Investment monitoring and simulation
A dual-horizon advisor that argues with itself before it recommends anything, then paper-trades its own advice.
-
Two horizons, and they disagree.
One agent grades fundamentals and a DCF, another grades momentum. The same holding can be a buy on one clock and a wait on the other, and the report says both.
-
A bull and a bear argue first.
Two researchers take opposite sides of the same position before anything is recommended. The disagreement goes into the output rather than being smoothed out of it.
-
Then it takes its own advice, on paper.
Position sizes come from real account NAV, deterministically. Every recommendation is entered into a simulated book, graded at entry, and scored against simply holding.
- 01 Built a dual-horizon multi-agent investment advisor — DeepSeek-graded long-term (fundamentals/DCF) and short-term (momentum) signals, a bull/bear researcher debate, deterministic position sizing off real IBKR NAV, and an options overlay, delivered nightly per user via Telegram
- 02 Fixed a position-sizing defect in an automated trading system that recommended stock purchases up to 1.7x the account's available cash, by implementing cash-aware sizing caps across equity and options order sizing.
- 03 Redesigned an LLM-based short-term trading signal grader to incorporate current holdings and prior-day signal history, fixing a HOLD recommendation state that was unreachable by design and measured at only 4.3% of signals versus 84.3% WATCH across 140 signal-days.
- 04 Added a paper-trading simulation layer — deterministic entry scoring (A-F grade, no LLM) on every trade, a Δ-vs-Hold dashboard across multiple simulated books, and a lessons-learned loop that feeds post-mortems back into future recommendations
Interface recreated with invented data — no production screenshot is published. Private system; walkthrough and code review available on request.
- Role
- Sole engineer
- Timeline
- 2026 — ongoing
- Stack
- Python · LLM agents · IBKR Flex · Telegram