Beta landing page liveCo-built with partnersPrivate client repo

AlgoDraft

AI trading strategy platform

AlgoDraft turns a trader's plain-English strategy into a structured, reviewable spec, asks about every missing or ambiguous rule, then generates a demo-only MetaTrader 5 Expert Advisor from tested templates. AI interprets; deterministic validation decides what is allowed. Includes a typed spec validator with SHA-256 approval, a Python reference backtester with reproducible manifests, and user accounts with a workspace.

Beta landing page

Built with: PythonFastAPITypeScriptReactViteMQL5Gemini APIPostgreSQL (Neon)DockerpytestRender
Type: Client & collaborative work
Source: Private client repository

How it decides

Strategy text → approved spec → demo EA

person
guesses / untrusted
decides
process
storage
third party
trust boundary

Boundary Where guessing stops and something deterministic decides.

AI interprets; deterministic validation decides what is allowed. Missing rules are never silently invented. repo docs
Live interpreter on 30 curated cases: 0 invented values, 100% recall. repo docs
Clarifying questions come from fixed templates, not from the model. repo docs
Approval is bound to a SHA-256 of the canonical spec. repo docs

The drawing, in words

AlgoDraft: Strategy text → approved spec → demo EA. Left of the redline is “Interpretation”; right of it is “Deterministic decisions”.

Trader — plain-English rules (Interpretation)
Interpreter — Gemini · structured (Interpretation)
Evidence guard — drops unbacked values (Deterministic decisions)
Validator — typed StrategySpec (Deterministic decisions)
Questions — templated, per gap (Deterministic decisions)
Approval — SHA-256 of spec (Deterministic decisions)
MQL5 EA — templates · demo-only (Deterministic decisions)
Backtester — reproducible manifest (Deterministic decisions)

Connections

Trader → Interpreter (text)
Interpreter → Evidence guard (draft)
Evidence guard → Validator
Validator → Questions (gaps)
Questions → Trader
Validator → Approval (valid)
Approval → MQL5 EA
MQL5 EA → Backtester

More screens

AlgoDraft landing page: Your trading idea, turned into exact rules.

The problem

Traders describe strategies in loose English. A model asked to turn that into code will happily fill gaps with values nobody chose. repo docs

What was built

A typed StrategySpec, a deterministic validator and SHA-256 approval of the canonical spec. repo docs
An interpreter that asks one templated question per missing or ambiguous rule. repo docs
An MQL5 Expert Advisor generator built from tested templates, and a Python reference backtester with reproducible manifests. repo docs
Accounts, owner-isolated strategies and a workspace UI. repo docs

My part

Co-built with partners. Individual contributions are not itemised here. stated

Limits and open work

Demo-only: no real-money trading and no broker connection. repo docs
The generated EA still needs a MetaEditor compile and a demo-account run (pending MT5 access in the project docs). repo docs

Decisions worth knowing

The model never has the last word
The interpreter only drafts. An evidence guard drops values the text doesn't support, and a deterministic validator decides what is allowed.
repo docs
Approval is pinned to a hash
Approval is tied to a SHA-256 of the canonical spec, so a changed spec is a different, unapproved spec.
repo docs
Measured, not assumed
The live Gemini interpreter was checked against 30 curated cases: 0 invented values, 100% recall.
repo docs

A closer look

text“Buy EURUSD when EMA 20 crosses above EMA 50 on M15. Risk 1 percent per trade.”
draftentry, timeframe, risk — extracted; stop-loss — missing
questionWhat stop-loss should each trade use?
speccanonical JSON, every value with its provenance
approvalsha256(canonical spec) — any edit makes a new, unapproved hash

Figure — from a sentence to an approved specrepo docs The sentence is the README's smoke-test strategy; the draft and question are an illustration of the documented pipeline, not recorded output.