Astralchemist
Technical founder and full-stack engineer. Quant verification systems and local-first developer tooling. · @Astralchemist
Quant developer who builds trading verification and backtesting systems end to end, now extending the same discipline into local-first tooling for AI agents.
▶ Every signal replayed against real bars, then judged on stop discipline
Risklimitsinatradingbotanda10%bufferinacreditmonitorarethesameinstinct.
One builder, two tracks. Read it in whichever order fits what you are evaluating.
Who I am, in two tracks
I build the infrastructure that decides whether a strategy is real.
I ship local-first tools with hard guarantees and small surfaces.
Seven projects carry the evidence
Each one states what it does, what was hard, what I would do differently, and where the proof is.
Forex signal verification
Verify Telegram trading signal channels before trusting them.
Python · FastAPI · Polygon · PostgreSQL
| TP1 only | +2.1% |
| TP2 | +3.4% |
| Trailing | +1.8% |
| Half at TP1 | +2.7% |
| Risk managed | +4.9% |
BTC order-book research, pre-registered
A deterministic Rust engine that records BTC perpetual order books on three venues, replays them exactly, and tested a market-making strategy that does not work at retail fees.
Rust · WebSocket feeds · L2 queue model · Python figures
$ cli fill-study sessions/A..D 1 200 50
drift at 1 s, ticks (share adverse)
binance bybit okx
A −55.8 (0.84) −59.0 (0.64) −45.4 (0.78)
B −62.5 (0.71) −66.0 (0.63) −50.1 (0.78)
C −55.3 (0.69) −56.9 (0.69) −53.6 (0.76)
D −48.0 (0.80) −51.8 (0.79) −41.3 (0.71)
with the 173-tick maker fee: ≈ −280 ppm a filltag prereg-session-E
hypotheses committed
before the holdout ran
gated round trips
≈ −46 USDT/BTC
ungated, same exits
≈ −46 USDT/BTC
verdict no edgePublic quant work
A strategy catalogue plus research tools, from indicator to execution to statistics.
Python · MQL5 · NumPy · QuantConnect Lean
| strategy | category | runs |
|---|---|---|
| mean_reversion/bollinger_rsi | mean reversion | yes |
| momentum/breakout_atr | momentum | yes |
| trend/sma_cross_filtered | trend following | yes |
| statarb/pairs_coint | stat arb | no |
| research/hill_regime | research only | yes |
Trading infrastructure
Backtest as a service: the boring parts, built so the same strategy is never computed twice.
Python · FastAPI · TimescaleDB · Redis · Docker
POST /v1/backtest pair: EURUSD tf: 1h strategy: sma_cross(20, 50) from: 2025-01-01 to: 2025-06-30 200 OK 1.9 s cache: MISS key: 9f3a1c… trades: 87 sharpe: 1.12 max_dd: -4.8% win: 54% rate: 60/min remaining: 59
POST /v1/backtest strategy: sma_cross(slow=50, fast=20) tf: 60m pair: eur/usd from: 2025-01-01 to: 2025-06-30 200 OK 41 ms cache: HIT key: 9f3a1c… trades: 87 sharpe: 1.12 max_dd: -4.8% win: 54% rate: 60/min remaining: 58
Semantic codebase graph over MCP
npx rig-constellation gives a coding agent a map of the codebase instead of grep. Local-first, no key, nothing uploaded.
TypeScript · MCP · Node · SQLite
Cloud-credit runway from the menu bar
Answers one question: how long does the prepaid GPU credit last at the current burn.
Kotlin · Compose Desktop · SQLite · Keychain
Commitment pane fed by MCP
Tell an LLM to call the dentist Friday, an MCP tool call fires, and a checkbox appears on screen. No backend, ever.
TypeScript · Rust · Tauri · MCP
Also on the account
Fixes in other people's libraries
Bugs found by reference-value tests and benchmarks, reproduced with failing tests, and reported with a fix. 4 of 8 libraries have shipped the fix; the links go to the release notes or the thread.
Findings from live verification
CreditWatch 1.0.0 was built and unit tested on Windows and Linux. It claims macOS only, because no human had run the installed binary on the other two. The code shipped. The claim waited.
CreditWatch 1.0.0, 2026-09-21
| platform | built | unit tests | human ran it | claimed |
|---|---|---|---|---|
| macOS | ||||
| Windows | ||||
| Linux |
Attempts, in the order they happened
Sorting work by what shipped hides the attempts that taught the most. This is every attempt, what it tried, what it taught, and what it turned into. The skills panel fills as you read.
attempt 01 of 15 · 2020
MQL5 and MT5 bots built to client specifications on short contracts, alongside full-stack work.
The client's risk rules are the real specification. The entry logic is only what they think they are paying for.
The habit of writing the risk limits first.
A live Expert Advisor with 5% max risk, a daily loss halt, and breakeven at 1:1.
The risk rules were the product. The entries were not.
The risk vocabulary every later system reuses.
Four strategy families, mean reversion, momentum, trend following, and stat arb, each folder self-contained with a backtest.
A catalogue rots unless every backtest runs on a schedule. One broken clone URL costs more credibility than a missing strategy.
The Trading Strategy Library.
Hill estimator regime detection, EVT, Hurst, Hawkes, and a St. Petersburg wealth simulation.
Research that is not a strategy must say so in its own README. It does.
Read a signal channel and mirror its trades.
The signals were the unreliable part, not the execution.
fx-bot.
An execution bot, then a dashboard on top of it.
Nobody wanted to run a bot. They wanted to know whether to trust the channel first.
TGX-bot.
Grouped by what they were used for
Nothing appears here that a featured project does not demonstrate.
The questions a reviewer asks
Answered directly, including the ones with an awkward answer. Anything else, write to me.
No verified live track record. Verification systems are the product. A 90-day pre-registered paper track, with kill conditions committed on day 0 and logs written by script, is the first build on the readiness plan. It missed its 2026-10-06 start.