Hello, I'm Ömer.

What I explored

Selected work spans residential renovation, bespoke interiors and commercial spaces, with decisions checked against the existing building and client feedback.

What I took away

• I keep concept dimensions separate from verified manufacturing dimensions until the appliance and site conditions are confirmed.
• I treat a render as a spatial communication tool, not proof that represented lighting or construction details have been built.

What I explored

I shaped reporting and analysis workflows that turn order and product data into clear profit insights while separating measured values from estimates.

What I took away

• I label commission as either a value supplied by Trendyol or a calculated estimate so sellers can see where each number comes from.
• For anonymous product research, I combined usage quotas with progressive cooldowns and learned to balance abuse controls with the experience of real users.

What I explored

I turned a large asset catalog into a usable product through search, taxonomy, and private project boards. I brought storage and secure file delivery together with payments, subscription entitlements, account management, and marketing workflows. I developed hybrid retrieval that combines embedding-based semantic search with keyword results.

What I took away

• Managing hundreds of thousands of assets taught me to separate catalog data from file access while coordinating search, storage, and permissions.
• Connecting payments, subscriptions, and download entitlements required designing the product across payment events, account state, and file delivery.

What I explored

I worked on an e-commerce experience connecting a Turkish product catalog with an editorial storefront, product details, and store administration.

What I took away

• I paired catalog data with usage scenarios and editorial cards; keeping product details consistent across the storefront made content design easier.
• I used Polaris for store administration and a separate design language for the storefront; designing each surface around its users made the experience clearer.

What I explored

I developed ingestion flows that parse different page structures and metadata formats into a shared content model, working across platform-specific extraction rules, content previews, and local archive search.

What I took away

• Different metadata and page structures across social platforms led me to build platform-specific parsers behind a shared normalization layer.
• Caching content snapshots and cancelling outdated work taught me how much feed responsiveness depends on the data lifecycle.

What I explored

I developed a store foundation that handles guest and member orders, inventory, and payment events within one commerce flow.

What I took away

• I kept order creation and payment confirmation inside short database transactions so inventory and order state stay consistent together.
• I routed store notifications through an event queue; retries and Telegram's limited delivery acknowledgement made separate notification tracking necessary.

What I explored

I shaped the first directory pilot across DPP sectors around source observations, editorial review, approved content releases, and a controlled publishing flow.

What I took away

• I did not infer a provider's capabilities from its category label; linking each field to a source and observation date kept uncertainty visible.
• I did not turn research directly into publication; editorial approval and fixed content releases made the directory's scope and claims reviewable.

What I explored

I explored how session and diary memories can support continuity without treating memory as model training. The implementation stores chat exchanges and diary entries, adds a weekly digest, and retrieves memories alongside entries for generated summaries. Emotion is represented as multiple scores; mood is a derived label, rather than a replacement for those scores.

What I took away

• I learned to preserve separate emotion dimensions instead of reducing a person's experience to one positive-or-negative label.
• I used retrieval and periodic digests to give summaries continuity, while keeping stored memories distinct from the language model itself.

What I explored

I connected schema-constrained story normalization with multi-reference image-to-image generation. I managed style references separately from story-specific visual context and brought editable intermediate output, asynchronous generation jobs, and result archiving into one studio.

What I took away

• I found that separating the fixed style reference from story-specific references makes each image input's role easier to control.
• I made the normalized story editable before image generation so I could correct narrative structure without spending image-generation credits.

What I explored

I routed location, route, and maneuver updates through a shared navigation state into a mirrored HUD. Separating the provider adapter from presentation supported orientation changes, route simulation, and display switching while guidance remained active.

What I took away

• Separating navigation state from its presentation let me switch to the mirrored HUD while preserving the active route.
• I validate mirrored HUD layout and guidance behavior on a physical phone as well as in simulator runs.

What I explored

I brought speech-to-speech inference, audio-driven facial animation, and asynchronous frame delivery into one conversation pipeline. I worked on VAD, conversational turn-taking, barge-in, and cancellation of stale audio and frame queues, evaluating stream stability separately from lip sync and portrait quality.

What I took away

• I learned to test interruption and resume behavior for stale audio or frames, instead of judging realtime performance only by response latency.
• I treat stream stability and portrait quality as separate engineering questions; a working pipeline does not solve facial identity, mouth shape, or listening behavior.

What I explored

I built around model operation and content production: ComfyUI workflows generate clips, while Remotion assembles selected material into reviewable videos. I kept production quality separate from technical checks such as resolution, audio transcription, and codec metadata; a successful render alone does not establish visual approval.

What I took away

• I learned to preserve the prompt, seed, source clip, and editable composition so I can trace and revise a result instead of treating it as a one-off generation.
• I review rendered frames and spoken content separately from metadata, because technical validity does not establish natural-looking or publishable quality.

What I explored

I explored repeatable preflight and platform-specific browser steps with a fail-closed checkpoint. Each post is tied to an exact video and destination; verified outcomes are reconciled to an append-only history. A later Pinterest publisher applies the same idea to fixed, reviewed batches.

What I took away

• I kept the final publish or schedule action behind a human preview, because automating repeatable browser steps must not imply blanket permission to publish.
• I check the media packet, history, and platform result together so retries cannot silently create duplicate posts.

What I explored

LiDAR review, Blender mesh inspection, CAD or SketchUp modeling and preliminary takeoffs connect existing conditions to design decisions and presentation views.

What I took away

• I inspect scan geometry against known dimensions before using it to place design elements or estimate quantities.
• I label suggested model dimensions as preliminary and keep them out of fabrication decisions until the source and site are confirmed.

What I explored

I developed a pipeline to collect changing product information, manage recurring updates, and serve research views at million-product scale. I balanced work queues, bounded retries, collection throughput, and data freshness. The service was paused in October 2026.

What I took away

• At million-product scale, data collection became more than a one-off scrape: scheduling, retries, and resumable work were essential to keeping the system operating.
• I kept stock, seller, and ranking changes as separate signals and learned that obtaining fresh data and drawing reliable sales conclusions are different problems.

What I explored

I engineered product and category features and modeled sales probability and volume in separate stages. Quantile regression captured sales ranges and calibration addressed confidence. Product-group splits kept the same product from leaking across training and validation. Predictions concern past 3-, 7-, and 30-day windows.

What I took away

• Results from overlapping product records did not establish generalization; I used product-level holdouts and feature ablations to test assumptions.
• Alongside point estimates, I modeled when confidence was insufficient, using quantile regression and confidence calibration to identify uncertain predictions.

What I explored

Campaign discovery feeds bounded website research and evidence-linked company records; contact enrichment is gated and email sending is not implemented.

What I took away

• I position visualization as helping design teams communicate their decisions, rather than making those decisions for them.
• I preserve the source behind each research claim and do not infer missing people or contact details from a company's website.

What I explored

I connected Tiyasis calculation engines to a conversational interface with distinct access boundaries for free tools and account-linked Premium data. I developed MCP tool calls, OAuth, and interactive cards together; the integration remains in private testing.

What I took away

• I connected the integration to Tiyasis' existing calculation engine instead of rewriting it, then added parity checks to keep results consistent across both interfaces.
• I presented product photos and results in cards within ChatGPT; keeping the test deployment private separated product feedback from a public release decision.

What I explored

Adapted from the Avenox Beyin approach, the system queues lifecycle events, summarizes conversations in the background, and retrieves project context for later sessions. I separated daily records, session summaries, and durable knowledge, using rebuildable full-text indexes over canonical Markdown sources.

What I took away

• I keep Markdown as the canonical record and treat search indexes as disposable, rebuildable views of that record.
• I make lifecycle hooks leave short, durable queue signals and move slower summarization into a background worker.