Cliffline — Benefits Navigator

An AI-powered benefits navigator and cliff calculator that models state-level eligibility, income volatility, and demographic enrollment uptake to guide low-income workers to financial stability.

Comment

Inspiration

Families relying on public benefits such as SNAP, Medicaid, and childcare subsidies face a complex, bureaucratic maze. The most punishing aspect is the benefits cliff: the point where a small raise or extra hours triggers a sudden, disproportionate loss in public assistance, leaving a family financially worse off.

Existing tools are either static listings that do not compute eligibility, or opaque black boxes that hide cliff risks. This lack of transparency forces workers to turn down career advancement out of fear. We built Cliffline to resolve this information asymmetry, turning policy complexity into clear, risk-aware financial paths.

What it does

Cliffline is an AI-powered benefits navigator and cliff calculator that helps users assess their eligibility and project financial scenarios:

  1. NLP Intake & OCR: Users describe their situation in plain English or Spanish, or upload pay stubs and W-2s. Gemini 2.5 Flash extracts demographics, income, and state details.
  2. State-by-State Rules Engine: Implements actual criteria across all 50 states and D.C. (no synthetic placeholders), incorporating KFF Medicaid expansion limits, SNAP Broad-Based Categorical Eligibility (BBCE), and Child Care and Development Fund (CCDF) activity rules.
  3. Interactive Cliff Slider & Scenario Comparison: Models how benefit allotments change at hypothetical wages. Users can compare two job offers side-by-side to see wages, benefit deltas, and net changes.
  4. SIPP Machine Learning Integration: Predicts enrollment likelihood using an XGBoost model trained on Census Bureau SIPP data, displaying participation likelihood and explainable SHAP contribution scores (e.g., "Income Band: +15%").
  5. 90-Day Volatility Projections: Plots probability bands (10th, 50th, and 90th percentiles) around household income using standard diffusion models: $$\sigma = \text{volatility} \times \sqrt{t}$$
  6. Safety Guardrail: Deterministic regex filters scan text inputs. If crisis keywords (self-harm, domestic violence) are detected, the app bypasses the benefits pipeline entirely to place the 988 Crisis Lifeline front-and-center.

How we built it

  • Frontend: React and Next.js App Router for a responsive, high-fidelity user interface.
  • LLM Layer: Server-side JSON-mode API calls to Google Gemini 2.5 Flash for intake parsing, with Groq Llama 3.3 as an active fallback.
  • Machine Learning: We analyzed longitudinal data from the Census Bureau SIPP survey. We trained XGBoost models to predict cohort enrollment behaviors. To eliminate runtime server cost and latency, we precomputed these predictions into an O(1) client-side cohort lookup database (sipp_ml_data.json).
  • Policy Engine: A strict, rule-based mathematical calculator built from official federal and state manuals, aligned directly with the methodology of the Atlanta Fed's CLIFF tool.
  • Fairness Audit: We wrote a Responsible AI fairness script (python/cliffline_parity_check.py) that checks prediction parity across demographic subgroups to flag any disparate-impact ratio under the 0.8 threshold.

Challenges we faced

  • Hydration Mismatches: Combining dynamic third-party HTML/Framer layouts with server-rendered React components initially caused hydration mismatch warnings. We resolved this by dynamically loading our visual landing page with server-side rendering disabled ({ ssr: false }).
  • Policy Modeling Diversity: Every state has its own formula for benefit allotments. For example, SNAP net income eligibility depends on allotments computed via: $$\text{Benefit} = \text{Max Allotment} - 0.3 \times \text{Net Income}$$ Ensuring our calculator correctly mapped the various state-specific deductions and FPL tables required strict mathematical validation.

Accomplishments we're proud of

  • Halting Hallucinations: We separated AI from policy. Cliffline uses LLMs only to parse user situations and document fields, while eligibility calculations are entirely deterministic, eliminating AI policy hallucinations.
  • Client-Side Latency: Compiling SIPP models into a compact lookup database allows the app to load cohort SHAP explanations, volatility bands, and participation probabilities instantly.
  • Typo-Robust Crisis Detection: The safety guardrail is deterministic, fast, and robust against spelling mistakes (e.g., "suicuide"), guaranteeing critical help is never missed.

What we learned

We gained deep insights into the socioeconomic inequalities embedded in public policy, specifically how non-expansion states (like Texas and Florida) lock childless adults out of healthcare at extremely low income levels compared to expansion states. We also learned how to implement rigorous Responsible AI parity checks using the federal four-fifths rule.

What's next for Cliffline

We want to integrate deeper state-level application APIs to enable one-click document submittal directly from the app, automate application form pre-filling, and expand our deterministic calculator to cover more localized county-level assistance programs.

Built With

Share this project:

Updates