← All work
AI Engineer2025

ReflexAI

ReflexAI pairs quantitative financial-risk diagnostics with retrieval-augmented AI reasoning grounded in George Soros's theory of reflexivity — analyzing how market feedback loops amplify financial fragility.

Live DemoGitHub
ReflexAI — AI stock and macro risk analysis platform
3
Data classes
3
Financial statements
5
Core capabilities
Annual
Data cadence
RAG
Reasoning layer
2025
Delivered

Try it

Live deployment

Running now

Pick a ticker. The app pulls the financial statements, runs the four risk diagnostics, and retrieves against the reflexivity corpus separately before it writes anything.

Open the app
Ask it
Whether a balance sheet survives stress
Not
Whether the stock goes up
Watch for
Where the two lanes disagree

The premise

A different question

Most tools answer

What happened?

This one asks

Why is risk building, and how might it evolve?

The loop it models

  1. Perceptions influence prices
  2. Prices influence fundamentals
  3. Narratives influence behaviour
  4. Behaviour feeds back into reality

Markets are not purely efficient. The premise is that these four statements form a cycle, and that the interesting risk lives in the cycle rather than in any single reading.

Three data classes

Different cognitive roles
  • 01

    Quantitative reality

    yfinance

    The baseline truth risk is assessed against

    • Annual income statement
    • Annual balance sheet
    • Annual cash flow statement
    • Lightweight market snapshot
  • 02

    Narrative and conceptual

    Soros knowledge corpus

    The lens the numbers are read through

    • Reflexivity as a framework
    • Boom and bust sequence structure
    • Feedback-loop reasoning patterns
  • 03

    Intent

    User queries

    What the analyst actually wants to know

    • Ticker-aware market context
    • Question routed against both layers above

Statements are facts. The corpus is a way of reading facts. Keeping them in separate retrieval paths is what stops a framework quote coming back as though it were a figure, and it is what makes the reasoning auditable.

Architecture

Two lanes, one join
ReflexAI architectureTwo independent retrieval lanes converging at synthesis. The quantitative lane pulls annual income statement, balance sheet and cash flow data from yfinance, normalises it, and computes four risk diagnostics: liquidity, leverage, profitability and narrative divergence. The conceptual lane holds a Soros knowledge corpus of reflexivity theory, boom and bust sequences and feedback-loop patterns, chunked and embedded for semantic retrieval. The two never mix during retrieval, which is what prevents a framework passage being returned as though it were a reported figure. A user query routes into both lanes, and only the synthesis step sees them together, producing a risk narrative that cites which statement drove it and which concept framed it. A feedback arrow returns from the narrative to the market snapshot, representing the reflexive loop the system models: perception influences price, price influences fundamentals, and the next reading starts from the changed state.QUANTITATIVE REALITYwhat the statements sayNARRATIVE AND CONCEPTUALhow to read themSOURCEyfinance· Annual income statement· Annual balance sheet· Annual cash flow· Market snapshotNORMALISEpandas· Line items aligned· Periods reconciled· Units standardised· Ready for diagnosticsDIAGNOSTICSRisk engine· Liquidity· Leverage· Profitability· Narrative divergenceCORPUSSoros knowledge base· Reflexivity framework· Boom and bust sequence· Feedback-loop patternsCHUNK + EMBEDVector index· Concept-level chunks· Embedded for retrieval· Framework, not figuresRETRIEVESemantic search· Top-k concepts· Scoped to the question· Kept separate from numbersUSER QUERY · INTENT LAYERticker-aware, routed to both lanesSYNTHESISRAG over both lanes· Numbers grounded in statements· Framing grounded in corpus· Cites which drove which· Explains the loop, not the levelTHE ONLY JOIN POINTRISK NARRATIVE· Where risk is building· Which loop is active· What would confirm it· Dashboard + statementsREFLEXIVE LOOP · perception moves price, price moves fundamentalsWHY ANNUAL DATAReflexive loops play out over years, not sessions. Annual statements strip the noiseintraday data would add. The cost is stated: nothing here answers a question about this quarter.WHY TWO LANESStatements are facts; the corpus is a way of reading facts. One combined index lets aframework passage return as though it were a reported figure. Separate paths stay auditable.
Two retrieval lanes, one join point, and the feedback loop that motivates the whole design.

Risk diagnostics

4 named dimensions

  • Liquidity risk

    Ability to survive stress without external financing, and cash adequacy against obligations.

  • Leverage risk

    How much of the balance sheet depends on borrowed capital holding.

  • Profitability risk

    Whether earnings quality supports the valuation being placed on it.

  • Narrative risk

    Where the story about a company has detached from what its statements show.

A current ratio does not tell you whether a company survives stress. Naming the risk makes the output something to act on rather than something to read.

Capabilities

5 layers

  1. 01Fundamentals layer

    Annual income statement, balance sheet and cash flow, normalised for downstream analysis.

  2. 02Risk diagnostics

    Liquidity, leverage, profitability and narrative risk, rather than generic ratio output.

  3. 03AI market reasoning

    Retrieval over the reflexivity corpus, so explanations are grounded in a framework rather than improvised.

  4. 04Ticker-aware context

    Market context attached to the specific company being examined.

  5. 05Pairs trading analysis

    Experimental. Included as a direction, not a finished capability.

Decisions

4 named tradeoffs
D-1

Annual data, deliberately

Chose
Annual statements only
Over
  • Quarterly filings
  • Intraday market data

Because

  • Avoids short-term noise entirely
  • Aligns with macro and structural analysis rather than trading signals
  • Reflexive loops play out over years, not sessions

Cost

Nothing here can answer a question about this quarter, by design.

D-2

Separate the corpus from the numbers

Chose
Two distinct data classes, retrieved separately
Over
  • One combined index of everything

Because

  • Statements are facts; the corpus is a way of reading facts
  • Mixing them lets a framework quote be retrieved as though it were a figure
  • Keeping them apart makes the reasoning auditable

Cost

Two retrieval paths to maintain and keep aligned.

D-3

Risk-focused diagnostics, not a ratio table

Chose
Four named risk dimensions
Over
  • The standard ratio set

Because

  • A current ratio does not say whether a company survives stress
  • Naming the risk makes the output actionable rather than descriptive

Cost

The dimensions are interpretive, so two analysts could weight them differently.

D-4

Explain the loop, not just the level

Chose
Model feedback between perception and fundamentals
Over
  • Report current values and stop

Because

  • Most tools answer what happened; the gap is why risk is building
  • A level tells you where you are, a loop suggests where it goes

Cost

Explanations are narrative, so they are harder to validate than a number.

Stack

Everything used

Reasoning

  • RAG
  • LLM synthesis
  • Retrieval over corpus

Data

  • yfinance
  • Annual statements
  • Market snapshot

Analysis

  • Python
  • pandas
  • Risk diagnostics

Framework

  • Reflexivity model
  • Boom-bust sequence

Interface

  • Structured dashboard

Experimental

  • Pairs trading analysis
See these across every project