Projects

Problem

Industrial sensors produce massive amounts of high-velocity time-series data. Traditional threshold-based monitoring fails to detect subtle, compounding failure states early enough to prevent downtime.

Approach

Instead of relying only on rigid rules, I approached this as a prototype for continuous analysis. The goal was to establish a moving baseline for normal behavior and flag deviations earlier than static thresholding.

System Design

• ▶Input: Raw industrial sensor metrics from simulated or streaming sources.
• ▶Processing: Preprocessing, sliding-window analysis, and anomaly scoring against changing baselines.
• ▶Output: Severity-based anomaly alerts for monitoring workflows.

Engineering Impact

• Structured preprocessing, scoring, and alerting into separate pipeline stages.
• Compared adaptive baseline logic against static-threshold monitoring.
• Designed outputs to support reviewable, severity-based monitoring decisions.
• Kept the prototype modular enough to refine individual steps independently.

Problem

Teams often have valuable structured data but still rely on manual query writing or analyst support to answer straightforward business questions.

Approach

I approached this as both an NLP and workflow problem. The goal was to turn plain-language questions into structured data exploration while keeping the request understandable, debuggable, and usable for non-technical users.

System Design

• ▶Input: Natural language questions, table context, and structured datasets.
• ▶Processing: Intent extraction, schema-aware query construction, and output formatting.
• ▶Output: Structured queries and readable result views for downstream analysis.

Engineering Impact

• Separated intent understanding, query generation, and output formatting into distinct steps.
• Made structured data access more usable for non-technical users.
• Kept request structure explicit for easier debugging and iteration.
• Designed outputs to be reviewed and refined rather than treated as a black box.

Problem

Financial analysts spend excessive time aggregating disparate market data and extracting insights from unstructured reports to make investment decisions.

Approach

The goal was decision support, not replacement. I explored how fragmented financial data could be structured into clearer signals, iterative forecasts, and usable summaries.

System Design

• ▶Input: Historical market data, live signals, and structured financial sources.
• ▶Processing: Pattern analysis across time-series data with iterative refinement.
• ▶Output: Dashboard-style insights, forecasts, and structured decision support.

Engineering Impact

• Organized fragmented inputs into repeatable processing and summary steps.
• Focused on readable, review-friendly outputs rather than opaque predictions.
• Explored iterative analysis across multiple time horizons.
• Framed the system as decision support that could be extended and improved over time.