Bilginer Sarı

    WORK INDEX

    Work

    FIVE PROJECTS

    Five projects, written up the way I actually worked on them: the signal that started the work, the decision that followed, the technical constraint that shaped it, and what the outcome taught me.

    Currently building ooba — an LLM training workflow and a passive health tracker for people who will not log anything by hand.

    INVENTORY

    All projects

    Fintech, consumer AI and internal tooling — each entry keeps the trade-offs in, not just the result.

    PROJECT 01

    2025 / Shipped

    Bundle — AI-focused redesign

    Bundle was drifting into a feature catalogue. We rebuilt the product around a single AI-assisted flow so a new user reaches a useful answer in one screen instead of five.

    ROLE / SCOPE

    Product lead — discovery, scope calls, design direction, launch

    DECISION TRACE

    01 / SIGNAL

    Support logs showed most sessions ended before the first result: people could not tell which tool to open.

    02 / PRODUCT DECISION

    Collapse the tool menu into one prompt-first entry point and let the model route the request.

    03 / TECHNICAL REALITY

    Routing accuracy was only good enough with a narrow tool set, so we shipped four routes instead of eleven.

    04 / OUTCOME / LEARNING

    Time-to-first-result dropped sharply, and the narrower route list turned out to cover most real requests.

    View case study

    PROJECT 02

    2024 / Shipped

    Midas — Advanced Trader Tools

    Experienced traders were leaving for terminals with faster order entry. We built a dense, keyboard-driven layer without breaking the simple app everyone else uses.

    ROLE / SCOPE

    Product manager — research with active traders, spec, rollout

    DECISION TRACE

    01 / SIGNAL

    A small group of high-volume users produced most order flow and complained about clicks per trade.

    02 / PRODUCT DECISION

    Ship an opt-in advanced mode rather than making the default app denser for everyone.

    03 / TECHNICAL REALITY

    Real-time depth data was expensive per session, so advanced mode streams only while the panel is visible.

    04 / OUTCOME / LEARNING

    Order entry time fell for the target group with no measurable regression in the standard experience.

    View case study

    PROJECT 03

    2024 / Shipped

    Midas — Detailed Search

    Search returned tickers but not intent. We rebuilt it to answer questions like “dividend-paying industrials” instead of only matching symbol strings.

    ROLE / SCOPE

    Product manager — query analysis, ranking rules, evaluation

    DECISION TRACE

    01 / SIGNAL

    Roughly a third of searches were natural-language phrases the exact-match index could not serve.

    02 / PRODUCT DECISION

    Add a filterable, intent-aware layer on top of the symbol index instead of replacing it.

    03 / TECHNICAL REALITY

    Full semantic ranking was too slow at peak, so we pre-computed filters and kept semantics as a re-rank.

    04 / OUTCOME / LEARNING

    Zero-result searches dropped, and the pre-compute pattern became the default for other list views.

    View case study

    PROJECT 04

    2025 / In progress

    ooba — LLM Trainer

    A training workflow for teams who have domain data but no ML staff: upload examples, run a fine-tune, and compare outputs side by side before promoting a model.

    ROLE / SCOPE

    Founder — product definition, evaluation design, build

    DECISION TRACE

    01 / SIGNAL

    Teams had usable data sitting in spreadsheets and no safe way to test whether fine-tuning helped.

    02 / PRODUCT DECISION

    Make evaluation, not training, the centre of the product — you promote a model only after a visible diff.

    03 / TECHNICAL REALITY

    GPU cost forced short runs, so defaults favour small adapters over full fine-tunes.

    04 / OUTCOME / LEARNING

    Reviewers trusted the output because every promotion is tied to a comparison they can read.

    View case study

    PROJECT 05

    2026 / In progress

    ooba — Health Tracker

    Most tracking apps ask for daily discipline nobody has. This one reads what is already being recorded and only asks a question when the data is ambiguous.

    ROLE / SCOPE

    Founder — product definition, data model, build

    DECISION TRACE

    01 / SIGNAL

    Logging drop-off happened in the first week, long before any trend could be shown.

    02 / PRODUCT DECISION

    Import passively from existing sources and reserve manual input for genuinely missing context.

    03 / TECHNICAL REALITY

    Source data is irregular and gappy, so the model reports ranges rather than false precision.

    04 / OUTCOME / LEARNING

    Retention past week one improved once the app stopped demanding daily entry.

    View case study