An end-to-end AI-powered platform that aggregates federal funding opportunities from multiple sources, processes them through an LLM-based evaluation layer, and presents ranked, contextualized recommendations to decision-makers.

Organizations pursuing federal grants face a fragmented, high-volume landscape of opportunities. Without systematic intelligence, relevant opportunities are missed and resources are misallocated to poor-fit applications.

Built a multi-stage pipeline: (1) automated ingestion from Grants.gov and agency APIs, (2) document parsing and semantic chunking, (3) vector embedding and retrieval layer, (4) LLM scoring against organizational profile criteria, (5) ranked output interface with reasoning traces.

RAG architecture with FAISS vector store, OpenAI embeddings, structured LLM prompting for multi-criteria evaluation, Python pipeline with Airflow orchestration, React frontend for opportunity review.

Active and in production use. Undergoing expansion to incorporate eligibility pre-screening and collaborative scoring workflows.

Prompt structure for multi-criteria scoring requires iterative calibration. Retrieval quality is the primary bottleneck — chunking strategy matters more than model choice for domain-specific documents.

A modular document processing system that ingests PDFs (including scanned documents), applies OCR, and uses structured LLM prompting to extract key fields, dates, requirements, and narrative summaries into a relational database.

Federal documents, FOANs, and policy briefs are dense, inconsistently formatted, and typically locked in PDF form. Analysts spend hours manually extracting basic information before any real analysis can begin.

Pipeline stages: document ingestion → format detection → OCR (Tesseract/AWS Textract) → text cleaning → LLM extraction with schema-defined output → structured storage → API access layer for downstream systems.

AWS Textract for scanned documents, custom post-processing for layout reconstruction, JSON schema-constrained LLM extraction, PostgreSQL storage with full-text search, FastAPI layer for integration.

Production deployment processing 50–80 documents per week. Integrated as upstream feed for the Opportunity Intelligence system.

Schema-constrained outputs are essential for reliability. OCR quality varies dramatically with scan quality — pre-processing steps (deskew, denoise) have outsized impact on downstream extraction accuracy.

A structured database system backed by NLP pipelines that ingests qualitative research data (interview notes, survey responses, stakeholder comments) and classifies, tags, and stores them for systematic analysis and retrieval.

Qualitative research data is inherently unstructured and resists systematic analysis. Research teams were manually coding thousands of inputs — a slow, inconsistent, and non-scalable process.

Designed a codebook-informed classification schema, trained a multi-label text classifier on annotated training data, built ingestion pipeline, and created a relational schema linking themes, sources, and metadata for flexible querying.

Fine-tuned BERT-based classifier for thematic coding, spaCy for entity extraction, PostgreSQL with JSONB for flexible schema, Metabase for analyst-facing dashboards and query interfaces.

Used in production for two major research cycles. Framework generalized and applied to NTIA listening session analysis.

Codebook design is the most important upfront investment. Ambiguous categories create downstream classification noise that's hard to correct at scale.

A data pipeline and interactive dashboard suite providing program managers with real-time operational insights, performance tracking, and resource utilization analytics across a multi-site program.

Program leadership was operating with lagging, siloed data — making resource allocation decisions based on weekly reports rather than current operational reality. Bottlenecks went undetected until they became crises.

Built ETL pipeline aggregating data from 5 operational systems, designed a dimensional data model, created role-differentiated dashboard views (executive summary, operational detail, regional comparison), and built alerting for threshold breaches.

Python ETL with dbt for transformation, PostgreSQL data warehouse, Tableau for dashboards, custom alert logic for anomaly detection, role-based access controls.

Deployed and in active use by 30+ program staff. Expanded to include predictive utilization forecasting.

Dashboard adoption depends more on UX and stakeholder involvement in design than on analytical sophistication. Build with end-users, not for them.

A machine learning model trained on historical challenge program data to forecast submission likelihood, volume estimates, and predicted quality tier — enabling program teams to target outreach and scale evaluation resources appropriately.

Challenge programs consistently faced unpredictable submission volumes — either overwhelming review capacity or missing participation targets. Planning was based on intuition rather than data.

Engineered features from historical program data, registration signals, outreach metrics, and program design characteristics. Built ensemble model with calibrated probability outputs and confidence intervals for planning use.

Gradient boosted trees (XGBoost), SHAP for interpretability, scikit-learn pipelines, calibration with Platt scaling, prediction intervals via conformal prediction.

Used for two challenge program cycles. SHAP-based explanations adopted by program team as a communication tool for leadership reporting.

Interpretability is not optional for operational models. Stakeholders need to understand why a forecast says what it says — black-box predictions erode trust and limit adoption.

A comprehensive economic impact study measuring the short- and long-run returns to federal broadband and community media center investments, using difference-in-differences and instrumental variable methods to establish causal estimates.

Federal broadband and CMC programs lacked rigorous, causal evidence of economic return. Policy debates were driven by advocacy rather than evidence, undermining funding allocation decisions.

Assembled longitudinal dataset from FCC, Census, BLS, and program-specific records. Applied difference-in-differences with event study design to estimate treatment effects on employment, income, and business formation. Used distance-to-infrastructure as instrumental variable.

Broadband expansion associated with 4–7% employment growth in treated census tracts over 5-year windows. CMC funding showed concentrated impacts on youth employment and small business formation within 3-mile radii of funded centers.

Published as program evaluation report. Methods framework adapted for ongoing BEAD program monitoring design.

Causal identification requires patient data assembly. The quality of the counterfactual is what determines whether policy audiences trust the analysis.

Applied the qualitative research database framework to analyze thousands of public inputs submitted to NTIA listening sessions, identifying dominant themes, geographic patterns, and stakeholder-type differences in broadband policy priorities.

NTIA received thousands of written and verbal inputs across listening sessions. Manual review was too slow and too subjective to produce systematic intelligence for policy drafters.

Adapted thematic classification pipeline to NTIA-specific codebook. Processed ~4,500 inputs, identified top themes by stakeholder type and geography, and built summary visualizations for policy staff consumption.

Multi-label NLP classification, topic modeling (BERTopic), geographic aggregation to CBSA level, stakeholder type clustering, Tableau dashboard for policy staff.

Analysis delivered and used in policy drafting process. Methodology documented for replication in future comment periods.

Speed of delivery matters for policy relevance. An analysis that arrives after the decision window closes has zero impact, regardless of quality.

A policy analysis examining ambiguities in BEAD program design and how variation in state-level implementation choices could produce divergent coverage and equity outcomes — even with identical federal funding levels.

BEAD's flexible state-driven design creates wide implementation latitude. Stakeholders lacked a clear framework for understanding how different state choices interact with federal requirements — and what's at stake for underserved communities.

Mapped key program design decision points, modeled coverage scenario outcomes under different implementation choices using FCC broadband data and Census geography, and produced comparison analysis across state implementation plans.

Scenario modeling with FCC Fabric data, spatial analysis in Python (geopandas), comparative implementation plan review, structured policy framework development.

Published analysis. Ongoing work tracking state implementation decisions against predicted coverage outcomes.

Policy analysis has the most impact when it gives decision-makers a clear framework — not just conclusions. Structure and accessibility matter as much as analytical rigor.