Pascal MCP Server

The Fund's AI Brain

Pascal connects your internal notes, broker reports, expert calls, filings, and financial data into one entity-resolved, temporally-aware intelligence layer — exposed as MCP tools to any AI agent.

5

Source types connected

860+

Companies covered

10

MCP tools exposed

1

Canonical entity per company

Analyst Notes

Internal memos

Broker Reports

Sell-side research

Expert Calls

Transcripts

Public Filings

SEC / exchange

FactSet

Consensus data

Pascal MCP Server

Doc Intelligence Entity Master & Ontology Temporal Indexing Security & RBAC Orchestration Harness

Any AI Agent

Any LLM · Any Agent · Any Client

The Problem

Same company, different names across every source

Six documents about NVIDIA from five source types. Each system uses a different name. No system connects them.

NVDA NVIDIA Corporation — 6 documents, 5 sources

Expert Call · Former Amkor VP (Apr 15)

Transcript says "Nvidia" and "Blackwell". Tagged under "Advanced Packaging Trends" — not NVDA. Contains Blackwell order data 30-40% above Street estimates. Critical supply chain signal invisible to any NVIDIA search.

✗

Internal Analyst Note (Mar 25)

Analyst writes "NVDA". Broker report says "NVIDIA Corp". Filings use "NVIDIA Corporation". Earnings transcript references "Jensen Huang". Four names for one company — four separate searches.

Searching any single name misses documents from every other source.

Temporal Awareness Problem — Affects Every Query

Query: "Give me the latest thesis on NVIDIA"

Two analyst notes exist: Feb 15 (pre-earnings, "if GM below 72%, downgrade") and Mar 25 (post-GTC, "overweight, GM 71-72%"). A generic search returns whichever scores highest on semantic similarity — which may be the older one.

✗

Why it matters

The analyst changed their view after earnings. Without pre-indexed timestamps, "latest" is guessed by the LLM. The wrong note gets returned — silently. Same query on two days may return two different notes.

Without pre-indexed timestamps, every "latest" query is a guess.

Cross-Source Blind Spots — Signals Your Own Data Contains

Expert Call: Packaging exec mentions Broadcom

Broadcom winning custom ASIC deals from hyperscalers — "the real threat, not AMD." This signal exists only in one expert call. Not in any analyst note or broker report.

✗

Result

Analyst thesis focuses on AMD as the competitor. Broker reports agree. The Broadcom signal — from your own licensed data — is never surfaced because it lives in a call not tagged to NVIDIA and not cross-referenced to the analyst's coverage.

Cross-document discovery is impossible when each source is searched in isolation.

The Foundation

Entity Master — Built at Ingestion, Before Any Query

Every document ingested gets entities extracted and resolved to a canonical form. Name variants, timestamps, and source metadata are structured at ingestion time — not guessed at query time.

Canonical Entity	Name in Source	Type	Source	Document	Date
NVIDIA Corporation	NVDA	Company	Analyst Notes	Post-GTC Thesis Update	Mar 25
NVIDIA Corporation	NVIDIA Corp	Company	Broker Report	Blackwell Ramp (MS)	Apr 10
NVIDIA Corporation	Nvidia	Company	Expert Call	Former Amkor VP	Apr 15
NVIDIA Corporation	Jensen Huang	Company	Public Filing	Q4 FY26 Earnings	Feb 20
NVIDIA Corporation	NVDA	Company	FactSet	Consensus Estimates	Current
NVIDIA Corporation	NVDA	Company	Analyst Notes	Pre-Earnings Positioning	Feb 15
Blackwell	B300	Product	Analyst Notes	Post-GTC Thesis Update	Mar 25
Blackwell	Blackwell GPU family	Product	Broker Report	Blackwell Ramp (MS)	Apr 10
Blackwell	B200/B300	Product	Expert Call	Former Amkor VP	Apr 15
Microsoft	MSFT	Company	Analyst Notes	Post-GTC Thesis Update	Mar 25
Microsoft	Azure	Company	Broker Report	Blackwell Ramp (MS)	Apr 10
Broadcom New	Broadcom	Company	Expert Call	Former Amkor VP	Apr 15
SK Hynix New	Hynix	Company	Expert Call	Former Amkor VP	Apr 15

"Jensen Huang" → NVIDIA Corporation and "Azure" → Microsoft — these resolutions happen at ingestion, not query time. The "Name in Source" column shows exactly why cross-source search breaks without this table.

The Foundation · Continued

All Sources Per Entity — What Pascal Queries Against

Every alias extracted, every document linked, every source tracked — grouped by canonical entity. This is the structure that makes cross-source queries work.

NVDA

NVIDIA Corporation

6 documents across 5 sources

Aliases NVDANVIDIA CorpNvidiaNVIDIA CorporationJensen HuangNVDA US Equity

Analyst NotesLatest: Mar 25

Post-GTC Thesis Update (Mar 25) — Overweight, GM 71-72%
Pre-Earnings Positioning (Feb 15) — "If GM <72%, downgrade"

Broker ReportsLatest: Apr 10

Blackwell Ramp Faster Than Expected (MS, Apr 10) — GM 73%

Expert CallsLatest: Apr 15

Former Amkor VP (Apr 15) — Blackwell 30-40% above Street

Public FilingsLatest: Feb 20

Q4 FY26 Earnings Transcript — GM 73.5% actual

FactSetCurrent

Consensus: Q1 GM 72.8%, Q2 GM 73.1%, Rev $45.0B

Blackwell

Product

AliasesB300B200/B300Blackwell GPU family

Broadcom

Discovered

SourceExpert Call onlyASIC competitive threat

SK Hynix

Discovered

AliasesHynixHBM3e supplier

Broadcom and SK Hynix were discovered from the expert call — entities that don't appear in any analyst note or broker report. Cross-document discovery, not just cross-document search.

Case 1 · Entity Resolution

Six Names, One Entity — Resolved at Ingestion

"Show me all research on NVIDIA" NVDA

Without Pascal

✗Analyst note uses "NVDA" — found only if you search that exact string

✗Broker report says "NVIDIA Corp" — different string, separate search

✗Expert call says "Nvidia" in a call tagged "Packaging Trends" — not NVDA. Blackwell supply signal missed entirely.

✗Earnings transcript references "Jensen Huang" — no generic search maps a person to a company ticker

Six documents about NVIDIA. A single search returns 1 or 2 at best.

With Pascal

✓"NVDA," "NVIDIA Corp," "Nvidia," "Jensen Huang" all map to NVDA in the entity master

✓Expert call content extracted at ingestion — "Blackwell" mapped to NVDA. Call is linked regardless of its original tag.

✓One query returns all 6 documents across all 5 source types — unified view with citations

All 6 documents retrieved. Full cross-source synthesis.

Case 2 · Cross-Source Discovery

Competitive Signals Hidden in Your Own Data

"What competitive threats to NVIDIA has our analyst not flagged?" NVDA

Without Pascal

✗Analyst note mentions AMD as the competitor. Broker report agrees.

✗Expert call from a packaging exec says Broadcom is winning custom ASIC deals — but this call isn't tagged to NVIDIA

✗SK Hynix HBM3e allocation data exists in the same call — another supply chain signal invisible to any NVIDIA search

Thesis focuses on AMD. The Broadcom signal — from your own licensed data — is never surfaced.

With Pascal

✓Doc intelligence extracted Broadcom from expert call content and linked it to the NVIDIA entity graph as a competitive signal

✓SK Hynix extracted and linked — HBM3e fully allocated to Blackwell. Supply chain intelligence from a single source.

✓Cross-source comparison: analyst + broker focus on AMD. Expert call says "the real threat, not AMD" — flagged as a single-source signal not present in sell-side research.

Broadcom surfaced as a competitive threat your analyst hasn't flagged. Cross-document discovery.

Case 3 · Temporal Awareness

Two Analyst Notes — Knowing Which One Is Current

"Give me the latest thesis on NVIDIA" NVDA

Without Pascal

✗Two notes exist: Feb 15 ("if GM below 72%, downgrade to neutral") and Mar 25 ("overweight, GM 71-72%")

✗Semantic similarity may rank the Feb 15 note higher — it contains the phrase "NVIDIA thesis" more prominently. Silently returns the wrong one.

✗"Latest" has no formal meaning. Non-deterministic — same query may return different notes on different days.

Analyst builds on a superseded note without knowing it was updated.

With Pascal

✓list_filing_periods(NVDA) returns actual dates from DB: analyst notes Feb 15 and Mar 25.

✓Mar 25 selected as latest. Feb 15 flagged as Superseded — analyst updated their view post-earnings.

✓Deterministic and auditable. Same query, same date, always returns the same result.

Latest note returned with superseded detection. No guessing.

How the table is used: The Date column in the entity master is indexed at ingestion. list_filing_periods queries this column directly — returning real database dates, not LLM-inferred ones. "Latest," "pre-earnings," and "last quarter" all resolve to actual timestamps.

Query Resolution · Step by Step

How Pascal Decomposes and Routes a Research Query

Each step handles one responsibility. Entity resolution → temporal grounding → per-source retrieval → synthesis with conflicts surfaced.

"What are the latest thesis points on NVIDIA?"

NVDA

1

resolve_user_context

"NVIDIA" → canonical entity across all 6 name variants

2

list_filing_periods TEMPORAL GROUNDING

Anchor "latest" to real database dates — no hallucinated time periods

3

plan_search(analyst_notes, latest=true)

Two analyst notes exist (Feb 15, Mar 25). Timestamp selects Mar 25. Feb 15 is superseded.

4

plan_search(broker_reports) + plan_search(expert_call)

Broker: Morgan Stanley Apr 10. Expert: Amkor VP Apr 15 — Broadcom signal not in sell-side.

5

search_public_filings + query_financial_data

Q4 FY26 earnings (Feb 20) + current consensus from FactSet

6

synthesis CONFLICT DETECTION

Compile answer — surface disagreements, attach timestamps and citations per source

The Differentiator

Why Not Just Connect Horizontal LLMs to Each Source Directly?

Individual MCPs work within one source. They break the moment a query crosses two.

Capability	Individual MCPs	Pascal MCP
Entity resolution	"NVDA" ≠ "NVIDIA Corp" ≠ "Nvidia" ≠ "Jensen Huang." Four MCPs treat these as four different entities.	Entity master maps all 12 name variants to one canonical entity. Deterministic lookup at ingestion.
Cross-source discovery	Expert call about packaging — not tagged NVDA. NVIDIA search returns nothing. Broadcom competitive signal invisible.	Doc intelligence extracts "Blackwell" → mapped to NVDA. Broadcom and SK Hynix surfaced from content, not tags.
"Latest" queries	Semantic similarity may return a 6-month-old note. "Latest" is guessed by the LLM, non-deterministically.	Timestamps indexed at ingestion. Feb 15 note superseded by Mar 25. Deterministic and auditable.
Conflict detection	Four sources disagree on gross margin (71-72%, 73%, 73.5%, 72.8%). No system connects them.	Conflicts flagged across all sources. PM sees 4-source comparison with provenance on every number.
Context window	Must dump entity mappings into context on every query. Quality degrades past ~150k tokens.	All metadata pre-indexed. Resolution is a lookup — zero context overhead per query.
Security	N connectors = N governance surfaces. Each manages its own permissions, identity, and audit.	One security boundary. Unified RBAC, PII redaction, audit logging, coverage restrictions.

The Brain After 6 Documents

Built automatically at ingestion. No analyst configured it.

11

Entities Resolved

12

Name Variants Mapped

5

Sources Connected

3

Conflicts Detected

2

Entities Discovered

1

Superseded Version

Six months later, the analyst leaves. A new analyst joins.

Day 1: they query "NVIDIA thesis." Pascal returns the full picture — 3 active conflicts, the Broadcom competitive signal from a packaging exec, temporal versioning across analyst notes. Every connection the previous analyst built over months is inherited instantly.

The data layer compounds today. The memory layer is next. Together, that's institutional memory that doesn't walk out the door.