Inspiration
Retail traders juggle charting tools and chatbots. One side is measurable but mute; the other is fluent but ungrounded. Frankenstein was the right shelf: sew a deterministic TA engine to a modern agent that explains, plans research, and names risks.
What it does
Computes SMA/EMA/RSI/MACD/Bollinger and discrete signal labels; builds an AnalysisBrief (summary, research steps, risks) via LLM or offline template mode; serves /api/analyze for UIs; production face https://quantagent.dpdns.org/en; ships with .kiro steering, specs, and save hooks for Kiro IDE.
How I built it
Python 3.11+, pandas for indicators, FastAPI for the scar, optional OpenAI-compatible client for the brain. Spec-driven layout under .kiro/specs/market-analysis/ before feature code. Pytest covers indicator bounds and API smoke tests.
Hackathon source branch: kiroween-2026 (https://github.com/sharpBrother/quantagent/tree/kiroween-2026). The judges’ GitHub field uses the repo root https://github.com/sharpBrother/quantagent.
Challenges
Keeping the two halves honest: no LLM inside the TA path; graceful template fallback when keys or networks fail; clear "not investment advice" framing.
Accomplishments
Runnable local API + fixture data, MIT license, hackathon docs pack, Frankenstein metaphor reflected in module boundaries.
What I learned
A typed signal pack is a better "scar" than a megaprompt. Kiro-style requirements→design→tasks reduce accidental vibe coupling.
What's next
Richer multi-timeframe packs, better evals on brief quality, and tighter wiring from this open backend into the live QuantAgent UI.
Log in or sign up for Devpost to join the conversation.