Tessera

A live Indian-market news terminal + a multi-agent equity research engine + an AI layer that makes every number on the page able to explain itself. Hover any P/E, ROE, Promoter % on a company page; click for a deeper look. Researchers run a four-analyst LangGraph pipeline with a bull/bear debate that issues a structured FinalDecision (rating, target, stop, horizon, risks, invalidation triggers).

                Frontend · Next.js 15 + React 19 + Tailwind v4
                          (3-pane shell · ⌘K palette)
                                    │  X-User-Id
                                    ▼
        ┌────────────────────────────────────────────────────┐
        │              FastAPI · Python 3.12                 │
        │  /api/markets · /api/news · /api/company · /api/ai │
        │  /api/research · /api/watchlist · /api/alerts      │
        │  in-process queue worker (3 concurrent runs)       │
        │  in-process notifier (Twilio WhatsApp drain)       │
        │  in-process TierPoller (Tier 1/2/3, 30s/2m/30m)    │
        └────────────────────────────────────────────────────┘
   ┌──────────┬──────────┬─────────┬──────────┬──────────┐
   ▼          ▼          ▼         ▼          ▼          ▼
 Supabase  Anthropic  Twilio  yfinance  Upstash  10 news
 PG 17     Sonnet/Hk WhatsApp  + NSE    Redis    RSS feeds
 +pgvector  + AI                        REST     (3 via GN)

Updates every 60s. Composite system status: unknown.

1. Four analysts run in parallel — Fundamentals (Screener + yfinance), Technical (3×4 indicator matrix), News (last-14-day classified tagged news), Sentiment (aggregated headline sentiment).
2. Bull and Bear researchers debate (1–2 rounds), each citing the analyst reports.
3. The Research Manager issues a structured InvestmentPlan picking the strongest points from each side.
4. The Trader turns the plan into an entry/target/stop TraderProposal.
5. A three-way risk team (Aggressive / Conservative / Neutral) critiques it.
6. The Portfolio Manager issues a FinalDecision: rating, confidence, position size, horizon, key risks, invalidation triggers. Persisted and resolved nightly with realised return + alpha vs NIFTY 50.

Every numerical cell on the company page is wrapped in an AskTessera surface. Three latency tiers keep cost flat:

• Pre-baked — field-level hover explanations generated at each fundamentals refresh, Postgres-cached by (ticker, field, value_bucket, refresh_date). Lookup is a PK hit, ~5-15ms. Haiku batched ~$0.004 per company.
• Cold-cached — concept tooltips, news impact lines, daily “what changed” widgets. Generated lazily on first hit, cached forever or 24h.
• On-demand — multi-agent research deep dive. Sonnet ~$0.20-$0.40 per run, budget-capped by TSR_DAILY_LLM_USD (default $15).

• Quotes-API service forked from 0xramm/Indian-Stock-Market-API (MIT)
• Multi-agent pipeline inspired by TauricResearch/TradingAgents
• Fundamentals from Screener.in (robots.txt-compliant scrape), plus yfinance + nsepython
• Indices from niftyindices.com (NSE) and BSE bhavcopy
• News: MoneyControl, Economic Times, LiveMint, Business Standard, Reuters, Bloomberg, Yahoo Finance, Google News, NSE + BSE corporate announcements (RSS / official feeds)
• UI: Tailwind v4 (OKLCH design tokens), TanStack Query, cmdk command palette, lightweight-charts v5, Lucide icons, Geist Sans/Mono
• Backend: FastAPI, SQLAlchemy 2.0, Pydantic v2, Alembic, LangGraph, Anthropic SDK
• Hosting: Render free tier (main API + quotes-api), Supabase (Postgres + pgvector), Upstash (Redis REST), Vercel (Next.js), Twilio (WhatsApp sandbox), GitHub Actions (cron + keepalive)

• GitHub repo (commit history shows every phase A-J of v3 + B-E of v4 + A-E of v5)
• /admin/costs — cost transparency dashboard (token-gated)
• /decisions — every research decision the engine has ever made