Senior AI Orchestrator: LLM Routing to Python Quant Library

A hedge fund quant is awake at 2am. He types: "Give me 1s and 0s for when the Big 7 tech stocks are in regime 1 but the rest of the NDX 100 is in regime 0." When are the generals marching but the army isn’t following? You might assume this means aligning 100 equity time-series, handling missing bars, resampling to common frequency. It doesn’t. You call 100 functions. Same interface: price array in, binary state out. The alignment and warmup are encapsulated. Not your problem. Thirty seconds later he has a single boolean vector.

A macro trader is reading on a Sunday morning. She types: "Combine Gold, TLT (bonds), and VIX into a single fear index. Give me 1s and 0s." One number that tells her when all the fear assets agree. You might assume this means reconciling trading hours, holidays, futures vs ETFs. It doesn’t. Three function calls. Compose the outputs. Each function is stateless and self-contained. Same signature. Same behavior.

A crypto trader doesn’t trust short timeframes. He types: "BTC on 5-minute bars says 1. BTC on daily bars says 0. Flag every moment they disagree." When is the short-term lying about the long-term? You might assume multi-timeframe alignment and timestamp logic. It doesn’t require that. Two function calls. Compare outputs. The functions are agnostic to each other’s timeframe.

A pairs trader is building a position that doesn’t exist as a ticker. She types: "Long energy sector, short financials. Treat that spread as a single price. Give me regime." Regime detection on a synthetic position. You might assume synthetic instrument construction and spread pipelines. You compute the synthetic time series. Pass it to the function. The function is agnostic to the asset class. It just evaluates price action.

A quant blends government data with market data. He types: "Average the regime state of: FRED Financial Stress Index + Unemployment Claims + 10Y-2Y Yield Spread. Give me 1s and 0s." A single macro regime indicator from three government sources. You might assume weekly data mixed with daily, irregular timestamps, release date handling. Each function handles its own series independently. Compose the outputs.

A trader tests a stock with almost no history. She types: "This stock IPO’d 6 months ago. Only has 500 bars of 5-minute data. Give me regime." You might assume lookback requirements and edge case handling. The functions don’t require massive burn-in periods. 500 bars of history is enough. Same interface. Same behavior.

A quant runs across an entire index. He types: "Run regime on all 30 Dow stocks. Tell me how many are in regime 1 right now. Track that count as its own time series." A single number — how many soldiers are marching — updated every bar. You might assume thirty parallel pipelines and state management. Thirty parallel function calls. A cross-sectional sum of the binary states. Done.

A trader gets noisy output. She types: "The output is too noisy on 5-second bars. Accumulate it, run again, give me the regime of the regime." Recursive self-application until stable. You might assume custom smoothing and parameter tuning. Run the output through the same function again. No parameters. Loop until stable.

What exists: One hundred Python functions. Same interface: prices: array[float] → states: array[int]. No parameters. No tuning. Stateless. Self-contained. Temporal alignment encapsulated. Warmup encapsulated. Resampling encapsulated. Any asset. Any timeframe. Uniform interface. The math is frozen. The library is tested. The interface is uniform.

What you’re building: An LLM-driven orchestration layer between natural language and a deterministic Python library. Parse the query. Route to functions. Compose outputs. Evaluate stability. Loop if needed. Return results with version provenance.

What you’re not building: Not the math — frozen. Not the data pipeline — encapsulated. Not predictions — not the point. Not a chatbot — query in, results out.

To apply: Do not send a portfolio. Do not send past work. Answer these two architecture questions in your proposal. Proposals without these answers will be archived.

• Question 1: Three Python functions. Same signature: prices: array[float] → states: array[int]. A user types: "Run A and C. Average their outputs." Describe your approach to: Parsing the intent. Routing to the functions. Composing the outputs into a single valid array. No code. Just your reasoning in 4-5 sentences.
• Question 2: Same functions. A user types: "Too noisy. Run it again until stable." You need a recursive loop that evaluates output stability, decides whether to re-run, and terminates safely. Describe: How you mathematically define "stable" for a binary array. What your exact exit conditions are. How you prevent infinite loops. No code. Just your reasoning in 4-5 sentences.

Contract duration of 1 to 3 months.

Mandatory skills: Python, Large Language Model, AI Model Integration, ChatGPT API Integration, Model Deployment.