

Intent detectionLead captureRecommendationsEmbeddable

PHPJavaScriptKeyword NLP Intent DetectionLead CaptureCRM

🚀 Live Demo

● Live on this site

🎧

AIPHPChatbot

AI Customer Support Bot

Support agent that raises tickets and hands off to a human on negative sentiment.

Production-ready AI customer support agent with ticket creation (saved to JSON), real-time sentiment detection, automatic human handoff on negative sentiment, and email notification. Handles order tracking, refunds, password resets, and more.

Ticket creationSentiment detectionHuman handoffEmail notification

PHPJavaScriptKeyword NLP Ticket SystemSentiment NLPHuman Handoff

🚀 Live Demo

● Live on this site

🏢

AIRAGChatbot

AI Internal Company Assistant

Internal knowledge assistant with document upload and role-based access.

Enterprise-grade internal AI assistant with document upload (PDF/TXT/MD), RAG-style keyword retrieval, source citations, PIN-based role access control (Admin/HR/Dev), and answers grounded directly in uploaded company documents.

Document uploadCited retrievalRole-based accessPIN gating

PHPJavaScriptKeyword NLP RAGDocument UploadRBAC

🚀 Live Demo

● Live on this site

🔎

PHPVision AI

Image Analyser (PHP GD)

Extracts dimensions, dominant colour, brightness and EXIF from an upload.

Upload any image and extract dimensions, dominant colours, brightness analysis, aspect ratio, and EXIF data using PHP GD. Replaces TensorFlow/Keras with pure PHP, zero API cost, zero ML setup, runs on any shared hosting.

Colour analysisBrightness metricsEXIF extractionZero dependencies

PHPGD Library Base64 EncodingcPanel

🚀 Live Demo

● Live on this site

📄

AIFastAPI

AI Resume Analyzer, cPanel Edition

The resume analyser rebuilt to run on shared hosting with no model server.

Full-stack resume analyser with ATS scoring, skill gap analysis, cover letter generator, 8-week learning roadmap, and salary prediction. Converted from Ollama + sentence-transformers to pure Python TF-IDF scoring + template-based responses, zero API cost, deployable on cPanel via Phusion Passenger.

ATS scoringSkill-gap analysisCover letterSalary estimate

FastAPIKeyword NLP PythonpdfplumbercPanel

● Client project · delivered

🤖

AIRAG

BitWithBite AI RAG Assistant, cPanel Edition

The RAG assistant rebuilt for shared hosting using keyword retrieval.

Production RAG chatbot for BitWithBite that answers customer questions from a live knowledge base. Converted from FAISS + Ollama to keyword-based retrieval, returns matching document excerpts directly. Supports PDF/DOCX upload, conversation history, gap logging, and cPanel deployment via Phusion Passenger.

Knowledge retrievalSource citationsShared-hosting compatible

FastAPIKeyword NLP RAGPythoncPanel

🚀 Live Demo

● Live on this site

❤️

UpworkNonprofit

Charity Website Development | Fundraising Platform for Nonprofits

Donation platform for a children's education nonprofit.

Developed a dedicated charity website to support children in need through a nonprofit organisation. Designed an intuitive, secure donation platform to facilitate contributions. Integrated interactive features to encourage community involvement and volunteer sign-ups. Automated donation tracking and reporting to optimise campaign management and transparency.

Secure donationsCampaign pagesResponsive build

PythonWeb Development API IntegrationAutomationResponsive Design

● Client project · delivered

May 2024 · Upwork Project

🛍️

E-CommerceFull UI

BigShop, Full E-Commerce Web Platform

Complete storefront UI: catalogue, cart, checkout, wishlist and account.

Complete multi-page e-commerce platform UI with product catalogue, shopping cart, checkout flow, wishlist, order history, user account management, blog, and transactional email templates. Fully responsive across all devices.

Product catalogueCart & checkoutOrder historyEmail templates

HTML5CSS3 BootstrapjQueryOwl Carousel

● Client project · delivered

Web Development · Full-Stack UI

🎓

EdTechFull UI

E-Educators, Online Learning & Course Marketplace

Course marketplace UI with browsing, cart, checkout and order tracking.

Full-featured online education platform UI covering course listings, category browsing, shopping cart, checkout, my account, order tracking, blog, and contact, designed for scalable e-learning delivery.

Course listingsCart & checkoutOrder tracking

HTML5CSS3 BootstrapjQueryResponsive Design

● Client project · delivered

Web Development · EdTech Platform

❤️

NonprofitPayment API

Donate for Kids, Payment-Integrated Charity Site

Charity site with a live payment gateway accepting real donations.

Charity donation website for children's education with integrated payment gateway. Features donation page, about, and contact sections, built to accept real online donations and raise funds for kids affected by the pandemic.

Payment gatewayDonation flowResponsive build

HTML5CSS3 JavaScriptPayment GatewayFont Awesome

● Client project · delivered

Web Development · Charity Platform

🌍

CharityRazorpay

E-Educators Foundation, Fundraising & Donation Platform

Fundraising platform with live Razorpay integration and campaign pages.

Full-stack charity fundraising website for "The E-Educators Foundation" with live Razorpay payment integration, campaign pages, donation CTA, and a responsive layout, focused on education support for post-pandemic children.

Razorpay integrationCampaign pagesDonation CTA

HTML5CSS3 JavaScriptRazorpay APIPoppins UI

● Client project · delivered

Web Development · Fundraising Platform

📚

EdTechResponsive

Responsive Online Study Website

Responsive study platform with course listings and teacher profiles.

Clean, fully responsive online study platform with course listings, teacher profiles, login modal, and contact section. Built with modern CSS and smooth UX, demonstrates front-end design and mobile-first development skills.

Course listingsTeacher profilesMobile-first layout

HTML5CSS3 JavaScriptFont AwesomeGoogle Fonts

● Client project · delivered

Web Development · Front-End Design

Research Projects, Actually Run

Three research projects run end to end on a CPU-only laptop. Every number below came from the actual run, not from memory. Where a result is weak, a method has a flaw, or a figure comes from a stand-in rather than the real thing, I say so instead of leaving it out.

🧱

Fine Crack Detection, U-Net Segmentation

Pixel-level crack segmentation for structural inspection · trained from scratch

✔ TRAINED & EVALUATED

A U-Net trained to segment hairline cracks at the pixel level, the fine-structure problem where most detectors fail, because a crack can be one or two pixels wide against a noisy surface. Trained on a synthetic dataset of 101 labelled image/mask pairs, evaluated on a held-out test split the model never saw during training.

0.785Dice coefficient

0.658IoU

0.763Precision

0.858Recall

0.808F1 score

1.89MParameters

Held-out test predictionsInput · ground truth · prediction, on images excluded from training

Training curveLoss and test Dice/IoU across all 20 epochs

Run configuration: 101 image/mask pairs split 80 train / 21 test · U-Net (3 encoder blocks, 1,885,697 parameters) · combined BCE + Dice loss weighted 0.3/0.7 per the project's own config · Adam · flip augmentation · images downscaled 512→192px · 20 epochs in 36 minutes on an Intel i5-8350U CPU, no GPU · seed fixed at 42 for reproducibility.

Honest reading of the result: recall (0.858) exceeds precision (0.763), the model finds most real cracks but over-traces slightly. For structural inspection that is the right direction to err, since a missed crack costs far more than a flagged false positive. The visible gap between prediction and ground truth is partly an artefact of the labels: the ground-truth masks break faint cracks into dashes, while the model connects them. A higher score would come from full 512px resolution and longer training, both GPU-bound, not method-bound.

🚗

Real-Time Vehicle Detection & Tracking

Motion-based detection with multi-object ID tracking · run on real traffic footage

✔ EXECUTED ON REAL FOOTAGE

A classical computer vision pipeline: MOG2 background subtraction isolates moving vehicles against a static camera, contour filtering removes noise, and a Euclidean-distance tracker assigns and maintains an identity for each vehicle across frames. Run headlessly over 900 frames of real 1280×720 traffic footage.

900Frames processed

1280×720Source resolution

30 fpsFrame rate

715Detections made

3Max simultaneous

0Training required

Detection & ID assignmentBounded vehicles with persistent IDs inside the ROI

Foreground maskThe MOG2 motion signal behind each detection

Re-run, Aug 2026Frame 265 · tracking 2 · 25 IDs assigned so far

Run configuration: MOG2 background subtractor (history 100, varThreshold 40) · minimum contour area 1400px · fixed region of interest · Euclidean-distance tracker with a 25px association radius · no training, no model weights, no GPU, runs in real time on CPU.

Re-run on 9 August 2026: I ran the pipeline again straight from the parameters committed in main.py, meaning MOG2 with history 100 and varThreshold 40, the same fixed region of interest and the same 1400px area filter. It reproduced exactly 73 IDs across the first 900 frames. Run over the full clip it handles 9,184 frames and assigns 595 IDs at 141 fps on CPU, comfortably ahead of the 30fps source. So the figures below are reproducible rather than remembered.

A limitation worth naming: the tracker assigned 73 IDs across 900 frames, which considerably overstates the true vehicle count. The association radius is fixed at 25 pixels, so when a vehicle moves faster than that between frames, its centroid jumps too far and the tracker issues a fresh ID. This is a real weakness of naive centroid tracking and the reason production systems use Kalman filtering or IoU-based association such as SORT/DeepSORT. I'm reporting the ID count rather than presenting it as a vehicle count, because the two are not the same thing here.

🏗️

AI-CAD, Sketch-to-Energy Building Design

Seven-stage generative design pipeline · rebuilt from a stalled client handover

✔ RUN END TO END

You draw a building; the system tells you what it would cost to run. It finds similar reference CAD drawings, estimates depth from the closest match, reconstructs a 3D solid from that single view, voxelises it into a storey-by-storey model, writes an EnergyPlus thermal model and ranks the options on a Pareto front. Every stage has both a research backend and a dependency-free fallback, so the whole thing runs on a laptop with no GPU and no paid APIs. That is how the numbers below were produced.

0.934Precision@1

0.613mAP

2.6sPer candidate (CPU)

5/5Sketches completed

4/5Pareto-optimal

28/28Self-tests passing

Inferred depthSilhouette relief from a single CAD drawing

Voxelised massing6 storeys · 1,251 m² GFA · 3,946 m³

Pareto frontSimilarity vs energy, the actual trade-off

Energy by end useHeating · cooling · lighting and equipment

Run configuration: 288-image, 12-category CAD reference database · retrieval benchmarked leave-one-out (query each image, relevant = same category) · 8 candidates retrieved with MMR diversification, top 5 carried through · voxel grid 20 at 1m pitch · 20×20×20m target massing · London degree-day climate preset · 13.1 seconds for all five candidates on CPU, no GPU · fixed config hash recorded in the run manifest.

The validation gate paid for itself: 3 of the 5 reconstructions came out of the shape stage with non-manifold edges and a stray second component. Earlier versions just logged a warning and moved on. Voxelisation closes almost any surface, so those broken meshes would still have produced floor areas that looked perfectly reasonable. The gate now repairs them, checks again, and rejects whatever it cannot fix.

What these numbers are, and what they are not: the retrieval metrics come from a synthetic stand-in dataset built to match the structure of the client's original 3,341 drawings, which were never shared. They measure the descriptor, not performance on the real corpus. The energy figures come from the screening model rather than an executed EnergyPlus run. The IDF generator, runner and results parser are all finished and I checked the generated model object by object, but no EnergyPlus install was available on this machine. The tool says so itself: energy stays labelled a "screening estimate" until a real simulation replaces it.

Read the full case study →   covering the original failure, the four faults behind it, and the bug that made a hollow shell look like a solid building.

📁 The rest of the research archive

What else is in the folder, what state it is actually in, and where it can't be shown running here. Updated August 2026 after a recovery pass took most of these from "untested" to "verified" — with the two corrections that pass forced noted in place.

Research & Publications

Peer-reviewed computational biology work, a deep learning master's thesis, and ongoing applied AI research, the through-line is using machine learning to solve problems that are expensive or impossible to brute-force.

🔎 Research Interests

📄 Peer-Reviewed Publication

2023 · Peer-Reviewed Journal

Design and Assessment of a Novel In Silico Approach for Developing a Next-Generation Multi-Epitope Universal Vaccine Targeting Coronaviruses

Rasheed MA, Raza S, Alonazi WB, Ashraf MA, Navid MT, Aslam I, Iqbal MN, Rahman SU, Riaz MI.

Microorganisms, 2023, 11(9):2282 · Impact Factor: 4.5 · DOI: 10.3390/microorganisms11092282

Computational Biology Vaccine Design Immunoinformatics Coronavirus IF 4.5

My Role: Experimentation, participated in conducting experiments, contributing directly to the research process and data generation. Assisted in drafting and revising the manuscript for publication.

🔗 View on DOI 📄 Download PDF 📋 Copy Citation 🧾 Copy BibTeX

📊 Results & Figures from the Paper

The full methodology and result figures as published in Microorganisms (MDPI, open access, CC BY). Click any figure to enlarge.

Figure 1Methodology flowchart, the full in-silico prediction pipeline

Figure 2Vaccine construct, epitopes and adjuvant joined into one peptide

Figure 3Predicted 3D structure & quality validation (I-TASSER)

Figure 4Molecular docking, vaccine construct against human TLR2

Figure 5Immune simulation, antibody titres & T-cell response over time

Figure 6Molecular dynamics, stability of the vaccine–TLR2 complex

Figures reproduced from the authors' own open-access article (Microorganisms 2023, 11(9):2282), published under the Creative Commons Attribution (CC BY 4.0) licence.

🔬 Experimentation📊 Data Generation✍️ Manuscript Drafting 🔁 Manuscript Revision🧬 Immunoinformatics📚 Literature Review

🎓 Master's Thesis

🧭 Research Timeline

Professional Journey

AI Tools Running in Production

Nineteen AI tools built and deployed on BitWithBite, most of them in PHP, running on standard shared hosting with local models and no paid API dependency. These are not demos. They are live, in use, and cost nothing per request to run.

Academic Background

Courses & Certificates

Preply Student Reviews

A selection of verified student feedback from Preply. The full rating history is public on the Preply profile linked below.

Freelance & Technical Work

Independent client work across AI/ML, computer vision, Python development, data analysis and technical writing. A selection of engagements below.

28 Browser Learning Games

A full games engine built for BitWithBite, every game runs in the browser with no install, no plugins and no backend dependency. Built to make practice feel like play, with XP, streaks and leaderboards driving retention.

Awards & Scholarships

Let's Work Together

Available for freelance projects, full-time roles, and collaboration on AI/ML or web development projects.