Build log · BTC Market Monitor

From one laptop to btcmm.app in eight days

Between October 1 and October 8, 2026, a paper-trading bot running on a single laptop became a monitored, test-gated, cloud-hosted web app with sign-in and its own domain. This page traces that change through 135 recorded change requests (CRs).

System

Kalshi 15-min crypto paper bot + dashboard

Window

2026-10-01 → 2026-10-08

Changes

CR-0001 → CR-0135

Release

1.1.2-beta

Trading

Paper only, no live order path

135change requests, each with a record and release note

501automated tests: 278 Python, 159 Vitest, 64 Playwright

3test layers: logic, screen components, full browser runs

8days from laptop to a public domain

Section A · System context

Before and after

October 1 · everything on one machine

As found

No automated tests in CI (the checks GitHub runs on every change), and no alerting.
The bot, the data and the public site all depended on one laptop staying awake.

October 8 · a small, managed cloud system

As built

Every service restarts itself and reports its health, and problems raise an alert.
Sign-in decides what each person sees: viewers get a read-only view, and only the admin has controls.

Section B · Change rate

135 change requests, by day

Section C · Change log

How the system changed, day by day

Oct 1CR-0001 – 0016

Foundation

CR-0001Code repository, change records, a verify step and a read-only preview
CR-0015Profit Taking Protocol added to the strategy engine
CR-0002–05Faster dashboard refresh; Whale Watch markers
CR-0016Whale Watch layout restored alongside the new protocol

From here on, every change goes through a branch, a written record and a release note. That paper trail is what made the fast pace below safe.

Oct 2–3CR-0017 – 0040

Data correctness

CR-0018Price index clock alignment; one shared forecast evaluation
CR-0022Independent BTC price proxies from three exchanges
CR-0035Private GitHub repository; Claude takes over development from Codex
CR-0037Window timing follows market time, not the computer's clock
CR-0039Depth-weighted proxy prices refreshed every second

The bot trades 15-minute windows, so time is part of the data. A clock running seconds slow skews results, so the logic moved to market time.

Oct 4CR-0041 – 0073

Data lifecycle and CI

CR-0041Faster Whale Watch queries through a database index
CR-0043Memory-leak audit
CR-0062Data retention: 7 days in detail, older data compressed into an archive
CR-0067Automated checks on every change, and one shared, test-gated release script
CR-0068–70Dependency security audits and code linting made mandatory
CR-0073Automatic dependency updates; production releases need green checks

Raw price ticks were nearly all of a database growing about 1 GB a day. Retention keeps its size steady, which made a small cloud server practical.

Oct 5CR-0074 – 0082

Test pyramid

CR-0075Component tests for the dashboard
CR-0076Full browser tests; coverage targets that only go up
CR-0077Replay harness pins every strategy decision on recorded days
CR-0078One-command release

The replay harness is the key safety net. Any change that alters a trading decision fails the build unless the change is intended and explained.

Oct 6CR-0083 – 0104

Cloud migration

CR-0091Read-only web access for viewers; controls for the admin only
CR-0095–96Production moved to a dedicated cloud server with a few minutes' switchover
CR-0097Bot health alerts and scheduled backups
CR-0098Releases deploy automatically from the code repository
CR-0101The main branch moves only after a change's checks pass
CR-0102–04Locked dependency versions, website health alerts, architecture doc

The data moved in two steps, a full copy and then the last few minutes, so the switch was quick and nothing was lost.

Oct 7CR-0105 – 0134

Identity and hardening

CR-0106More health alerts, an operations runbook and an alert test
CR-0107–10New sign-in system: shipped switched off, then turned on, then moved to production
CR-0112Abuse protection on the data API
CR-0116Safer automatic server updates
CR-0123Viewers share the same data responses, which lightens the server's load
CR-0124Terms, privacy policy and disclaimer, with tracked acceptance
CR-0127–29Version numbers, leaner build pipeline, overload guards
CR-0130Claude Code skills, review agents and safety hooks

The busiest day. Viewers got their own accounts, and the server got the safeguards a public site needs.

Oct 8CR-0135 – 0139

Public identity

CR-0135The site moves to btcmm.app; old links redirect with their path kept
CR-0135Sign-in moved to the new domain with every account kept
OpsContact email on the new domain
CR-0139Pre-launch hardening: sign-in checks fail closed, browser security headers, a log of admin actions, tighter abuse limits, unknown addresses go to the first page

Sign-in paused for a few minutes during the move, and the bot never stopped.

Section D · Review against architecture pillars

What each pillar looks like now

Pillar	October 1	October 8	Key CRs
Reliability	One laptop; the site stops when the laptop does	Dedicated cloud server, services that restart themselves, scheduled backups	0096 0097 0116
Security	One shared gate in front of the whole site	Per-user accounts, sign-in checked on every data request and failing closed, read-only by default, admin actions logged, browser security headers, secrets kept in a managed vault	0091 0109 0110 0125 0139
Operational excellence	Manual releases and restarts	Checks on every change, automatic deploys, a written runbook, health alerts	0067 0098 0101 0106
Quality	A single verify script	501 tests across three layers, a replay harness for every strategy decision, coverage that only goes up	0075 0076 0077
Performance	Database growing about 1 GB a day	Data retention with archive, indexed queries, shared responses, polling pauses in hidden tabs	0041 0062 0116 0123
Cost	Tied to a personal machine	A small, right-sized server and free tiers elsewhere; build pipeline usage cut by about 70%	0096 0128

Section E · Ground rules

Rules that held all week

Paper only

No real money moves

The bot simulates trades against live market data. It has no path to place real orders.

Recorded changes

Every change has a record

Each CR has its own branch, written record and release note, and passes the checks before it ships. Each one can be rolled back.

Calm releases

The bot keeps running

Dashboard updates ship on their own. The bot restarts only when its code changes, and only just after a 15-minute window ends.

Section F · Beyond

How this would scale

The current system is sized for a paper bot and a few dozen viewers. Each step below starts on a measurable trigger, not a date. The cheapest fix that removes the next bottleneck comes first, and the hard requirements for real money come last.

Stage 0 · today

One server, done well

Tens of viewers · paper only

One small cloud server runs the bot and the dashboard
Embedded database with retention and weekly backups
Managed sign-in, edge network, health alerts
Tested, automatic deploys

Recovery point: 7 days
Recovery time: ~30 min
Monthly cost: lowest

Stage 1 · harden

Separate and protect

>100 active viewers, or anyone outside the invite list

Bot on its own instance, isolated from viewer load
Continuous database backup to object storage
Shared API snapshots cached at the edge for 1–2 s
Infrastructure as code, a staging copy, blue/green dashboard deploys
Central logs and a latency target (SLO) with alerts

Recovery point: minutes
Recovery time: ~10 min
Monthly cost: ~2–4×

Stage 2 · scale out

Stateless and pushed

>1,000 concurrent viewers, or API p95 over target

Stateless API containers that autoscale behind a load balancer
Live updates pushed over streams instead of 1-second polling
Managed Postgres across two zones, plus a read replica
Ticks archived as columnar files, queried on demand
Two availability zones; the bot keeps a warm standby

Recovery point: <1 min
Recovery time: <5 min
Monthly cost: ~20–40×

Stage 3 · real money

Isolate the trading path

Any live order path, or users' own exchange keys

Trading engine in its own account and network
Pre-trade risk limits, position caps and a kill switch
Immutable order ledger and daily reconciliation
Per-user keys in a hardware-backed vault
Second-region recovery, on-call rota, outside security test, legal review

Recovery point: 0 orders
Recovery time: <15 min
Monthly cost: people-led

Recovery point is how much recent data a failure can lose. Recovery time is how long until the site is back. Costs are relative to today.

Stage 2 target architecture

To be

What breaks first	Why	Early warning	Fix (stage)
Memory	The bot and the dashboard share one small server	Memory alerts	Run the bot on its own server (1)
Polling fan-out	Each open tab asks for fresh data several times a second, so load grows with every viewer	Rate-limit hits, server CPU	Edge-cached snapshots (1), push streams (2)
Chart rendering	Charts are built on the server, so CPU grows with viewers	Slow chart requests	Cache at the edge (1), render in the browser (2)
Single server	One server in one location; a restore takes a while	Server health checks	Continuous backup (1), two zones and a standby (2)
Embedded database	One writer suits one bot; many bots or tenants need concurrent writes	Write waits	Managed Postgres (2)
Plan tiers	Free tiers cap users and protection rules	User-count alerts	Upgrade as traffic justifies (1–2)

Principle

Scale reads, not the bot

Viewers only read. Caching and pushing shared snapshots handle almost all growth, and the single bot never sees viewer traffic.

Principle

Keep the safety nets

The replay harness, checks before main and change records carry into every stage, so a bigger system still can't ship an untested trading change.

Principle

Money changes the rules

Stages 1 and 2 are about uptime and cost. Stage 3 adds isolation, audit and legal duties, and it isn't worth starting without a reason to trade live.

Independent research project; not affiliated with or endorsed by Kalshi. Results are from simulated paper trading. Simulated results have inherent limitations and do not represent actual trading; no account will or is likely to achieve similar results.