Research
Intelligence System
An AI engine that reads, sees and remembers the world's institutional research, and turns it into decisions you can source.
The edge is buried in ~950 pages of research. Every single day.
~58 reports/day × ~16.5 pages × ~3 min/page. September 2026 intake.
One engine that reads, sees, remembers, and argues back.
Six skills. One research brain.
From raw PDF to structured, sourced analysis in about a minute.
Not one model. An orchestra of them.
- Right model for each jobfrontier quality where judgment matters, speed and cost where it doesn't
- Health-aware routingautomatic failover across cloud and local providers
- Capacity gatespause, probe, resume, and never silently degrade quality
- Validated outputsevery answer schema-checked, repaired or rejected
- Model-agnosticthe next frontier model slots in without a rebuild
Production-grade. Kill it mid-run, and it just resumes.
Recoverable
Every stage checkpointed. Crash, kill or power cut: rerun and it picks up exactly where it stopped.
Deduplicated
Content fingerprints across the whole archive. The same note forwarded three ways is stored once.
Self-correcting
A QA pass repairs known failure classes, such as wrong dates or garbled sources, before a batch is called done.
Validated
Every model answer checked against a schema. Invalid output is repaired or rejected, never stored.
Transactional
Each report commits atomically. No half-written records, no orphaned charts.
Observable
Per-stage timings, quality flags and an automatic recap for every batch.
Five generations. Each one faster, cheaper, smarter.
Everything lands in one queryable research brain.
One brain, infinite outputs, now with agents on top.
Built for how a PM actually works.
Conviction heatmap · who is bullish on what, this week
Fed & Rates
Updated automatically as each new report lands
Finds meaning, not keywords
Self-contained, shareable, charts embedded
Publishing every trading day.
"Is the 30Y Treasury at ~5.10% a good investment?"
"The missing catalyst, confirmed labor weakness, is not a footnote; it is the gating condition."
A human signs off. The system learns from every sign-off.
Nothing leaves unreviewed
Every publication stops at a draft. Every number is checked against live market data before release.
Corrections become rules
A mis-flagged report today becomes an importance criterion tomorrow. Recurring issues become automatic fixes.
Memory compounds
House views are rewritten with every new report, and changes of mind are logged as divergences.
Grounded by design
Claims are tied to a source report. Numbers are grounded against data. Uncited claims are flagged.
September FOMC: 14 houses, 3 agents, one view.
49 notes · Sep 10–25, 2026
The Fed restarted hiking and 16 of 18 officials see another 2026 move. Terminal 4.25–4.50% by 1Q27 is credible while oil-driven inflation persists.
The dots are hawkish; the data isn't. Core PCE was revised down to 3.19%, already under the Fed's 3.4% Q4 path. One more move, then done.
~33bp is priced by year-end and the long end rallied on the hike: a credibility trade. First hikes lifted equities in 4 of 7 past cycles.
Oil near $106–110 is the swing factor. If it holds, the PCE revision won't last.
Base case 4.00–4.25% by year-end. The hawkish tail needs oil to stay above $105.
Would change our mind: core PCE re-accelerating above 3.4%.
Markets depth. AI breadth. Built to last.
Markets
- Reads 200+ research, insights and news sources
- Thinks in rates, flows, positioning & catalysts
- Publishes a daily market close
AI
- An orchestra of state-of-the-art models
- Vision that reads 65,000+ charts
- Agents that debate and form a view
Engineering
- 1,048 tests · recoverable · validated
- One queryable research brain
- Infinite outputs, always sourced
Every report read. Every chart seen. Every decision sourced.