A video-avatar AI assistant for Hardline Nutrition that speaks Turkish 24/7, recognises returning customers by face, and sells across a 150-product catalogue with live price and stock.
A Sales Advisor That Never Closes and Knows Its Customers
Hardline Nutrition needed a customer experience that works at every hour and every level of demand: answering product questions, tracking orders and shipments, and recommending like a knowledgeable sales advisor — in natural Turkish.
We built Hardiye: a real-time talking AI assistant with a video avatar. The customer speaks; Hardiye listens, reasons over the live catalogue and answers by voice. It recognises returning customers by face without storing photos, searches products by meaning rather than exact names, scans QR codes and barcodes, and looks up shipments.
The Hard Part
A voice that feels human leaves no room for lag or misunderstanding. We built the line as a streaming problem, measured a sentence's mouth-to-ear journey step by step, and tracked the backend and avatar shares separately. A possibly misheard product name or phone number triggers a spoken confirmation; precise requests go to a structured query in milliseconds, vague ones to semantic search.
How We Built It
Real-time conversation and video avatar over WebRTC
Face recognition by embedding, without storing photos
Speech-to-speech tests with an AI playing the customer, plus latency telemetry
An automated release pipeline in Frankfurt; a separate product API for the client's IT team
The Result
Live since June 2026. Live price and stock across all 150 products, 155 green automated tests. The first greeting dropped from 5 seconds to 1.5, a repeated product query from 380 milliseconds to zero, and hybrid search answers on the server in 4 milliseconds.
Let's produce a similar outcome together
If you have a need like this reference project, let's talk.
A multi-tenant CRM/ERP for pilates, yoga and personal-training studios: one product from booking to payment, with an AI assistant that never leaks data to the model.
A Studio's Entire Operation in One Product — and an AI It Can Safely Talk To
Studios usually run customers, packages, bookings, staff and payments on disconnected tools. Motice brings them into one product: a business panel, staff and client apps, platform administration and an AI assistant — on web, iOS and Android from a single codebase.
Motice is our own product and the clearest showcase of our engineering system. Each business's data is separated at the database level; the AI assistant, Ask Motice, works through a single door that cannot exceed the user's permissions and never shows customer names to the model.
The Hard Part
Adding AI to business software means putting customer data in a prompt. That is why we built Motice Shield: names and record IDs turn into pseudonyms before reaching the model, and every model call passes a fail-closed egress gate. The model may only use identifiers it has actually seen — a rule we added after measuring that one invented branch ID could make 11 bookings show up as "0".
How We Built It
Forced database-level tenant isolation with a coverage test
A single-door tool registry for UI and AI; approval cards for risky writes
A pseudonym vault and an LLM egress gate that demands contract evidence
One codebase shipping to web, iOS and Android
A four-lane CI/CD pipeline and our agentic development loop
The Result
799 unit and 131 end-to-end test files, a PR gate cut from 13.7 minutes to 0.6, and more than 260 spec, plan and review documents. Motice is the living proof of the engineering system described on this site.
Let's produce a similar outcome together
If you have a need like this reference project, let's talk.
A system that runs education management and the Turkish proficiency exam on one platform — from application to verifiable certificate — for the Yunus Emre Institute, with centres in 69 countries.
An Exam Result Cannot Be Wrong
The Yunus Emre Institute had two large operations: Turkish courses delivered at centres around the world, and an international Turkish proficiency exam. Every step — from application to classroom, exam hall to result, appeal to certificate — had to be traceable, permission-bound and auditable.
We built the education and exam systems under a shared identity layer but as two separate domains. Each domain owns its database; services trust each other through local token verification without calling a central server, and events flow without being lost or processed twice.
The Hard Part
Paper exams are still real. We wrote an optical reader that reads legacy scanner files and PDF/image forms: it aligns the image with registration marks, sends ambiguous marks to a person instead of guessing, and never computes the score itself. For writing papers, the grader and the approver cannot be the same person — a rule written both in the application and in the database.
How We Built It
Shared identity, a separate database per domain, local service-to-service token verification
Transactional outbox: event flow without loss or duplication
An optical form reader with registration-mark alignment and human review of ambiguous marks
A maker-checker rule enforced in the application and the database
Hash-chained audit logs, versioned certificates and a verification portal
The Result
A platform carrying an operation that spans exam centres, halls, sessions, entry documents, grading and certificates, protected by roughly 1,400 test files. Every result and every certificate can be verified after the fact.
Let's produce a similar outcome together
If you have a need like this reference project, let's talk.
For the Yunus Emre Institute: a content network running the websites of 93 centres in 69 countries from one codebase, a culture and arts platform that opens to each country, and a bilingual digital magazine.
One Codebase, 93 Centres, 69 Countries
Each of the institution's centres around the world needed its own website, language and editors — yet brand, security and approval had to be managed from one place. We built a network of three platforms: a multi-site institutional content system, the Turkish Art Platform and a digital culture magazine.
The content system gives each centre its own subdomain; country roles see only their own sites, and content passes a seven-stage approval running from country to headquarters. The culture and arts platform manages artists, works, events and applications per country; the magazine publishes with per-issue access rights.
The Hard Part
Scale and multilinguality together. We wired machine translation into the workflow, but wrote a separate protection layer so template variables and editor markup survive translation intact. For the culture platform's maps we avoided dependence on a third-party service and hosted the basemap on our own servers.
How We Built It
A subdomain-based multi-site structure with a 10-level role model
Seven-stage content approval from country to headquarters, with revisions and previews
A machine translation workflow that protects variables
Shared user management through the institutional directory (LDAP); hash-chained audit
A digital magazine with per-issue access rights and branded PDF generation
The Result
93 centres in 69 countries publish from one codebase. The content system was built through more than 500 pull requests, and its end-to-end tests are planned, written and repaired by agents. The culture platform and the magazine were delivered hardened by security scanning.
Let's produce a similar outcome together
If you have a need like this reference project, let's talk.
An app and admin panel against gambling addiction that combines network-level protection on iOS and Android, risk assessment and supporter follow-up.
A Blocklist Heals No One. A System Can.
For someone trying to quit gambling, the hardest moment is when a single tap can take them back. BGB was designed to protect that moment: it blocks gambling sites on the device and the network before a connection is made, measures the user's risk, and keeps a supporter close.
We built the app, the admin panel and our own DNS infrastructure together. When protection is switched off, the supporter the user chose is notified; account deletion follows data-protection law with a cancellable window.
The Hard Part
Two platforms, two realities. On Android we wrote a native module for on-device DNS protection that restarts itself after a reboot. On iOS we generated signed, per-device DNS-over-HTTPS profiles, and we detect that a profile is active from the queries reaching our own resolver. We did not hide the platforms' protection limits; we documented them.
How We Built It
On-device DNS protection on Android with an offline matching fallback
Signed, per-device DNS-over-HTTPS profiles on iOS
Our own DoH resolver: per-device tokens and a large-scale blocklist
A risk questionnaire (0–100), supporter alerts, account deletion under data-protection law
The Result
A store-ready mobile product with subscriptions. The blocklist was designed to grow to a million entries; more than 6,600 rules match in constant time.
Let's produce a similar outcome together
If you have a need like this reference project, let's talk.
A B2B2C platform connecting real-estate advisors with residential developments, managing portfolio sharing, apartment locks, quotes and commission flows.
Two Advisors Cannot Sell the Same Apartment at Once
Keyperties opens developers' residential projects to verified real-estate advisors. An advisor shares a project with a client through a personalised portfolio, prepares a quote, the client accepts and the developer approves — and at every step commission and stock must stay correct.
We built the platform on a multi-tenant structure that keeps the data of all three parties — advisors, developers and platform operations — apart; quotes are rendered as PDFs on the server, and customer acceptance is verified with a one-time code.
The Hard Part
The sales process is a state machine. When an apartment enters a quote it is locked for a limited time; an accepted quote holds the lock until the developer decides; availability is re-checked at acceptance. Price revisions are versioned — the price of an open quote is frozen at that moment — and a revision drops the commission of the earlier acceptance.
How We Built It
Database-level tenant isolation with a leakage test for every table
Time-limited apartment locks, a quote state machine and scheduled sweepers
Versioned price revisions and a commission ledger recorded at the moment it is earned
Customer acceptance by one-time SMS/email code
Server-rendered quote PDFs and plan-based quotas
The Result
All 811 core and 134 end-to-end tests green; every tenant table is guarded by an "A writes, B cannot see" test.
Let's produce a similar outcome together
If you have a need like this reference project, let's talk.
Unitic's own collective AI brain: a system that serves company memory, knowledge and identity to every agent over MCP — and knows when each thing was true.
The AI Brain We Built for Ourselves
The strongest proof that a company builds AI is the AI it runs on itself. APEX is Unitic's collective brain: the single, durable source of company memory, knowledge and identity. Our conversation, coding and content agents connect to it over MCP.
We separated intelligence from the surfaces. Instead of scattered assistants each knowing half the story, there is one brain: it remembers, it knows, and it carries identity. It can read and write work items in Linear, access documents in Google Drive, and answer the team through Google Chat.
The Hard Part
Keeping memory from being wrong. We built a bi-temporal memory that supersedes knowledge without deleting it; near-duplicates are never written twice, and retrieval scores meaning, importance and recency together. Embeddings are generated locally, and knowledge with a retention period drops out when it expires.
How We Built It
An MCP server: recall, remember, context; Linear and Drive tools
Bi-temporal memory, deduplication and exponential recency scoring
OAuth 2.1 for people, fail-closed access for agents
An audit record and a PII-masked trace on every call
Cost-capped scenario tests
The Result
Unitic's agents run on the same brain today. The agentic development method described on this site relies on coding agents knowing the company's decisions — and that knowledge comes from APEX.
Let's produce a similar outcome together
If you have a need like this reference project, let's talk.
A web and mobile platform that manages mentors' calendars, bookings and institutional website integration for İTÜ Ginova, Istanbul Technical University's entrepreneurship centre, and similar institutions.
Two Bookings for the Same Slot — Impossible at the Database Level
In mentoring programmes, institutions often approve booking traffic by hand: the mentor's availability, the founder's request, the institution's approval chain. Mentorual takes the institution out of that daily chain — mentors manage their own calendars, and founders reach the same booking engine from the institution's website, a subdomain or the mobile app.
We are building the platform as a multi-tenant structure offered to institutions: each institution's data is separate, and institutional websites such as Ginova's connect to availability and bookings through an API. Sign-in is through LinkedIn and Google.
The Hard Part
Booking conflicts are prevented in the database, not in code: a second booking for the same time range is impossible because of a constraint. Concurrent updates are protected by version checks and repeated requests by idempotency keys. Institution isolation is applied in two independent layers — a fail-closed permission chain and database row-level security.
How We Built It
Double-booking protection by database constraint, with idempotency and version checks
Two-layer institution isolation and leakage tests on a real database
Transactional outbox and a sealed audit chain
A contract gate that catches backward-incompatible API changes
An availability and booking API for the institution's website; web and mobile
The Result
A platform where institutions leave the booking approval chain and double bookings are architecturally impossible. The next version plans mentor recommendations with a language model hosted in Turkey, for data residency under data-protection law.
Let's produce a similar outcome together
If you have a need like this reference project, let's talk.
