asepharyana
3882c5aef0
- Introduced AGENT-SPECS.md detailing the specifications for seven agents including their inputs, processing steps, and outputs. - Created API-REFERENCE.md outlining the Sectors API v2 endpoints, parameters, costs, and usage. - Developed API.md to specify backend routes, request/response structures, and error handling. - Established ARCHITECTURE.md to describe the project layout, conventions, scheduler, and citation pipeline. - Added DATA-MODEL.md to define the database schema, tables, and seed strategy. - Compiled PLAN.md to outline the project concept, problem statement, unique features, and implementation timeline. - Created README.md as an index for documentation with links to all relevant files. - Documented ROUTINES.md detailing the seven automated routines, their schedules, inputs, detection logic, and delivery formats. - Introduced TECH-STACK.md to specify the technology choices and rationale for both backend and frontend components.
3.7 KiB
FlowSight Roadmap
What is built, what is next, and what has been deliberately declined. Reordered when evidence says the order is wrong.
Nothing here is a date. Items move to TODO.md when they are next up.
Shipped
- Plan + API reference from live schema. docs/PLAN.md (concept, routines, verifiable AI, accuracy ledger) + docs/API-REFERENCE.md (all 70 v2 paths with params, costs, per-cycle credit budget). Source: schema.json + docs site.
Next
Hackathon core (maps to TODO Now, in order)
Foundation — data in, health out. Scaffold (FastAPI + Next.js), Sectors client with credit counting + param narrowing, SQLite schema (§8 tables), historical seed. Health endpoint proves the pipeline breathes.
Ingestion — the 30-minute heartbeat. Scheduler pulls universe sweep (close/),
market context (top-changes minimal combos, most-traded, idx-total, brokers/top),
per-watchlist depth (broker-summary/top, foreign-flow, daily), incremental events
(news/filings/suspensions), quarterly freshness (since=). Redis caches slow-moving
reference data. Credit spend stays inside the §12 budget.
Agents — seven specialists, one synthesis. Smart Money Tracker and Broker Intel are built first (the moat: no competitor fuses broker × foreign flow). Sentiment (Adaptive RAG), Fundamental, Technical, and Event Catalyst follow the same input→detect→score contract. Master Synthesizer correlates cross-signal agreement into conviction, flags conflicts, and sizes positions by risk profile.
Routines — data becomes habit. Routine engine with cron schedules turns agent output into deliveries: Morning Briefing (07:30), Accumulation Radar (30-min), Foreign Reversal Watch, Insider Tape, Earnings Countdown, Dividend Calendar, Weekend Review. Each run is recorded; each delivery carries citations.
Trust — verify, then track. Every number cites endpoint + snapshot timestamp; stale data is labeled, never hidden. Accuracy Ledger records each recommendation and resolves it at +30d against actual return; agent weights follow the ledger.
Surface — see it, run it, export it. Dashboard with live agent panel (SSE), Routine Manager, Institutional Screener, Alert Engine UI, One-Click Report with interrogation scoped to citations, Portfolio Risk + accuracy table. Demo seed + script + deck close the loop for judges.
Post-hackathon
- SGX extension (buybacks + short-sell have no IDX equivalent — new angles).
- KLSE basic coverage through the same screener.
- Mining vertical: commodity price → miner watchlist linkage.
- WhatsApp delivery; per-user keys and watchlists; backtest harness for rules.
Later
- Push-first mobile. Native-feel PWA with push for radar alerts; web stays primary.
- Community routines. Shareable routine templates (e.g. "dividend hunter", "foreign follower") with fork counts.
- Multi-market synthesis. One briefing spanning IDX + SGX + KLSE positions.
- LLM cost control. Spend ceiling per routine run; cheaper model for sentiment triage, strong model only for synthesis.
Declined
- Order execution / trading. Read-only intelligence; executing trades adds regulatory surface no hackathon needs.
- Price prediction models. Directional forecasting competes on accuracy claims no 48h build can defend; the ledger tracks recommendations, not price targets.
- Real-time tick streaming. Sectors data is end-of-day granularity; pretending otherwise would fake the product. 30-min cycles match the source.
- A general chatbot. Conversational UI exists as a scoped sidebar only; the product is routines that run without being asked.
- v1 API support. v1 returns 410 Gone; no compat layer will be built.