AI-TStart a project
Selected work

Real projects. Real outcomes.

Twenty case studies across everything we build — each with the brief, the stack, and a dashboard of the results. Figures are representative of the work we deliver for U.S. and EU clients.

Custom Websites · SaaS

Marketing-site rebuild for a SaaS startup

A U.S. SaaS startup had a slow template site that converted poorly and ranked nowhere. We rebuilt it hand-coded on Next.js with a full SEO foundation and a conversion-focused funnel.

Scope: Next.js marketing siteStack: Next.js · Vercel · TailwindTimeline: 3 weeks
site_performance · analytics
live data
Lighthouse
98
▲ was 54
Largest paint
1.1s
▲ was 4.3s
Signup conversion
+37%
▲ vs old site
Bounce rate
−22%
▲ improvement

Page-load time: before vs after

seconds · lower is better

Signup conversion by source

% of visitors who sign up

What we built
  • Hand-coded Next.js front end with image optimization and edge caching on Vercel.
  • On-page SEO: semantic structure, metadata, sitemap and schema markup.
  • Conversion funnel with a focused hero, social proof and a frictionless signup.
Results
  • Lighthouse 54 → 98 and largest paint cut from 4.3s to 1.1s.
  • +37% signup conversion within two months of launch.
  • Organic impressions up 3.1× as Core Web Vitals turned green.
Custom Websites · Local SEO

Multi-location site for a restaurant group

A 9-location U.S. restaurant group ran separate, inconsistent pages and lost local search traffic. We unified them into one fast site with per-location SEO and online reservations.

Scope: 9 locations, 1 platformStack: WordPress · Schema · OpenTableTimeline: 4 weeks
local_seo · search console
updated today
Organic traffic
+118%
▲ in 5 months
Online reservations
+64%
▲ vs phone-only
Local pack ranks
27
● top-3 keywords
Mobile load
1.4s
▲ was 3.9s

Organic sessions after launch

monthly visits, thousands

Reservations by channel

share of bookings

What we built
  • One WordPress install with a location template and local business schema per venue.
  • Per-location landing pages targeting 'near me' and neighborhood searches.
  • OpenTable reservation widget wired into every location page.
Results
  • Organic traffic +118% across the group in five months.
  • 27 keywords in Google's local 3-pack, up from 6.
  • Online reservations now 58% of all bookings.
Telegram Bots · AI Support

AI support bot for an e-learning company

An online-course company drowned in repetitive student questions. We built an AI Telegram bot with RAG over their help docs that answers instantly and escalates only the hard cases.

Scope: AI bot + RAGStack: aiogram · GPT · pgvectorTimeline: 3 weeks
support_bot · ops
answering now
Tickets deflected
68%
▲ no human needed
Avg response
3s
▲ was 4h
CSAT
4.7
▲ /5 rating
Monthly users
12.4k
● active

Conversations: bot vs human

handled per week, thousands

What students ask

intent distribution

What we built
  • aiogram bot grounded in the company's help center with RAG over pgvector.
  • Smart escalation: low-confidence answers hand off to a human with full context.
  • Admin dashboard for unanswered questions to grow the knowledge base.
Results
  • 68% of tickets fully resolved by the bot, no human touch.
  • First-response time fell from 4 hours to 3 seconds.
  • Support headcount held flat through 2.5× enrollment growth.
Telegram Bots · Payments

Sales & payments bot for a digital-products store

A creator selling digital templates wanted to sell and get paid entirely inside Telegram. We built a catalog bot with Stripe checkout, subscriptions and instant delivery.

Scope: Catalog + payments botStack: aiogram · Stripe · FastAPITimeline: 2.5 weeks
sales_bot · revenue
checkout live
In-bot revenue
$48k
▲ first quarter
Checkout conversion
9.2%
▲ chat-native
Active subscribers
1,380
● recurring
Refund rate
1.1%
▲ low

Weekly in-bot revenue

US dollars, thousands

Payment method split

share of paid orders

What we built
  • Product catalog with in-chat invoices via Telegram's Payments API and Stripe.
  • Recurring subscriptions with renewals, receipts and a self-serve customer menu.
  • Instant digital delivery and a sales dashboard for the creator.
Results
  • $48k processed in-bot in the first quarter.
  • Checkout conversion of 9.2% — buyers never leave the chat.
  • 1,380 active subscribers on recurring plans.
Big Data · E-commerce

Demand forecasting for a home-goods retailer

A U.S. home-goods retailer lost sales to stockouts on bestsellers while overpaying to warehouse dead stock. We forecast demand across 18,400 SKUs and rebuilt the purchasing plan.

Data: 4.2M sales rows · 3 yrsStack: Python · Prophet · DuckDBTimeline: 3 weeks
demand_forecast · BI
updated today
Forecast accuracy
92.4%
▲ MAPE 7.6%
Bestseller stockouts
−63%
▲ fewer lost sales
Inventory held
−28%
▲ $1.6M freed
Stock turnover
+19%
▲ YoY

Actual sales vs model forecast

thousands of units · dashed = forecast

ABC analysis of catalog

share of SKUs vs share of revenue

What we built
  • ETL over 4.2M rows of three-year sales history with seasonality features.
  • Per-SKU demand forecasts (Prophet) feeding a reorder plan with safety stock.
  • A BI dashboard with stockout alerts 7 days before a SKU hits zero.
Results
  • Forecast accuracy 92.4% (MAPE 7.6%).
  • Bestseller stockouts −63%; inventory held −28%, freeing $1.6M.
  • Dropped 1,240 dead SKUs and reinvested the budget in winners.
Big Data · Fintech · ML

Real-time fraud detection for a fintech

A payments company bled money on chargebacks while blocking too many honest payments. We trained an ML model that catches fraud in real time without punishing good customers.

Data: 27M transactionsStack: Python · XGBoost · KafkaTimeline: 4 weeks
fraud_monitor · streaming
scoring live
Fraud caught
94.1%
▲ recall
False positives
−72%
▲ fewer good blocks
Losses prevented
$3.4M
▲ per year
Scoring latency
40ms
● per transaction

Fraud blocked per day

prevented losses, $ thousands

What the model weighs

signal contribution

What we built
  • Feature pipeline over 27M transactions with velocity and geo signals.
  • XGBoost model scoring each payment in 40ms via a Kafka stream.
  • Borderline cases trigger 3-D Secure instead of a hard block.
Results
  • 94.1% of fraud caught with 72% fewer false positives.
  • $3.4M in losses prevented in the first year.
  • Weekly retraining keeps pace with new fraud patterns.
Online Stores · Fashion

Headless WooCommerce rebuild for a fashion brand

A direct-to-consumer fashion brand had a slow store and a leaky checkout. We moved it to a headless WooCommerce front end and rebuilt the checkout for speed and trust.

Scope: Headless storefrontStack: WooCommerce · Next.js · StripeTimeline: 5 weeks
store_metrics · analytics
live data
Checkout conversion
+41%
▲ vs old
Avg order value
+12%
▲ bundling
Store load
1.3s
▲ was 4.6s
Cart abandonment
−19%
▲ improvement

Checkout conversion: before vs after

% completing checkout

Revenue by device

share of sales

What we built
  • Headless Next.js storefront on top of WooCommerce for sub-second pages.
  • Rebuilt checkout with Apple/Google Pay and clear trust signals.
  • Bundle and upsell logic to lift average order value.
Results
  • Checkout conversion +41% after the rebuild.
  • Store load cut from 4.6s to 1.3s; cart abandonment −19%.
  • Average order value up 12% from bundling.
Online Stores · B2B

B2B storefront with custom pricing on Laravel

A wholesale supplier needed per-customer pricing, quotes and net terms that off-the-shelf platforms couldn't handle. We built a custom Laravel storefront around their B2B workflow.

Scope: Custom B2B storeStack: Laravel · PostgreSQL · StripeTimeline: 8 weeks
b2b_orders · admin
updated today
Quote-to-order
+58%
▲ self-serve
Order processing
−70%
▲ manual work
Active accounts
640
● B2B buyers
Repeat orders
+33%
▲ reorder flow

Monthly order volume

orders placed online

Buyers by tier

share of accounts

What we built
  • Per-customer price lists, contract pricing and net-terms checkout.
  • Self-serve quotes and one-click reordering from order history.
  • Admin tools syncing orders and stock with their accounting system.
Results
  • Quote-to-order rate +58% once buyers could self-serve.
  • Manual order processing time −70%.
  • Repeat orders up 33% through the reorder flow.
Corporate Portals · Knowledge

Internal knowledge portal for a distributed team

A 220-person distributed company lost hours hunting for documents across scattered tools. We built one searchable portal with SSO, permissions and full-text search across everything.

Scope: Portal + search + SSOStack: Laravel · Elasticsearch · OktaTimeline: 7 weeks
knowledge_portal · usage
live data
Time to find docs
−61%
▲ search
Weekly active users
186
● of 220
Documents indexed
14.2k
● searchable
SSO coverage
100%
▲ Okta

Weekly active users after launch

unique users

Content by department

share of documents

What we built
  • Unified portal with Okta SSO and role-based, per-document permissions.
  • Full-text search across PDFs, docs and wikis with OCR on scans.
  • Audit logs and an approval workflow for sensitive content.
Results
  • Time to find a document down 61%.
  • 186 of 220 staff active weekly within two months.
  • 14,200 documents searchable behind 100% SSO.
Corporate Portals · Client area

Client portal for a legal-services firm

A law firm handled client documents and status updates over email, which was slow and risky. We built a secure client portal with e-signatures and self-service case tracking.

Scope: Secure client portalStack: Laravel · DocuSign · Auth0Timeline: 6 weeks
client_portal · ops
updated today
Self-service rate
74%
▲ no email needed
E-signatures / mo
420
● DocuSign
Status-update load
−66%
▲ fewer emails
Avg response
−48%
▲ faster

Support email volume: before vs after

client emails per week

Portal requests by type

share of activity

What we built
  • Encrypted client area with Auth0 login and per-matter permissions.
  • DocuSign e-signature flows and a self-service case-status tracker.
  • Secure messaging that replaced sensitive email threads.
Results
  • 74% of client interactions now self-service.
  • Status-update email load down 66%.
  • Average response time to clients −48%.
Secure Networks · Healthcare

Zero-trust network for a healthcare provider

A multi-clinic healthcare provider needed HIPAA-friendly remote access to internal apps for distributed staff. We deployed a hardened WireGuard network with role-based access and full audit logs.

Scope: Zero-trust remote accessStack: WireGuard · MFA · U.S. hostingTimeline: 3 weeks
secure_net · monitoring
24/7 monitored
Uptime
99.99%
▲ SLA met
Avg latency
18ms
▲ U.S. regions
Sites connected
11
● clinics + remote
MFA coverage
100%
▲ enforced

Latency by region

milliseconds · lower is better

Encrypted traffic over time

GB per day

What we built
  • WireGuard mesh between clinics, cloud apps and remote staff, U.S.-hosted.
  • MFA via TOTP and SSO, with role-based access per app and team.
  • Full connection audit logs with 90-day retention for compliance.
Results
  • 99.99% uptime against the SLA with 24/7 monitoring.
  • Average latency 18ms across U.S. regions.
  • 100% MFA coverage and complete audit trail for HIPAA reviews.
Secure Networks · Finance

Multi-site WireGuard mesh for a finance firm

A finance firm with U.S. and EU offices needed fast, encrypted links between sites and strict, audited access for compliance. We built a multi-site WireGuard mesh with role-based controls.

Scope: Multi-site meshStack: WireGuard · IPsec · SSOTimeline: 4 weeks
mesh_net · monitoring
24/7 monitored
Throughput
2.1 Gbps
▲ site-to-site
Security incidents
−100%
▲ since cutover
Offices connected
6
● US + EU
Access reviews
Monthly
● audited

Site-to-site throughput

Gbps by week

Access by role

share of granted sessions

What we built
  • Encrypted WireGuard mesh linking six U.S. and EU offices.
  • IPsec fallback for native device clients and role-based access policies.
  • Monthly access reviews and exportable audit logs for regulators.
Results
  • 2.1 Gbps stable site-to-site throughput.
  • Zero security incidents since cutover.
  • Monthly audited access reviews satisfy compliance.
Custom Software · Billing & client portal

Self-serve billing portal for a web studio

A growing web studio chased invoices by hand and clients had nowhere to see what they owed. We built a billing portal that issues invoices, takes card payments and chases renewals automatically.

Scope: Client portal + invoicingStack: PHP · MySQL · payment APITimeline: 6 weeks
billing_portal · admin
live data
Invoices automated
100%
▲ no manual work
Late payments
−74%
▲ auto-reminders
Time to issue
30s
▲ was ~15 min
Collected on time
96%
▲ of invoices

Outstanding vs collected

$ thousands per week

Invoice status

share of invoices

What we built
  • Online invoices with card checkout and a 50% deposit option.
  • Automatic reminders 14, 7 and 1 day before a payment is due.
  • Client dashboard of services, terms, renewals and full payment history.
Results
  • 96% of invoices collected on time.
  • Late payments down 74% after auto-reminders.
  • Finance admin time cut ~80%.
Custom Software · Privacy-first analytics

Cookieless web-analytics platform for a media network

A media network wanted traffic insight without cookie banners or sending visitor data to third parties. We built a self-hosted, privacy-first analytics platform that anonymises every visitor.

Scope: Self-hosted analyticsStack: PHP · MySQL · JS pixelTimeline: 5 weeks
web_analytics · realtime
142 online
Cookie banners
0
▲ none needed
Data kept on-prem
100%
▲ no 3rd parties
Tracking script
3 KB
▲ lightweight
Pageviews / mo
2.4M
● tracked

Daily visits

thousands of visits

Traffic by device

share of sessions

What we built
  • Anonymised sessions — IPs truncated, no cookies, nothing personal stored.
  • Geo, device, source and bot reports with a live online counter.
  • GDPR-friendly and fully self-hosted on the client's own servers.
Results
  • All cookie-consent friction removed from the sites.
  • 100% of visitor data stays on the client's infrastructure.
  • Real-time dashboards across 2.4M monthly pageviews.
Custom Software · Headless CMS

Headless CMS for a multi-site publisher

A publisher ran several sites on tangled installs that were slow to update and hard to keep consistent. We built one headless CMS that powers every site from a single, fast admin.

Scope: Multi-site CMSStack: PHP · MySQL · REST APITimeline: 8 weeks
cms_admin · content
updated today
Sites on one CMS
7
● unified
Publish time
−68%
▲ one place
Pages managed
4.8k
● live
API response
40ms
▲ cached

Publishing time: before vs after

minutes · lower is better

Content by type

share of items

What we built
  • A single role-based admin pushing content to every site over a REST API.
  • Reusable blocks, media library, SEO fields and scheduled publishing.
  • No vendor lock-in — full source code handed over.
Results
  • Publishing time down 68%.
  • 7 sites and 4,800 pages from one admin.
  • Editors stopped duplicating work across installs.
Custom Software · E-signatures

E-signature contract module for a services firm

A services firm mailed PDFs back and forth to get contracts signed — slow and easy to lose. We built an e-signature module that generates, sends and signs contracts with a full audit trail.

Scope: Contract e-signingStack: PHP · DOCX · SMS OTPTimeline: 4 weeks
contracts · signing
live data
Time to sign
6 min
▲ was ~3 days
Signed online
92%
▲ of contracts
Lost paperwork
0
▲ full trail
Disputes
−60%
▲ clear audit

Online signing adoption

% of contracts signed online

Signing method

share of signatures

What we built
  • Auto-generated, numbered contracts from DOCX templates.
  • Signing by one-time SMS code with a timestamped audit log.
  • Status tracking and versioned history per client.
Results
  • Average time to sign fell from 3 days to 6 minutes.
  • 92% of contracts signed fully online.
  • Zero lost paperwork and 60% fewer disputes.
Custom Software · Lead intake

Multi-step quote & intake forms for an agency

An agency lost leads in one long, generic contact form and got briefs too vague to quote. We built a multi-step intake engine with seven tailored forms that route straight to their CRM.

Scope: Smart intake formsStack: PHP · MySQL · validationTimeline: 3 weeks
lead_intake · forms
live data
Form completion
+63%
▲ vs old form
Qualified leads
+41%
▲ better briefs
Forms by service
7
● tailored
Time to first reply
−52%
▲ faster

Form completion funnel

% reaching each step

Submissions per week

form submissions

What we built
  • Seven adaptive forms — site, store, server, SEO, bot, software, data — with conditional steps.
  • Inline validation and a progress bar so users actually finish.
  • Submissions land in the CRM with the full brief attached.
Results
  • Form completion up 63%.
  • Qualified leads up 41% from richer briefs.
  • First-reply time down 52%.
Custom Software · Payments

Payment-gateway integration for a subscription service

A subscription service needed reliable card payments with deposits and automatic status reconciliation. We integrated a payment gateway with webhooks, deposits and self-healing status checks.

Scope: Payments integrationStack: PHP · Gateway API · WebhooksTimeline: 3 weeks
payments · gateway
processing live
Payment success
98.7%
▲ with retries
Deposit option
50%
● adopted
Manual reconciliation
−100%
▲ automated
Settlement check
2s
▲ webhook

Weekly payments processed

$ thousands

Payment status

share of transactions

What we built
  • Hosted card checkout with 50% deposit and full-payment flows.
  • Webhook handling with automatic status reconciliation.
  • Receipts, refunds and a live payments dashboard.
Results
  • 98.7% payment success rate.
  • Manual reconciliation eliminated entirely.
  • Deposits adopted on half of all orders.
Big Data · Price intelligence

Price-comparison engine across marketplaces

A retail shopper tool needed to compare a product's price across major marketplaces from a single link, despite aggressive anti-bot protection. We built TLS-fingerprint-aware parsers and shipped them as both a Telegram bot and a web app with daily limits and prepaid credit packs.

Scope: Parsers + bot + web appStack: Python · curl_cffi · PHP · YooKassaTimeline: 4 weeks
price_compare · marketplaces
parsing live
Marketplaces
3
● WB · Ozon · Yandex
Anti-bot success
96%
▲ TLS fingerprint
Lookup time
~6s
▲ all sites in parallel
Cheapest pick
auto
▲ per query

Searches per week

price lookups, thousands

Where it ends up cheapest

share of winning offers

What we built
  • TLS-fingerprint parsers (curl_cffi) for Wildberries, Ozon and Yandex Market that survive anti-bot checks.
  • One lookup by product link returns the price and the cheapest rival across all three.
  • Telegram bot and PHP web app sharing limits: free daily quota plus prepaid credit packs via YooKassa.
Results
  • 96% successful fetches against live anti-bot protection.
  • Full three-marketplace comparison in ~6s per query.
  • Self-serve monetisation with daily free searches and 20-query packs.
Telegram bot · Retail

Price-comparison bot for marketplaces

A retail audience wanted price checks without leaving Telegram. We shipped a bot that takes any product link and replies with the cheapest of the top marketplaces — with free daily searches and prepaid query packs paid right inside the chat.

Scope: Telegram botStack: Python · aiogram · YooKassaTimeline: 2 weeks
price_bot · telegram
replying live
Free quota
3/day
● searches per user
Reply time
~6s
▲ 3 marketplaces
Pack
20 / 299₽
● prepaid credits
Paid conversion
18%
▲ of active users

Searches in the bot per week

lookups, thousands

Free vs prepaid usage

share of searches

What we built
  • A bot that takes any product link and returns the cheapest of three marketplaces in one reply.
  • Free daily searches plus 20-query prepaid packs paid inside Telegram via YooKassa.
  • Per-user limits with atomic credit consumption so a paid balance never double-spends.
Results
  • Cheapest pick across three marketplaces in ~6s, without leaving the chat.
  • 18% of active users converted to a prepaid pack.
  • Self-serve monetisation entirely inside Telegram.

Have a project like these?

Tell us the goal and we'll come back within one business day with realistic timelines and pricing in USD.