Problem: Healthcare payers can't identify which members will become high-cost (>$30k/year) until the cost has already happened — too late for proactive care management.

Solution: A LightGBM + XGBoost ensemble on 336 engineered features (cost ratios, code diversity, HCC risk scores), with the decision threshold tuned to maximize recall — catching over 82% of true high-cost members before costs occur.

Result: Recall 0.825 · F1 0.563 · 🏅 Softtec 2026 ML Competition, FAST NUCES Lahore

Problem: Reading long research papers to find one fact is slow, and asking a general LLM instead risks a confident, hallucinated answer with no way to verify it.

Solution: A RAG pipeline — PDF split into page-aware chunks, embedded locally (zero-cost, no API calls), retrieved by cosine similarity, and answered by an LLM instructed to cite the exact page it used.

Result: Every answer is grounded in the actual document, with inline (p. 3)-style citations

Problem: Small e-commerce teams outgrow spreadsheets for inventory but can't justify the cost or complexity of a full ERP system.

Solution: A full-stack SaaS where every number that matters — forecasts, reorder quantities — is computed deterministically; an LLM is used only to narrate the trend, never to invent a figure. Real-time Socket.IO alerts, full RBAC.

Result: 42 backend tests, CI pipeline, and graceful degradation if Redis/Groq are unavailable

Problem: Students, applicants, and visitors couldn't get quick, role-relevant answers about University of Layyah — info was scattered across static pages nobody reads in full.

Solution: A bilingual (EN/UR) AI assistant grounded in real scraped university data, with role-based responses (student/applicant/visitor) and an offline keyword-matching fallback when the API is unreachable.

Result: 🏅 Top 10 out of 23+ teams, South Punjab Generative AI Hackathon 2026

Problem: Most people can't understand American Sign Language, creating a real communication barrier with no accessible, real-time translation tool at hand.

Solution: A CNN trained on class-balanced, augmented ASL image data (with horizontal flip intentionally disabled, since mirroring changes a sign's meaning), served through CLI, image-upload, and live webcam inference.

Result: Shared inference code across all three interfaces, unit tests + CI pipeline

Problem: Manual shelf audits are slow and easy to skip, so stockouts often go unnoticed until a customer complains or a sale is lost.

Solution: A real-time YOLOv8 detection pipeline that counts items per frame and triggers color-coded low-stock alerts — architected so the detection model can be swapped for a custom-trained, SKU-specific model with zero code changes.

Result: CLI-configurable (video/model/threshold), includes a ready-to-use custom fine-tuning script

Problem: Small event organizers need real ticketing infrastructure — booking, seat tracking, revenue — but existing options are either an expensive platform or a form bolted onto a calendar.

Solution: A multi-role (attendee/organizer/admin) booking platform with seat-locked bookings that can't overbook, JWT auth with httpOnly + CSRF-protected refresh cookies, and a revenue analytics panel.

Result: 11 passing pytest tests, one-command Docker setup (Postgres + Flask + Vite together)

Problem: Campus/office cafeterias running orders manually leads to mistakes and zero visibility into what's actually selling.

Solution: A full-stack ordering system with role-based access (admin/customer), bcrypt-hashed auth, a live order-tracking flow, and a revenue analytics dashboard.

Result: Coupons, verified reviews, and full cart-to-checkout flow

Problem: Small grocery businesses need an online storefront, but affordable e-commerce options rarely come with real security or delivery logistics built in.

Solution: A full-stack platform (42+ products, 12+ tables, 50+ endpoints) with bcrypt hashing, login-attempt lockout, delivery-slot scheduling, and an admin analytics dashboard.

Result: ~5,900 lines of code, transaction rollback support, stock validation before every order

Problem: Static portfolios make visitors dig through sections to find what they actually want to know about a candidate.

Solution: A portfolio with a real AI chat assistant (Groq-powered, grounded in my actual background) that answers any question conversationally and attaches project cards automatically when relevant.

Result: Runs as a single Vercel Edge Function — no separate backend to host

Problem: Most people avoid tracking personal finances because every tool demands a signup and a subscription just to log an expense.

Solution: A zero-backend, browser-based finance dashboard — full transaction CRUD, category budgets with color-coded alerts, and multi-currency support, all persisted locally.

Result: CSV/JSON export, print-ready reports, dark/light mode

Problem: Small service businesses (salons, clinics, gyms) need appointment scheduling without paying for a heavyweight SaaS subscription.

Solution: A single-file dashboard with a FullCalendar-powered visual schedule, complete appointment CRUD, and dynamic accent-color theming — drop the file anywhere and it runs.

Result: Honest in-memory data model, clearly documented rather than disguised as persistent

Problem: Small e-commerce businesses need a modern, engaging storefront to compete with bigger brands, but generic templates look dated and don't convert.

Solution: A React + Vite storefront with a glassmorphism UI, real-time search/category filtering, cart + wishlist management, and a persistent dark/light theme.

Result: Full checkout flow with order confirmation, cart/wishlist persisted via localStorage

Problem: Restaurant startups need a premium, trustworthy landing page to convert visitors into orders, without the cost of a full custom build.

Solution: A fully responsive landing page with live menu filtering, an animated FAQ accordion, cart-add simulation with toast feedback, and scroll-triggered animations via the Intersection Observer API.

Result: Sticky nav with active-section highlighting, smooth-scroll navigation throughout