Samih · Software developer
I build offline-first commerce systems and quantitative trading research: software that has to stay correct when the network drops and the market moves.
samih.online
Life & work ledger
Printed
samih.online
Thank you · scroll for the details ↓
Every sale lands in a local SQLite database first. Sync is background work that can fail, retry and catch up. It is never allowed to block the till.
Stock is an append-only ledger: what is on the shelf is always the sum of what happened. Client-generated UUIDs make every push safe to retry.
Encrypted nightly exports, and a drill that deleted tables, restored them and compared a fresh backup byte for byte: 29 of 29 identical.
Walk-forward folds, deflated Sharpe, bootstrap intervals. Each trading repo states its status plainly: what is built, what is tested, and what is not.
An offline-first point of sale and a public marketplace on one Postgres backend. Stores keep selling when the internet drops; neighbors browse what is in stock and reserve items for pickup.
CI/CD with a manual production gate · encrypted nightly backups · hourly health checks
Flutter · SQLite
One Flutter codebase for Windows, Android and web, with native Fluent UI on Windows and Material 3 on Android. Cash-drawer shifts, returns, expenses, reports, running customer balances, owner and cashier roles, and a hardware layer for ESC/POS printing, USB and camera barcode input and cash drawers (unit-tested).
Offline-first
Append-only rows (stock, sales, returns, expenses) merge as a union, deduplicated by UUID. Editable data is last-write-wins. Pulls use a per-table watermark. Two offline tills can sell the last unit; the ledger stays correct and shows negative stock for the owner to resolve.
Supabase
Plain, numbered SQL migrations: row-level security, SECURITY DEFINER RPCs, per-device auth tokens, scheduled reservation expiry. Schema ships backend-first, to dev before prod, with a typed confirmation to touch production.
Marketplace
Next.js 16 and React 19, server-rendered so search engines can index it. Stores, people, products and services; reserve-for-pickup; a household shopping list that updates in real time; a license storefront and admin.
Operations
Nightly encrypted backups with weekly permanent archives and a tested restore. Hourly health checks that also probe the security gate. Importers that move real shops off OpenSourcePOS and a Delphi/Firebird system. Checksummed Windows and Android builds.
Two tills share one bottle of olive oil. Take both offline, sell it twice, then reconnect. Ledger rows merge as a union, editable fields are last-write-wins, and every push is safe to retry.
Windows POS
Olive oil 1 L · stock here
oversold
Postgres
Sum of all ledger rows
oversold
insert … on conflict (id) do nothing
Android POS
Olive oil 1 L · stock here
oversold
Event log
A simplified model of the real engine, which adds per-table watermarks, per-device tokens and Postgres RPCs (sync_push, sync_pull).
Reinforcement-learning experiments on forex, engineered like production systems: safety rules in code, leakage tests, walk-forward validation. It is all research. None of it is advice, a track record or a product.
Python · PyTorch
Research code
A graph encoder over 28 pairs and 8 currencies feeds a PPO policy. A safety layer (exposure caps, max-hold, volatility circuit breaker, VaR cap, forced flatten) runs identically in training and inference.
MQL5 · MetaTrader 5
Backtest-only research · never run live
One shared Q-table decides once per H1 bar on seven USD pairs. State comes from triangulated pivot and RSI lines plus a PCA currency-strength meter. Illegal actions are masked out, not learned.
MQL5 · MetaTrader 5
Research code v2.00 · trading off by default
Eight instruments (seven USD pairs and gold), one independent Q-learner each, deciding once a day. A fixed-random temporal graph encoder that is never trained feeds the state; a rule engine hedges currency exposure.
How I treat results
“Research / backtest only. Not a proven profitable system” is the first thing in QTrader, TQLearning records no results at all, and forex_rl_v4 keeps a progress file that says what is still untested.
Limits are enforced by the same path in training, testing and live use, so a backtest cannot quietly assume a rule the live system would not follow.
QTrader’s state space went from 243 states to 54 because most cells never reached a trustworthy visit count. The redesign lifted the usable share from about 40% to about 79%, and a deviations table records why.
A tiny trader (27 states × 3 actions) learns on a synthetic market with the same update rule my MQL5 experts use: Q ← Q + α (r + γ · max Q′ − Q). The last 30% of bars are a frozen, out-of-sample test. Then try twenty random markets and pick the best: the trap that deflated Sharpe exists to catch.
Q-table · how much better each action looks
Learning starts when this lab scrolls into view.
Best of 20 random markets · each dot is one market’s out-of-sample P&L
Synthetic data with a visible seed, so every run is reproducible. P&L is in units of one bar’s volatility; costs are 0.03 per unit of position change. This demonstrates a method and a trap. It is not a strategy, a backtest or advice.
After paying for its lessons, the agent mostly learns to sit out, which is the right answer. When it does trade, it is chasing phantom patterns.
Same code, same random numbers, but each return leans a little toward the last. The table finds it, and the frozen policy keeps earning on bars it never saw.
Try enough random markets and one will look like skill. That is why I freeze the policy, test walk-forward across folds, and correct for how many things I tried.
Milestones, not commit counts. A number of commits says little about what was actually shipped.
The account opens with C notes on threads, processes, sockets and sed: the operating system underneath everything else.
An Expo and Next.js monorepo with a hybrid local/server ledger: idempotency, an offline queue, reconciliation. Alongside it, a C core compiled to WebAssembly and wired into a Telegram Mini App demo.
Flutter POS, Supabase backend and Next.js marketplace scaffolded. The first working POS screens (products, stock, customers, reports, returns) land the same day.
A license storefront, a browsable store directory that needs no login, and an outbox-to-Postgres sync deduplicated by event ID.
GNN and PPO trading research with a safety layer and a walk-forward harness, in a public repository.
Reserve-for-pickup with expiry, a live shared shopping list, and a canonical product catalog that de-duplicates marketplace search.
Two tabular Q-learning experts for MetaTrader 5, each with an honest status section instead of a results table.
Device authentication on every sync call, separate dev and prod environments, and CI/CD with a manual production gate. Domains live on Cloudflare.
Windows installer, portable build, Android APK and web app, with checksums and bilingual install notes.
This site: a portfolio that runs its own experiments.