DepthSight

An autonomous AI trading platform with a visual strategy builder, self-learning MCP memory, and a Slack Agent that backtests and optimizes algorithmic bots right from your chat.

Comment

Inspiration

Algorithmic trading is notoriously difficult for retail traders. It usually requires writing complex code in Python or PineScript, managing backtesting infrastructure, and constantly battling strategy overfitting. We asked ourselves: What if we could build a system where AI does the heavy lifting, acting as an autonomous quantitative researcher working alongside the trader?

What it does

DepthSight is a next-generation algorithmic trading platform and AI Co-Pilot. It empowers users to build, test, and deploy trading bots through a powerful ecosystem:

  • Visual Strategy Editor: A drag-and-drop canvas to build complex trading logic (filters, foundations, position management) with absolutely zero code.
  • AI Autopilot Agent: An autonomous loop (Generate -> Backtest -> Learn -> Repeat) that builds strategies, tests them on our proprietary engine, analyzes the KPIs, and rewrites the logic to optimize profitability.
  • MCP Memory Server: A custom Model Context Protocol implementation. The AI has semantic memory—it stores past successful parameter combinations and historical mistakes, ensuring it gets smarter with every optimization run and prevents overfitting.
  • Native Slack Agent: You can control the entire DepthSight platform directly from your Slack workspace. Run /depthsight optimize [asset], and watch the AI generate strategies, run distributed backtests, and render beautiful performance cards (with equity curves and PnL) right in your chat. You can even deploy the bot to live trading with a single button click in Slack!

How we built it

DepthSight is a full-stack, distributed system:

  • Frontend: Built with React, Next.js, and TailwindCSS for a premium, highly interactive drag-and-drop strategy canvas.
  • Backend & Engines: Powered by Python and FastAPI. We built two proprietary backtesting engines (a blazing-fast Vector engine and a Precision Kline-based engine) distributed via Celery and Redis.
  • AI Architecture: Driven by Google Gemini. We implemented a multi-agent system consisting of a Generator (writes the strategy JSON), a Critic (analyzes results), and a Strategy Advisor (queries the MCP memory and provides strict guidelines).
  • Memory (RAG): We built a standalone MCP Memory Server that uses semantic search to fetch trading rules from PostgreSQL.
  • Slack Integration: Built using the Slack Bolt API, featuring dynamic SVG-to-PNG chart rendering for backtest reports and interactive UI blocks.

Challenges we ran into

One of the biggest challenges was the "Over-filtering" problem in the AI Autopilot. Because the AI Strategist retrieved multiple successful patterns from the MCP memory (e.g., use Volume Filter > 2.5, use 1H Local Levels, use Squeeze momentum), the Generator would sometimes combine ALL of them into a single strategy using logical AND. This made the strategy so strict it would only trigger 2 trades a year! We solved this by refining the context window, tweaking the MCP prompt instructions, and adding a /depthsight clear command to easily reset the agent's context.

We also had to overcome significant engineering hurdles to synchronize long-running, asynchronous Celery backtests with real-time WebSocket progress updates on the frontend and polling loops in Slack.

Accomplishments that we're proud of

  • The Autonomous Loop: Watching the AI write its own trading logic, backtest it in seconds, read the drawdown/PnL, and successfully rewrite a better version of itself is pure magic.
  • The Slack Integration: We turned a complex trading terminal into a conversational experience. Rendering accurate equity curves and interactive "Deploy to Live" buttons inside a Slack thread feels incredibly futuristic.
  • The MCP Implementation: Successfully standardizing how our AI agents read and write long-term trading memory.

What we learned

We learned that connecting different AI agents via the MCP protocol allows for incredibly powerful long-term persistence, transforming a stateless LLM into a continuously improving quant researcher. We also learned that in UX, real-time visual feedback (like rendering live progress and charts) is absolutely critical when dealing with complex, long-running AI tasks.

What's next for DepthSight

We plan to expand our live trading integrations (Binance, Bybit), add support for more LLM providers to the Agent pool, and build community features where users can share their AI-generated strategies via social leaderboards.

Built With

Share this project:

Updates

Submission history