Experiments and forecasts that teams actually use

Problem

Product and operations teams wanted to test more ideas, but underpowered tests and slow, inconsistent readouts meant results often arrived too late or too vague to act on.

Approach

Standardized experiment design around a clear hypothesis, a primary metric with guardrails, and power and sample size set before launch. Built a GenAI-assisted tool that recommends test designs, interprets A/B results, and writes a plain-language summary. Alongside it, built Bayesian time series forecasts of key payments metrics in SQL, PyMC, and statsmodels.

Result

Product and operations teams adopted the experimentation tool, Finance and Strategy use the forecasts for planning and resource allocation, and readouts now go straight into roadmap and decision reviews.

• SQL
• Experiment design
• Power analysis
• Bayesian forecasting
• Tableau

An agentic copilot for financial operations

Problem

Analysts spent hours reading filings, transaction logs, and dispute records to answer routine compliance questions, and a single-chain LLM setup broke midway too often to trust.

Approach

Designed a multi-agent system in LangGraph: a planner breaks the question down, an executor calls tools and retrieval, and a validator checks every answer before it returns, with deterministic routing and recovery when a step fails.

Result

Analysis time dropped 60% and unhandled agent failures fell 55% versus the single-chain version.

• LangGraph
• GPT-4
• RAG
• FastAPI
• LangSmith

Measuring impact when you can't run an A/B test

Problem

Some programs roll out to everyone at once, so there's no randomized control group, and simple before-and-after comparisons kept overstating impact.

Approach

Built a toolkit of causal methods, including Difference-in-Differences, Synthetic Control, and Double ML, that builds a credible counterfactual for what would have happened without the program.

Result

Measurement error fell 30% compared with the previous attribution approach, and leadership got estimates with honest uncertainty ranges that informed which programs to expand and which to rethink.

• Synthetic Control
• Diff-in-Diff
• Double ML
• A/B testing
• PyMC

Document intelligence for clinical records

Problem

Clinical notes, insurance documents, and medical reports were reviewed by hand, which was slow and hard to scale in a regulated environment.

Approach

Built a platform that extracts entities, classifies documents, and answers questions over records using RAG, with an evaluation framework tracking hallucination rate, accuracy, and latency before anything shipped.

Result

Processing time fell 50%, manual review dropped 35%, and reliability of AI-generated insights improved 30%.

• LangChain
• Hugging Face
• FAISS
• Pinecone
• Azure AKS

Production fraud and risk models that stay healthy

Problem

Fraud and risk models score high-volume payment data, and patterns shift constantly, so a model that looks great offline can quietly degrade in production.

Approach

Built classification and scoring models with XGBoost, LightGBM, and PyTorch, with feature engineering on transaction, account, and third-party vendor signals. Added calibrated scores, drift and data quality checks, automated evaluation before every release, and dashboards in Prometheus and Grafana.

Result

Production regressions fell 45% and incident detection got 50% faster, and risk and operations teams adopted the monitoring.

• XGBoost
• LightGBM
• PyTorch
• Feature engineering
• MLOps

Making LLM inference faster and cheaper

Problem

As usage grew, response times and token costs grew with it.

Approach

Added Redis caching for repeat requests, compressed prompts, and moved long-running work onto async task queues, with Prometheus and Grafana dashboards to watch the effect.

Result

p95 latency fell 28% and token usage fell 32% without losing output quality.

• Redis
• Async queues
• Prompt compression
• Prometheus
• Grafana

Consulting across industries

At Hexaware I worked with enterprise clients across e-commerce and marketplaces, streaming and consumer apps, fraud and trust and safety, logistics, and financial services. The work included recommendation and ranking models, pricing and promotion experiments, fraud and anomaly detection, and demand forecasting, usually presented straight to client leadership.

Improved model performance 25% through feature engineering and tuning, cut data processing latency 35% with Spark pipelines, and reduced manual reporting effort 70% with automated dashboards.