Projects

Independent Projects
In progress

Reinforcement Learning from Verifiable Rewards

RLVR Harbor Reward Shaping LLM Agents Python
  • Document extraction environment: an agent reads PDF invoices and receipt emails and returns key fields as structured output, with PII masked before any text leaves the machine.
  • Every rollout is graded by field, numeric, and line-item F1 verifiers, composed into one scalar reward.
  • Credit underwriting environment: an agent decides on synthetic loan applications, end to end — applicant in, verified reward out.
  • Deterministic verifiers check DTI arithmetic, policy application, decision label, and cited evidence.
Live demo

Autonomous Data Science Agent

LangGraph LLM Agents Python scikit-learn Streamlit
  • Closed loop: EDA → cleaning → feature engineering → modeling → evaluation → critic.
  • An LLM-as-a-judge critic reviews each run and its structured recommendations drive the next iteration — no hardcoded rules.
  • Streamlit app: upload a CSV, pick a target, and watch every step run live.
  • Exports the best iteration as a standalone Jupyter notebook.
Live

AnayaVerse

React TypeScript Vite Web Audio API Web Speech API
  • Little Learners: interactive ebooks for young children, with tap-to-speak narration.
  • Little Maestro: a music playground where every instrument is synthesized in the browser — no audio files, works offline.
  • Designed, built, and deployed end to end.
Professional Projects

AI-Powered HTML Receipt Parser

Product Strategy Cross-Functional Leadership Hugging Face AWS
  • Replaced a high-cost third-party vendor with an in-house AI platform.
  • Novel ML approach reads fields straight from receipt HTML, bypassing traditional OCR pipelines entirely — granted a US patent.
  • Delivered multimillion-dollar annual savings.
  • Supports 10M+ users and 600+ brand partners.

RAG Pipeline for Product Match Intelligence

Revenue Strategy Vector Databases FAISS ANN Search
  • Reframed product matching from an accuracy problem to a revenue-aligned one.
  • Failure analysis surfaced a Pareto distribution in unmatched items.
  • Two-track strategy lifted coverage 45% across millions of daily line items.

User Engagement & Retention Analytics

SQL Python pandas A/B Testing Tableau Snowflake
  • Composite engagement scoring across millions of active users.
  • Cohort-based retention analysis and churn prediction models.
  • Experimentation layer for A/B testing product changes.