Five years. Six products.

The Turkish retail brokerage market lacked a serious, modern equivalent to what Groww and Zerodha built for India — a single platform where retail traders could research, watch, trade, and run algorithms across asset classes without stitching together five different tools. Existing Turkish offerings either felt like dated banking portals or limited mobile-only apps. The client wanted to lead the category, not enter it as another option.

The product needed to do three things at once: feel as approachable as a fintech consumer app, behave with the reliability and density of a professional trading terminal, and stay performant during the high-volume open and close windows of the Borsa Istanbul session.

Design-system-first. Every screen was built off a single Figma source — color theory, typography, spacing, density, and dark mode — so the platform read as one product across watchlist, charting, order entry, and back-office surfaces. Accessibility (contrast, keyboard navigation, focus management) was a baseline requirement, not a polish phase.

UX before features. Trading apps fail when the user has to think about the interface during a fast market. Reduced clicks-to-trade across every flow, tuned hover and feedback states for confidence under pressure, and used motion sparingly — present where it confirms an action, absent everywhere else.

Real-time data, honestly handled. Built a WebSocket subscription layer that reference-counts channels per symbol so the same instrument open across watchlist, chart, and order ticket consumes one socket, not three. Order book and price ticks tuned to update without ever shifting the user's focus.

Charting via TradingView. Integrated TradingView's advanced charts to give traders a familiar, professional charting surface — drawing tools, indicators, multi-timeframe analysis — without rebuilding what TradingView already does best. Custom overlays handled the platform-specific data (positions, fills, algo events) on top.

Algo-trading surface. Designed the UX for the algo-trading flow as a first-class feature, not a developer escape hatch — strategy builder, backtest viewer, and live monitor are accessible without leaving the main shell.

The client was running accounts payable and accounts receivable manually across hundreds of distributors. Invoices arrived as PDFs and email attachments, got entered into spreadsheets, then re-entered into QuickBooks. Approvals were chased over email. Errors surfaced weeks later during reconciliation. The cost was not the software — it was the people-hours and the mistakes.

The brief was to build a single web platform that accepted invoices at one end, ran them through the right approval chain, and emitted clean reconciled entries into the accounting software at the other end — without losing the control and review steps the finance team relied on.

One portal, four accounting backends. The platform sits in front of QuickBooks, QuickBooks Online, Xero, and Zoho Books, exposing a single consistent UI to the finance team regardless of which accounting system the client uses. Each integration handles its own auth, sync, and field mapping; the user never sees that complexity.

Upload-first invoice ingestion. Users upload AP invoices grouped by distributor in batch. The system parses, validates, and routes each invoice. Anything that fails validation lands in a Pending queue with the specific error highlighted, so a clerk can correct and resubmit without leaving the platform.

Multi-stage approval workflow. AP and AR both run through a configurable chain: Clerk → Marketing Manager → Payment Team. Each stage has its own portal view, comment thread, and audit trail. The Marketing Manager dashboard surfaces weekly batch reviews and bulk approval. The Payment Team sees only what's cleared the manager stage, with the reasoning attached.

Embedded reporting. Power BI dashboards are embedded directly inside the platform — distributor performance, approval bottlenecks, monthly payable trends — so the finance team gets analytics without leaving the workflow they live in.

Beyond AP/AR. The same architecture powers payroll automation, general ledger entries, email notifications across the workflow, and user management with role-based access (clerk, manager, payment, admin).

Enterprise decision-making at scale still runs on spreadsheets, scattered scripts, and one or two data scientists who can read them. The teams who actually need to act on the predictions — credit risk, marketing strategy, operations — are gated by tooling that requires Python, statistical fluency, and a willingness to wait days for a model to come back from the data team.

The brief was to compress that gap into a single desktop product. Analysts upload a dataset, define what they want to predict, walk through a guided workflow, and walk out with a deployable predictive model and the documentation to justify it — in hours, not weeks. Desktop-first because the data is sensitive: credit files, customer records, internal financials. Cloud round-trips were a non-starter for the target users.

Six steps, one screen at a time. The product is structured as a left-rail workflow: Data, Prepare, Train, Evaluate, Build Decision, Auto Documentation. Each step has its own canvas, its own state, and its own validation gate before the user can move on. The interface never shows a step the user is not ready for — fewer choices, clearer progress.

Visual model training via React Flow. The training canvas is a node-based decision tree built on React Flow, where users grow and prune the tree visually. Target variables are auto-detected and surfaced as branching options; linear and logistic regression nodes can be dropped in alongside tree splits. The user sees the model as a graph, not as code.

Offline-first by design. The Electron shell keeps the data, the model artifacts, and the documentation entirely local to the user's machine. No telemetry on dataset contents. IPC channels are scoped, the renderer is isolated, and secure state management keeps the model training state durable across sessions even if the app crashes mid-workflow.

Auto-documentation as a first-class deliverable. The sixth step generates a complete technical document covering data lineage, transformations applied, model architecture, hyperparameters, and evaluation metrics — exportable in a format the analyst's compliance team will accept. The documentation is not a side-effect; for most users in regulated environments, it is the actual deliverable.

Champion vs challenger model comparison. The Build Your Decision step lets users define a champion strategy and one or more challengers, then runs them against held-out data so the analyst can defend the chosen model with side-by-side performance, not vibes.