Tengen An AI trading platform that connects cause to effect.
Tengen is an AI algorithmic trading platform, built data-first. Its edge is connecting cause to effect as it happens: an earnings filing, an insider buying, a burst of news, linked to what prices did in the sessions after. Nothing trades until the data under it can be trusted.
- Role
- Sole builder
- Built as
- Operator console and Rust pipeline
- Trades
- On paper only
- Version
- The second build
- Receipts
- 275 commits · 75 test files · since 2 September
- events
- 1,284 on 212 days
- best horizon
- h3, from next open
- mean net
- +0.58%
- t, by day
- 3.42
- hit rate
- 0.57
- halves
- +0.61 · +0.55
| algo | ic | sharpe | weak link |
|---|---|---|---|
| earnings-drift-h3 | 0.041 | 1.41 | none |
| peer-follow-h2 | 0.019 | 0.88 | finding |
| swing-ridge-decile | 0.027 | 0.64 | costs |
| insider-window | 0.006 | 0.12 | model |
| news-burst-fade | 0.022 | — | no book yet |
A model is easy. Trusting what feeds it is not.
My first attempt built five model engines and no governance for the data feeding them. Every signal looked plausible and none could be checked: I could not say when a fact became knowable, whether a run had finished, or where a number came from. An edge built on that is a story, not an edge.
Data first, then cause, then money.
This build inverts the order. Every source is declared, every pull is planned and rehearsed before it fetches anything, and every run is a durable row. Every fact carries the moment it became knowable, so a backtest never sees tomorrow. Only then does it ask the real question: when this happens, what follows, and does it pay after costs? An idea climbs from a finding to a model, an algo and a book, and each step cites the one before it.
| known at (et) | ticker | event | score | side | size | entry | state |
|---|---|---|---|---|---|---|---|
| 14 Oct 16:05 | AVGO | beat 1.8σ | 0.92 | long | 2.0% | filled 15 Oct open | filled |
| 14 Oct 16:11 | JPM | beat 1.3σ | 0.81 | long | 1.6% | filled 15 Oct open | filled |
| 14 Oct 16:20 | NFLX | miss 1.6σ | 0.77 | short | 1.4% | working, 3 of 5 fills | working |
| 15 Oct 07:02 | UNH | beat 1.1σ | 0.64 | long | 1.1% | next open | queued |
| 15 Oct 07:15 | ADBE | beat 2.4σ | 0.95 | long | — | semis and software cap | held by risk |
| 15 Oct 07:31 | PEP | beat 0.6σ | 0.31 | — | — | below the 1σ floor | skipped |
| 15 Oct 08:02 | KO | known late | — | — | — | 8-K stamped after open | refused |
- the move
- weight 0.10 → 0.14, from the next session
- quorum
- 4 of 5 seats, met
- against
- 0.17 · operations
- veto
- risk holds it, not used
- written as
- control.decisions · new allocation row, the old one closed
- sitting cost
- 38 calls · 291k tokens in · 22k out
F3.2 is established on 1,284 events and both halves agree. The book has tracked its promise within 40 bp for 24 sessions.
finding:F3.2book:earnings-drift-h3Paper Sharpe 1.41 against 1.52 promised. The gap is slippage on the open, not the signal. 0.14 keeps it the third largest weight.
report:paper-w6sweep:alloc-0.10-0.20Gross stays under 1.4× after the move. The semis and software cap at 18% binds first, so ADBE stays held.
limit:sector-capmarks:15-octThree 8-K timestamps last week landed after the open. Hold the increase until the knowledge-time check is green again.
qa:K-TIME-07ticket:T-0412Not a trader’s kind. At 0.14 the opening auction absorbs every name in the cohort under 2% of its volume.
desk:shortfall-sep| level | what | where |
|---|---|---|
| act | minute bars corridor failed at 01:12: the vendor answered 429 past the throttle's retry budget | execution |
| watch | K-TIME-07 failing: 3 filings stamped after the session they claim to be known in | qa › knowledge time |
| watch | 1 reconciliation ticket waits on the operator: two sources disagree on a split ratio | curation › reconciliation |
- socket
- one connection, held by the process · opened by a click 14:21:40, never on load
- observed
- 48,212 events · 19,904 trades · 27,118 quotes · 834 second bars · 14 minute bars
- stored
- nothing: a passthrough, computed nowhere
Screens drawn with dummy data. No real customer, client or account appears.
What it does
Cause, linked to effect
The event bench takes something the store recorded, such as an earnings 8-K, an insider purchase or a burst of news, and measures what prices did over the next one to five sessions against the cohort, net of the round trip.
A ladder of evidence
Findings motivate strategies, strategies hold models, models feed algos and algos feed a fund. You enter each tier from the one above, on the evidence, never from a blank form, and the lineage graph draws the whole chain.
Three rings before money
An algo is backtested as of each decision time, then traded on paper on a real clock with its own balance sheet, then live-tested, promise against record. It moves forward one ring at a time, on evidence.
A council of AI members
Five seats, research, allocation, trader, risk and operations, sit on a schedule and govern the fund on paper. Each vote gives its reason and cites the rows it read, and risk holds a veto.
Pulls that rehearse first
Sources live in config, pulls are scheduled and resumable, and every verb shows its work list before it fetches. Vendor bytes land as received, with a hash, and curated data reconciles against disk.
A console that tells the truth
Did it run, how fresh is it, what broke. The console answers all three, watches the vendor’s stream live, and says plainly when a screen is empty and what would fill it.
More of the work
Sentinel
A Mac companion for AI agents
One window for what an AI agent needs from you, what it's doing and in what order, what it has learned, and what it left running. It keeps the Mac inside a budget, and every check is broken on purpose to prove it works.
Roomi
Pitch the product, not the deck · getroomi.com
A platform for deploying working demos into a pitch room: the demo, the narrative and the materials behind a link per viewer, with analytics and feedback coming back.
SMD
Social media distribution
One system that creates, distributes and measures social media for every project I run, and judges each post by the leads and the won work it produced, not the likes.
Daedalus
Co-founder and CTO · since 2022
Data engineering, data science and software architecture for B2B financial-services ventures. The client work is private; the approach is everything on this site.
Alfred
A chess coach that runs on your Mac
Paste a game or drop in a screenshot and it shows the best moves in order, explains them in plain words, plays them out on the board, and learns your patterns across games.