Atishay Kasliwal — Software Engineer, production AI and distributed systems.

Applied ResearchResearchNov 2024 — Present

FOMC Intelligence

NLP pipeline turning Federal Reserve communications into market signals.

Graduate Research Assistant, Stony Brook University

200K+
data points processed
3.7K+
time-series intervals
<300ms
P99 API latency
27%
simulated portfolio return

The problem

Federal Reserve language moves markets, but the signal is buried in hedged prose. The interesting question is not what the FOMC said — it is what changed since the last statement, and whether that delta is priced in.

Approach

Treat each statement, minutes release, and speech as a document in a time series. Extract sentiment and stance, diff against the prior release, and align the result to market data on the same intervals.

Architecture

  1. Ingestion: AWS Lambda pulls Fed publications and financial news on a schedule, landing raw documents in S3 with schema validation at the boundary.
  2. Processing: transformer-based sentiment and stance extraction, plus statement-over-statement diffing to isolate what actually changed.
  3. Storage: DynamoDB holds structured per-interval features across 3.7K+ time-series intervals.
  4. Serving: FastAPI exposes the feature store to downstream consumers at sub-300ms P99.
  5. Observability: Prometheus and CloudWatch track ingestion lag and pipeline failures.

Design decisions

Event-driven ingestion over a monolithic batch job

Fed publications arrive irregularly and the market reaction window is short. Lambda triggers keep end-to-end lag low and make partial failures recoverable per-document instead of per-run.

Diff against prior release rather than scoring in isolation

Absolute sentiment on central-bank prose is nearly constant — it is engineered to be. Almost all the information is in the change, so the pipeline is built around comparison, not classification.

Schema validation at ingestion, not at read time

Upstream formats change without warning. Failing loudly at the boundary keeps malformed documents out of the feature store, where they would be far more expensive to detect.

What was hard

  • Backtests are trivially self-deceiving. Any signal computed with information unavailable at decision time produces spectacular and meaningless returns, so the pipeline enforces point-in-time correctness on every feature.
  • Sustaining zero data loss under bursty load when several publications land at once.
  • Fed language is deliberately hedged; off-the-shelf sentiment models trained on consumer text score it as uniformly neutral and had to be adapted.

What I took from it

  • The modeling was the easy part. Correct point-in-time data plumbing was most of the work and all of the risk.
  • Reported returns from a simulation are a statement about the backtest, not about the market — and the site says so rather than implying a track record.

Stack

  • Python
  • FastAPI
  • AWS Lambda
  • S3
  • DynamoDB
  • NLP
  • Transformers
  • Prometheus

Other work

  • Product

    Atriveo

    Job-search platform with a 5.0-rated Chrome extension and live customers.

  • AI Systems

    Legal RAG

    Retrieval-augmented generation over legal filings, with citations that hold up.

  • Machine Learning

    MRI Tumor Viewer

    Brain tumor segmentation running entirely in the browser.