Kinetix Lens

What it is and why I built it

Kinetix Lens grew directly out of Kinetix. Kinetix worked well for me, and it gave me a pattern I trusted: let software do the watching, make every setup clear several checks, and only interrupt me when something earns it. I wanted to bring that same idea to the stock market.

SEC filings seemed like a natural place to start. Companies and their insiders are required to report certain events publicly, so when an officer or director buys their own company's stock, or a company discloses a major event, it lands in a public record. That's a signal software can read for me, instead of me digging through filings myself.

How it works

Kinetix Lens watches two kinds of SEC filings. Form 4 covers insider transactions, which is when a company officer or director buys or sells their own stock. 8-K covers material events, like acquisitions, officer changes, new agreements, and other things companies are required to report. Several times a day during market hours, an engine called Scout pulls these from the SEC's public EDGAR feed.

Scout filters out the noise first: preferred shares, mutual funds, placeholder tickers, and 8-K categories that aren't material. It then cross-checks each filer against Alpaca's list of tradable stocks to confirm the ticker is a real, tradable U.S. equity, and groups everything by ticker. Each ticker that survives gets one AI call that reads the raw filing data and produces a plain-language summary and a confidence rating. The result is a candidate list.

The AI is Claude Haiku, and it's steered toward facts: dollar amounts, how many insiders bought, what the filing actually says. It isn't asked for recommendations. The confidence rating is high, medium, or low, and each tier has a concrete definition. High means something like three or more different insiders buying within the same week, a single insider purchase over $1M, or insider buying paired with a material corporate event. Medium means a noteworthy 8-K without insider buying to back it up. Low means marginal relevance. The summary tells me what happened, and the rating tells me how significant it looks. The decision stays with me.

About 30 minutes after Scout's last run of the day, a second engine, the Main Engine, takes the candidate list and runs every ticker through four technical gates. All four have to pass. RSI has to land in a defined zone, neither overbought nor oversold, which screens out stocks that are already running hot or in free fall. The 20-day exponential moving average has to sit above the 50-day, with price above the 20-day, which confirms an established uptrend. MACD has to be above its signal line, meaning momentum is building instead of fading. And current volume has to beat the 20-day average, so there's real participation behind the move.

The gates are intentionally picky. In practice MACD is the strictest one, blocking most candidates when the market is pulling back. That's the point. The system is meant to sit quiet during runaway trends and corrections, and only fire during measured resumptions with volume behind them.

When a ticker clears all four gates, a formatted alert goes to Telegram with the ticker, the price, all four indicator readings, and Scout's AI summary, plus Approve and Skip buttons. Nothing executes without a tap. On approval, the trade goes through Alpaca, a regulated U.S. brokerage with a proper API, and a slippage shield blocks the order automatically if the price has moved too far since the alert fired. I get a confirmation in Telegram either way. That last step is still being wired up, and What's next below covers where things stand.

The Funnel

The part that still gets me is the filtering. The SEC feed dumps hundreds of filings a day, and most of it is noise: mutual fund share classes, preferred stock conversions, placeholder tickers that just say "N/A," tickers with spaces between the letters. The system chews through all of it, applies increasingly aggressive filters, hands the survivors to an AI for synthesis, then runs what's left through four technical gates that reject most of it.

Out of that whole funnel, maybe a handful of candidates a week survive to become an alert. Each one lands on my phone as a Telegram message, the whole SEC filing universe boiled down to "here's the one worth looking at, and here's exactly why."

The AI summaries have a deadpan quality, too. A typical one reads something like: "Three directors purchased a combined $2.1M in shares within a four-day window; 8-K filed same week disclosing a new partnership agreement." No hype, no "strong buy signal," just the facts. The confidence tier does the editorial work, and there's something satisfying about a system that refuses to editorialize.

What's next

Kinetix Lens is in development, and the backend comes first. Scout is live today, running on schedule against real SEC filings, with real AI scoring and real candidate lists produced every trading day. The Main Engine is running on schedule too, evaluating those candidates against the four gates. The whole system is in paper trading mode, so alerts fire but no real money moves.

The Approve and Skip buttons are the current build. Once they're wired up, the system will run end to end through Alpaca's paper trading environment for a validation period before any real capital is involved. There's no dashboard yet. For now I monitor everything through Telegram alerts and server logs.

Once the backend is solid, I'll build a UI on top of it and keep testing and fine-tuning. Eventually I'll decide whether to release it in the app stores or keep it a personal tool.

Tech stack

• PHP on HostGator shared hosting. Intentionally simple: no framework, no containers, runs on cron.
• Alpaca brokerage API for market data and trade execution (paper trading mode for now).
• SEC EDGAR public filings feed (Form 4 and 8-K).
• Claude Haiku (Anthropic) for AI synthesis of filing data.
• Telegram Bot API for alerts and approval buttons.