// Personal project
Project Kuber
I built a multi-agent trading intelligence system for the Indian stock market, delivered entirely through a Telegram bot instead of a website.
Tech Stack
How it works
Seven specialized agents, six of which make zero LLM calls, independently analyze a stock, and a shared deterministic core combines their outputs into a consensus signal that the AI is never allowed to override.
- Built five zero-LLM agents: Quant (technical, 10 indicators plus multi-timeframe analysis and a backtest lookup), Auditor (fundamental, a 10-point score across growth, margins, debt, and more), Oracle (sentiment, via a local FinBERT model, no cloud call), Regime (market direction, a 3-phase Sensex/sector/stock trend gate), and Risk (portfolio concentration, read-only).
- Built an Explorer agent for macro news and FII/DII flows that makes a single web search call, and an autonomous Nightly Screener agent that writes its own SQL query specs, validates them against the real database schema, and self-heals up to 4 times if a query errors out or returns nothing before falling back to a safe hardcoded query.
- Built a shared deterministic core: calculate_consensus() scores every agent's verdict from Strong Buy to Strong Sell, then a regime penalty gate tightens the call based on market and sector trend, it can only make a call more conservative, never upgrade it.
- Built six scheduled jobs on APScheduler: a morning brief, hourly intraday monitoring, an end-of-day report, the nightly autonomous screener, a prediction-grading job, and a weekly ranking job, pushing to Telegram and email concurrently.
Results
Specialized Agents
Typical Response Time
Why I built it
AI trading tools often let the model do the actual math, which makes the numbers hard to trust. I wanted a system where the financial calculations are always deterministic and verifiable, and AI only handles the explanation, not the arithmetic.
How I approached it
I split analysis across specialized agents, most of which run zero LLM calls, and combined their outputs through a shared deterministic core instead of letting an AI vote on the outcome, delivering everything through Telegram instead of a separate website.
Full System Architecture
The actual, code-verified architecture: the Telegram command layer, the LangGraph supervisor, all six scheduled jobs, the shared deterministic core, all seven agents, three databases, and the Nightly Screener's own self-healing query loop. Explore each part below, drag to pan and scroll or pinch to zoom.
