Skip to content
CASE STUDIES

Six systems shipped.

Everything below is either live in production or, in SoleVault's case, a design-and-engineering showcase we built to set our own bar for interaction quality.

01[✓] LIVE IN PRODUCTION

CloudDesk

CloudDesk sorts and routes support tickets in under 2 seconds at 91% accuracy, using a fine-tuned NLP model rather than keyword rules. Built for high-volume support operations where manual triage is the actual bottleneck.

accuracy91%
latency<2s / ticket
intent categories7
response templates42
BUILT WITH
DistilBERTFastAPINext.jsGemini AI

Fine-tuned DistilBERT classifies each ticket into an intent category; a FastAPI service routes it to the right queue and drafts a first-pass reply for the agent to approve or edit.

Agents work from a kanban or list queue with per-ticket history and template-backed replies instead of blank-slate composition every time.

91% accuracy is measured against a labeled holdout set, not a best-case demo run.

02[✓] LIVE IN PRODUCTION

SwiftDocs

SwiftDocs turns unstructured invoices and contracts into clean, structured data at 98% accuracy, built for finance and ops teams where manual re-keying is the real cost center.

accuracy98%
saves~30hrs / week manual work
BUILT WITH
Next.jsOCR pipelineLLM extractionstructured export

Documents are parsed through an OCR + LLM pipeline and normalized into structured fields (vendor, line items, totals, dates) with a review step before export.

A history view keeps every processed document queryable, so a team can audit what was extracted and when.

98% field-level accuracy is measured against manually verified documents, not the easiest cases.

03[✓] LIVE IN PRODUCTION

Sales Intelligence

Sales Intelligence forecasts revenue and segments customers into AI-driven behavioral personas across 1M+ rows of transactional data, built for teams moving from gut-feel targeting to data-driven prioritization.

processes1M+ rows
total revenue trackedPKR 1.60bn
orders analyzed395,806
querynatural language to SQL
BUILT WITH
Next.jsSQL warehouseforecasting modelsLLM query layer

An executive dashboard surfaces revenue trend, order status, and category performance at a glance, backed by a real transactional dataset, not seeded demo data.

A forecasting module projects revenue forward from trend; a segmentation model clusters customers into behavioral personas the sales team can act on directly.

The 'Ask Your Data' layer translates plain-English questions into SQL, so a non-technical teammate can query the warehouse an analyst would.

04[✓] LIVE IN PRODUCTION

Daraz Intelligence

Daraz Intelligence is a real-time market analysis tool for e-commerce decision-making. It tracks pricing and listing behavior continuously and flags sellers manufacturing fake discounts.

tracking600 products/day
caught47 sellers faking discounts
BUILT WITH
scraping pipelineprice-history engineanomaly detectionNext.js dashboard

A continuous pipeline tracks listing prices per product over time, building a real price history instead of a single snapshot.

The fake-discount detector compares a listing's 'was' price against its tracked history, flagging discounts that were never real to begin with.

An AI insights layer surfaces drop patterns and anomalies across the catalog for a pricing team to act on.

05[✓] LIVE IN PRODUCTION

HOS Trip Planner

A full-stack Hours-of-Service compliance and trip-planning tool for trucking. The core engine simulates duty status hour-by-hour, advancing only as far as the nearest of five independent regulatory thresholds, never applying rules after the fact.

regulatory thresholds simulated5
rolling compliance window8-day / 70-hour
log formatFMCSA-compliant SVG
exportmulti-page PDF
BUILT WITH
Django REST FrameworkReact + ViteGeoapify APIRender + Vercel

generate_hos_schedule() advances to whichever of five thresholds is nearest: the 11-hour driving limit, 14-hour duty window, 30-minute break trigger, a mileage-based fuel stop, and the 70-hour/8-day cycle limit.

The 70-hour rule runs on a rolling 8-day sum seeded with real prior-day values, so hours age out of the window exactly as the regulation intends.

Every event carries real coordinates sampled along the route via the Geoapify API (server-side only), so the map renders a distinct icon per stop type instead of generic pins.

The frontend renders an actual FMCSA-format daily log as an SVG grid (duty rows, quarter-hour ticks, plotted status line, per-category totals summing to 24), exportable as a multi-page PDF, alongside a 7-day cycle-hours form with proportional redistribution.

Backend is deliberately stateless (a fresh schedule per request), deployed separately from the frontend (Render/Vercel) and connected via an environment-driven API URL, not a hardcoded host.

06[✓] DESIGN & ENGINEERING SHOWCASE

SoleVault

SoleVault is a concept sneaker e-commerce build: our design-and-engineering showcase, not a client product. Built in Next.js 16 with Tailwind CSS v4. Every interaction below is real code: no third-party animation platforms, no pre-baked video pretending to be interactive.

hero sequence175 frames, scrubbed live
products individually anchored10
checkout flow3 steps, self-drawing progress
reduced-motion coverage100% of custom animation
BUILT WITH
Next.js 16Tailwind CSS v4Framer Motionanime.jsraw requestAnimationFrameCSS 3D transforms

A scroll-scrubbed product disassembly hero: a 175-frame sequence scrubbed live via video.currentTime against a hand-built smooth-scroll layer with a noise deadzone and a throttle that measures each visitor's actual display refresh rate.

A drag-to-rotate 360° turntable with real momentum on release, idle auto-rotation resuming from the exact position you left it (solved via a raised-cosine curve), a cursor-tracked glare, and a counter-drifting shadow for depth.

A pinned scroll-driven product story with tethered callouts individually anchored to pixel positions across ten products' actual photography, nothing generic.

FLIP-style grid reflow when filtering products, so surviving cards glide to their new position instead of hard-refreshing.

Cart and checkout with real physicality: the product photo flies into the cart icon with a comet-tail, a genuine three-step checkout with self-drawing progress, and a fully custom CSS-3D packaging animation on confirmation.

Framer Motion for anything tied to React's render cycle, anime.js for explicit timeline sequencing and SVG stroke-drawing, and hand-rolled requestAnimationFrame loops for scroll/drag physics, kept outside both libraries. Every custom animation ships a genuine prefers-reduced-motion fallback.