How it works · Technical overview

An analyst that shows its work.

Give DeepSieve a handful of names or URLs. It returns a structured intelligence table covering every institution, fund, and decision-maker it found — where every single cell can prove where it came from.

Your seed list
kkr.com/about
Tiger Global
Howard Marks
Deep research engine
Verified table
12 src
9 src
17 src
10
orchestrated stages in every run
2
passes: research, then verification against the source
Per cell
verdict — supported, contradicted, or not checked
15–60
minutes for a standard run
01Identify & anchor

First, it learns who you mean.

Paste URLs or simply type names — “KKR & Co” works as well as a link. A classifier decides what each seed is, using the entity types your own Blueprint defines. Anything you paste is read first, so the research starts from what you already know rather than from a guess about who you meant.

stage:scraping · classifying_seed · seed_enrichment
https://www.kkr.com/aboutcompany
Tiger Global Managementcompany
Howard Marksindividual
Known facts — verified before research begins
Seed read · kkr.com · 3 entities identified
Blueprint matched · institution · dedup passed
02Autonomous investigation

Then it reads the web — properly.

For each entity, an autonomous research agent plans its own investigation: what to ask, where to look, what to read next. It works the public web — company sites, press coverage, announcements — with entities researched in parallel, and writes a long-form, citation-rich dossier as it goes. If a research stream ever goes quiet, the orchestrator reconnects and presses on; it never quietly settles for a thinner model.

stage:investigating · investigating_entity
Live activity · one entity of many, in parallel
planmapping KKR’s fund complex & key people
search“KKR Asian Fund IV final close size”
readinvestors.example.com/news/q4-results…
readkkr.com/businesses/kkr-asia
donedossier complete — 47 sources cited
03Structured extraction

Findings become a table, not an essay.

A second pass reads each dossier and extracts it into structured records — institutions, funds, people, and how they connect — against a strict schema. Empty fields aren’t shrugged off: targeted follow-up searches go out to fill the gaps, field by field, until the table is as complete as the public record allows.

stage:extracting · gap_filling · linkedin_research
Entity
Type
AUM
HQ
KKR & Co. Inc.
company
$553B
New York
Asian Fund IV
fund
$15.0B
Singapore
↳ gap detected → targeted search → filled
04Evidence & verdicts

Every cell keeps its receipts.

Nothing lands in your table as a bare claim. Every cell is written to the database alongside evidence rows — the exact source URLs that back it. A second pass then re-reads those sources against the claim and marks each cell: supported, contradicted, corrected, or simply not checked. You are told which is which, in the table.

stage:persisting · qa_verifying
Companies · AUM · KKR & Co. Inc.
$553B
company newsroom · results announcement
kkr.com · Q4 earnings release
ft.com · coverage, Mar 2026
evidence attached · 3 sources source captured · press release
05Independent verification

A second AI checks the first one’s work.

Finally, a separate QA agent — one that played no part in the research — re-fetches the cited sources and asks a single question: does this page actually support what the table says? Every claim receives a confidence score and every entity a relevance score against your brief. The dials on the report page are this pass’s verdict.

stage:qa_verifying
Scored by an agent that didn’t write the answers.
Under the hood

And for the AI-native in the room…

langfuse · trace: deep_research.run
deep_research.runone trace per run · session_id = run_id
├─ investigateagentic SSE stream · idle watchdog + reconnect
├─ extractschema-constrained JSON · bounded retries
├─ persisttransactional UPSERT · per-cell evidence rows
├─ enrichmentPublic-web research · every cell source-linked
└─ qaindependent verification · confidence + relevance
No silent downgrades
A stalled stream is probed and reconnected on the same interaction — it is never swapped for a thinner model.
Parsed by contract, not hope
Extraction is schema-validated structured output with bounded retries, not free-text parsing.
Idempotent persistence
Transactional UPSERTs converge on re-runs instead of duplicating records.
Fully observable
Every search, token, and verification step lands in a single end-to-end Langfuse trace.

Research you can audit.

The output isn’t a chat answer — it’s a table you can defend in a meeting, with the sources to prove it.

Standard run · 15–60 min · every cell cited