Skip to content
FINQUILL://terminal

The Investment
Reasoning Engine

Your conviction: |

Falsifiable theses
Adversarial red-team
A graded record
THE ENGINE

Markets don't
compile
.

Vibe coding works because code has a compiler — the machine tells you, instantly and without mercy, when you're wrong. Markets tell you slowly, noisily, and sometimes lie for years. Finquill is the missing toolchain for conviction: capture the thesis, arm its kill criteria, red-team it before you size it, monitor it while you live your life, and grade the record when reality reports back.

01CAPTURE

Say it in plain English. The engine structures it into falsifiable claims with deadlines.

02STRESS-TEST

A panel of adversarial agents tries to kill it. Four lenses; default verdict: refuted.

03MONITOR

Always-on graphs watch your kill criteria against news, filings, earnings, and your own trades.

04RESOLVE

Claims come due. Reality grades them.

05RECORD

Brier-scored calibration plus process grades: was the decision right given what was knowable?

AI Response
grounded in your book

Thesis structured. 3 falsifiable claims: (1) top-3 hyperscaler capex guidance holds or rises through FY27 — resolve by each earnings; (2) Blackwell lead times stay >6mo through H1; (3) AI-accelerator mix >35% of DC revenue. Kill criteria armed — I'll flag guidance cuts >10% the morning they print.

Thesis captureFalsifiable claimsKill criteria
View ChartsFull ResearchGet Started
TRADING ON VIBES
A hunch from a headline
Conviction that hardens with the position
No record of why you entered
Hindsight rewrites the story
Zero learning per trade
REASONING WITH FINQUILL
A falsifiable thesis with a deadline
Red-teamed before it's sized
Kill criteria armed and watched
Graded on reasoning and outcome
Every trade teaches
Live demo · public mode

What our AI builds for you, in seconds.

Ask any ticker, any thesis. The platform composes a custom surface — KPIs, bull/bear cases, comparable history — that you can interrogate, refine, save, or share.

PLATFORM//SURFACES

Everything you need.
Nothing you don't.

Fifteen integrated surfaces on one reasoning engine. Built AI-native from the first commit.

Dashboard

What changed overnight for what you hold and watch — a belief-delta digest, not a quote wall.

Belief-delta view

Watchlist

Every name carries a live thesis. Always-on graphs wake on news, earnings, and your own trades.

Always-on monitors

Journal

A coach, not a diary — post-event interviews that separate your edge from your tilt.

Edge vs. tilt

Research

Cited theses with testable claims. Every number from a tool; every quote traceable to a source.

Citations everywhere

AIStudio

Ten specialists behind one conversation — a planner fans work out in parallel; one voice synthesizes.

10 specialists

Trading

Preview → confirm on every order. Typed confirmation when it's live and the size is real.

Safety-gated

Monitor

The world on one lens, your exposure on the other — quakes, storms, chokepoints, sanctions, touching your book.

Dual-lens OSINT

Calendar

Earnings, dividends, and macro events — with your exposure and your theses attached.

Exposure-aware

Explore

Public dossiers, trending research, and the firm's own graded record.

Open door
15
Surfaces
10
AI Specialists
35
Tool Modules
PUBLIC
The record → /board
AI//Classical ML

Some questions aren't language problems.

Ask a language model for a Sharpe ratio and it will give you a confident, plausible-sounding number — computed from nothing but the shape of your question. Finquill trains its agents to catch the difference: when a question is actually a statistics problem, it gets handed to a real model, not answered from the language model's memory. Eighteen classical ML and statistics tools are bound directly to the agents — extending what the LLM can do, not standing in for it. The LLM writes the sentence. It never invents the number.

Would this strategy have worked?
Routed to
Backtest engine
Cost-aware · in/out-of-sample

A moving-average crossover, tested on data it never saw. The strategy is picked using only the training window — the test window plays no role in the pick, so the result can't cheat.

What's the probability this happens?
Routed to
Calibrated classifier
XGBoost, Platt/isotonic-calibrated

A trained model, calibrated against data it wasn't trained on — so a 70% call actually resolves around 70% of the time. Ships with its own reliability curve and Brier score, not just a headline number.

What's volatility likely to do?
Routed to
GARCH & statistical tests
Volatility, cointegration, stationarity

The tools a quant desk reaches for — GARCH volatility forecasts, rolling beta, cointegration tests — run directly by the agent, not approximated by a language model guessing at the math.

Every one of these tools has a floor — ask for a probability on six months of data, and the answer is a refusal, not a guess. Try it in AI Studio, or ask the trading specialist directly.

PROOF//OUR OWN BOOK

We run it on ourselves.
In public.

An autonomous research firm runs this same engine on its own coverage — publishing beliefs pre-market with kill criteria attached, admitting changes of mind, and grading the record where anyone can check. Editions are hash-committed.

Scoreboard
withheld
12 resolved · floor is 20
Today's Tape
Daily Tape — 2026-08-17: AAPL, GOOGL, MSFT
2026-08-17 · pre-market edition
Receipts
Every edition hash-committed at publish. Tamper-evident, verifiable by anyone — no trust required.
RESEARCH ENGINE

Integrated research
environment.

Our intelligent RAG pipeline ingests SEC filings, earnings transcripts, news, and market data — then synthesizes it into actionable research you can query in natural language.

Multi-Source Ingestion

Continuously indexes SEC filings, earnings calls, analyst reports, and news into one queryable corpus.

Contextual Retrieval

Semantic search across your entire research corpus — ask a question and get answers grounded in real source documents.

AI Synthesis

Generates investment theses, competitive analyses, and risk assessments with full citation trails back to source material.

research / NVDA-thesis.mdCITED
Investment Thesis — structured for review

NVDA: data-center capex sustainable through FY27

Captured in plain English, then decomposed into claims that can actually be settled. Every figure is read out of a filing or transcript at generation time — a number the tools can't retrieve is refused, never estimated.

[1]
Top-3 hyperscaler capex guidance holds or rises
Resolves each earnings · 10-Q + call transcripts
[2]
AI-accelerator share of data-center revenue holds above its level at capture
Resolves quarterly · segment disclosures
[3]
Competing ASIC ramp does not compress realized ASPs
Resolves FY27 · your uploaded research
Kill criterion — armed
A combined guidance cut of more than 10% flags this thesis the morning it prints.
Every claim
Source-linked
AI//Prompt library

Think like Buffett. Or Dalio. Or Munger.

Finquill started as a prompt library — years of engineered frameworks that let you run any question through the minds of the investors you learn from. Switch lenses by tag, or stack multiple frameworks on one query.

Buffett & Munger
Durable moats & quality

Screen the way Berkshire does — moat strength, owner earnings, returns on tangible capital.

Example prompt
$Score my watchlist on Buffett-Munger quality criteria; flag any name where the moat has narrowed in the last 4 quarters.
Howard Marks
Second-level thinking

Look past the consensus. What does everyone already know, and where is the non-obvious read?

Example prompt
$What's the consensus view on NVDA right now, and what's the second-level read a careful contrarian would build from the same facts?
Ray Dalio
All-weather & regime balance

Pressure-test portfolios against growth, inflation, and deflation regimes.

Example prompt
$Score my portfolio for factor balance across growth, inflation, and deflation regimes — flag any regime exposure above 40% and suggest rebalancing.
Peter Lynch
Invest in what you know

Simple story, growing earnings, boring industry — the ten-bagger pattern, systematized.

Example prompt
$From the names on my watchlist, which match the Lynch 'simple story / growing EPS / under-covered' pattern? Rank by how closely they fit.
Ben Graham
Margin of safety

Defensive-investor criteria with today's data — quantitative, uncompromising.

Example prompt
$Run a Graham defensive-investor screen on my holdings. Flag any name outside his quantitative criteria with the specific line item that fails.
Michael Burry
Contrarian + catalyst thesis

Short-setup discipline: where is sentiment wrong and what's the catalyst to close the gap?

Example prompt
$Map the consensus bull case on CVNA, identify the three strongest counterpoints, and rank each by how catalyst-able it is inside 12 months.

Build your own. Codify your investment philosophy as a reusable prompt once, then apply it to every new name you research.

Library grows every release
AI//Context × model fit

Your context, routed to the right model.

Finquill enriches every query with your watchlist, economic calendar, and media feed — then picks the frontier model whose strengths fit the question. Different questions want different models.

Your watchlist + news feed
Routed to
Claude
Reasoning

Cross-portfolio implications of today's CPI print — reasoned through each of your names with their individual rate-sensitivity.

10-Ks + transcripts you've saved
Routed to
Gemini
Long context

Side-by-side comparison of the last four filings for three names you're researching — all in a single context window, nothing summarized away.

Your media feed + social signals
Routed to
Grok
Social zeitgeist

Live sentiment pulse on the meme-adjacent names in your watchlist, with retail positioning and social chatter refreshed in real time.

Override the routing any time — pin the model you want, or swap mid-conversation if a query needs something different.

AI//Model choice + frontier speed

You decide which model touches your work.

Open-source for full data sovereignty. Frontier closed-source for max capability. Diffusion models when you want ten draft theses to compare in seconds. Your call, query by query.

Your choice of model

Every major frontier model. Open and closed source.

Pin a specific model for a particular workflow. Default to the strongest available frontier for everything else. Switch mid-conversation if a query needs more horsepower.

OpenAI
Anthropic
Google
Meta · Llama
Mistral
xAI
Proprietary · in-house
Thorp— our in-house quant model, named after Edward Thorp. A fine-tuned LLM combined with TimesFM— Google's foundation model for time series forecasting — to ground probabilistic reasoning in real market structure. Currently in training; early-access users get first access as it ships.
Built for speed

Answers at the speed of thought.

We route to optimized inference compute (including Groq for ultra-low-latency models) so iterative research doesn't turn into a coffee break between every question. The faster the loop, the more questions you ask.

Inference latencylive sample
Frontier closed-source1240ms
Open-source · standard720ms
Optimized compute180ms

Managing client assets? Finquill has a dedicated advisor product — cited research, household-level portfolio reviews, and human-in-the-loop approval built for RIAs.

Explore Finquill for Advisors
$ finquill --init

Stop trading on
vibes
.

Capture your first thesis in plain English. The engine makes it falsifiable, red-teams it, and starts your record. Belief capture is never metered.

Free to start
Belief capture never metered
Analytics, not advice
Early access — actively onboarding new users