Agentic research infrastructurePublic and private markets
One workspace for agentic chat, governed ingestion, multi-factor frameworks, and portfolio-wide sweeps — grounded in filings, broker research, market data, and your own documents.
Sign in with your workspace account. Private libraries by default.
Illustration of a live research run: an analyst asks the Compounder Identifier v2 agent to score a company, seven evidence agents gather sources in parallel, scripts compute a 40-factor matrix, and the result is a 78 out of 100 scorecard with a full citation trail.
Score NVDA's compounding durability as of 2025-06-30.
sec-edgaryfinancedefeatbetaragEvidence-linkedGrounded in the systems your desk already trusts
SEC EDGAR, EDGAR 13F, YFinance, DefeatBeta, StockAnalysis, AKShare, TWSE, Firecrawl, Gangtise, RAG corpus, arXiv, Synology, Lark, Notion, YouTube
FR Copilot connects the conversational layer to a governed data plane. Every surface shares the same evidence layer, agent engine, and audit trail.
Threads keep their context, attachments, and citations. Switch agents, models, or knowledge scope mid-conversation without losing the trail.
Compounder v2 and Fallen Angel run evidence agents in parallel, compute with scripts, and reserve LLM judgment for the factors that need it.
Documents, arXiv papers, broker reports, YouTube, audio, and Synology folders are parsed, tagged, and indexed for retrieval.
Queue any agent across tickers and as-of dates. Stages run in order, every run is durable and retryable, and the sweep exports to CSV.
Personal libraries, group workspaces, and a governed shared corpus — searchable together or scoped per thread.
Compose orchestrators and subagents, attach MCP servers and skills, publish an immutable version, and fork from the gallery.
Schedule a thread with /schedule, refresh coverage with stock pools, or queue a portfolio sweep. The work runs on your cadence.
Research does not stop at the answer. Move from a question to a thread, from a thread to a knowledge library, from one name to the whole portfolio.
Score NVDA's compounding durability and flag what would change the thesis.
10-K FY25Q2 call14 brokersKeep the question, the files, the tools, and the evidence in one thread. Switch agents or scope without losing the trail, then take the answer out as a chart, a scorecard, or a share link.
Two research frameworks turned into repeatable, auditable runs. Same engine, opposite questions: what compounds, and what has fallen far enough to turn.
Forty orthogonal factors across ten categories. Evidence agents gather; scripts compute peer baselines and factor inputs; LLM subagents adjudicate the rest against a numeric rubric.
Thirty-six factors and a hard-gate screen separate a temporary mispricing from permanent impairment. Point-in-time discipline keeps the evidence honest.
Attach MCP servers, tools, and skills on a canvas. Publish immutable versions, share them with a group or the workspace, and fork anything from the gallery.
The same loop powers every packaged agent, custom agent, scheduled task, and sweep. Nothing important is left to a black box.
Resolve the issuer, market, and as-of date. Lock the point-in-time basis before any data is read.
Issuer · market · dateEvidence subagents run in parallel across filings, transcripts, broker research, market data, and the web.
Parallel evidence agentsScripts align history, normalize peer baselines, and prepare every measurable factor input.
Deterministic mathLLM subagents score judgment-heavy factors against a named rubric. A challenger audits the call.
Rubric + challengerReview the scorecard, override where warranted, then export, share, schedule, or run it across a sweep.
Analyst gateResearch infrastructure has to survive review. Access, versions, usage, and automation are first-class parts of the workspace.
Owner-scoped threads, immutable published agent versions, and admin-visible agent traces. Every output keeps the path back to its evidence.
Keep sensitive work in a private library, collaborate inside a group, and publish only what belongs in the shared corpus.
Model pricing is visible before a run, and every charge lands in a usage ledger. Buy credits once; spend them across chat, sweeps, and schedules.
Drive ingestion, headless chat, sweeps, and agent authoring from your own systems. Create a thread, attach files, ask a question, and download the artifacts — all with a per-user key.
bashcurl -X POST "$BASE/copilot/api/ask" \
-H "Authorization: Bearer qrc_your_key" \
-H "Content-Type: application/json" \
-d '{
"question": "Score NVDA compounding durability",
"thread_id": "$THREAD_ID"
}'Bring a question, a document batch, or a portfolio. The evidence layer, the agent engine, and the audit trail are already connected.
Already have an account? Sign in from the workspace. New users join by invitation.