Mohona Yesmin.

Education

Bachelor of Arts, Major: Computer Science, Minor: Math and Statistics

Bachelor of Arts, Major: Neuroscience and Behavior, Minor: Chemistry and Film. Extensive coursework in data analytics.

Experience

• Validated and scaled a multimodal Graph Neural Network pipeline (PyTorch, PyTorch Geometric GAT/GCN encoders, ESM protein embeddings) for toxin ion-channel classification; debugged a label-mapping defect silently corrupting class-level statistics and extended a persistent-homology analysis (GUDHI) from a 3-sample proof-of-concept to a validated 200-protein benchmark, achieving 87.5% test accuracy.
• Implementing a custom Gradient Reversal Layer (PyTorch autograd.Function) for domain-adversarial training to close the ~30-point generalization gap between random-split (~86%) and held-out-taxon (~55%) accuracy.

• Developed and deployed statistical models to predict building emissions trajectories across 200+ building portfolios, delivering interpretable, actionable LL97 compliance recommendations that drove business value for 50+ clients.
• Engineered large-scale data pipelines to scrape, transform, and aggregate multi-source energy datasets, building client-facing Power BI dashboards and supporting LL97 compliance filings and payment processing for 900+ properties in coordination with DOB systems.
• Automated client communications, document workflows, and file management across 1,000+ clients using Python-based tools, reducing manual processing time by more than 50%.

• Designed and deployed a JSON parsing agent to normalize LLM outputs across multiple data formats, ensuring consistent inter-agent communication and scalable multi-agent orchestration using RAG and agentic AI pipelines.
• Implemented FastAPI endpoints to integrate AI agents with live dashboards, enabling real-time data processing, decision support, and system transparency in production environments.
• Extracted, transformed, and validated backend SQL datasets by engineering new data columns and persistence logic to ensure accurate, reliable data storage across agent operations.

Previously: Fellow, Sep – Dec 2023 · Chapter Committee Leader (part-time), Dec 2023 – Apr 2024

• Designed and analyzed 100+ community surveys and maintained a 500+ member database, turning social-media, news and policy data into visualizations and reports for non-technical decision-makers.
• Researched climate change and climate disasters across South Asia, and turned the analysis into data visualizations and teaching materials for community climate education, work I later extended into my climate research project.
• Led multiple campaigns and designed leadership-development programs for youth and adult members; hosted workshops and trained members to lead their own sessions.
• Managed members and volunteers through conflict and crisis, keeping campaigns on track under pressure.

• Completed 500+ hours of hands-on training in Python and SQL, exploratory data analysis, classical statistical modeling (study design, model evaluation, linear and logistic regression) and machine learning, from decision trees and random forests to NLP and neural networks.
• Applied it end to end in three projects, each a written technical report plus a presentation built around a stakeholder question: a climate and life-expectancy analysis of the US and Bangladesh against 190+ economies; an automated valuation model for residential property with conformal price intervals and out-of-time validation; and a Reddit NLP classifier (TF-IDF, logistic regression, Naive Bayes) that I later rebuilt into ModSieve, a moderation-triage tool.

Projects

• Shipped merged PRs migrating Python clients off deprecated APIs to v1 per KEP-0004 across kubeflow/hub (merged: #3114, #3137) and kubeflow/sdk, regenerating OpenAPI clients and validating 270+ unit/E2E tests with pytest, mypy, ruff.
• Designed and shipped a decoupled upload_artifact utility (sdk #514, merged) (Python, Pydantic) supporting S3 and OCI object storage, isolating artifact uploads from registry logic per a community feature request.
• Building a Go client and backend proxy layer (hub #3136, open) routing UI traffic to v1 Model Registry, Model Catalog, and MCP Catalog APIs; co-authored the SDK 0.4.0 release blog post and became a Kubeflow org member on the strength of sustained contributions.

• Designing an app that reads uploaded PDFs aloud in an original synthetic voice (StyleTTS 2) with word-by-word synchronized highlighting, a retrieval-augmented “Ask Sera” assistant (LangGraph, pgvector) and reading-based recommendations.
• Built for its first 100 users and architected to scale to 10,000+: FastAPI (async), Postgres with pgvector, Redis caching, an SQS-backed TTS worker queue, and AWS ECS with a Kubernetes path. Scaling targets are defined per user tier; load-test results will be published as they are measured.
• Project page: github.com/myesmin/sera, with the engineering method behind it. Live demo will be linked here at launch.

• A six-notebook pipeline on 2024 data (Python, pandas, statsmodels, scikit-learn, SciPy): emissions drivers, backtested forecasts, carbon-budget scenarios to 2100, carbon liability and income at risk from warming, with an interactive Plotly dashboard, 31 figures and 35 tests.
• Found a 23× per-person emissions gap between the US and Bangladesh, forecast error rising to about 18% at nine years, and fast decarbonisation still overshooting a 2 °C budget.

• An automated valuation model for residential property — the class of model a lender uses to price collateral (Python, scikit-learn, LightGBM, DuckDB over parquet). Every estimate carries an 80% conformal price range whose coverage is measured against realised sales, validated out-of-time at 5.5% median error with 77% of homes within ±10%.
• Ongoing monitoring across six production quarters, a governance review against the 2024 Interagency AVM Rule, and a credit-risk model in which pricing valuation uncertainty adds 85% to expected losses. Packaged as a containerised HTTP service, with the results published as a static page.

• A first-pass moderation tool that resolves the posts it is confident about and sends the rest to a human moderator. Tuned to hold a fixed precision, it auto-resolves 22.5% of posts at 99.2% precision and scored 0.820 against a blind human baseline of 0.802. Labels come from real moderator removals collected hourly across 8 communities.

Certifications

Hugging Face AI Agents · McKinsey Forward Program

Leadership

• Young Climate Leaders of Color — Fellow, New York, NY (2025 – 2026)
• TEDxWesleyanU — Marketing Intern, Middletown, CT (2020 – 2022)
• National Organization for Rare Disorders (NORD) — Vice President, Middletown, CT (2020 – 2023)