Every recommendation comes with a reason.
Here’s how we build it.
MobilityLens doesn’t guess. Every plan, alert, and recommendation is assembled from your profile, your exact ZIP code, and authoritative research — then checked against six validation criteria before it reaches you. Not a chatbot. A decision-support system.
The recommendation pipeline
Seven stages. Every request. No shortcuts.
Request normalized
user_id, tier, locale, zip, task_type, trace_id captured and validated before any processing begins.
Stage classification
Anonymous → Free → Paid → Premium determines which features unlock, which model tier routes, and what data can surface.
Task classification
Structured query? RAG retrieval? Agent reasoning? A cheap classifier routes the request before any expensive model call.
Memory assembly
Session (Redis) + Profile (PostgreSQL) + Knowledge (Qdrant, location-filtered) assembled before the LLM sees any prompt.
LLM reasoning
vLLM/Mistral for small and mid-complexity tasks. Claude API for complex agent routes. Local inference is always attempted first.
Safety validation
Benefits cliff check, citation verification, PII redaction, tier permissions — six checks run on every response.
Response assembly
A typed AIResponse object is returned — never raw LLM text to the UI. Every output is structured and auditable.
Steps 3–5 route through six expert personas before the output is assembled.
Meet the six personasWhat the system knows before it reasons
Three distinct data layers assembled before any model call. The depth of the input is why the output is specific — not generic.
Your Profile
EMS scores, four-capital assessments, household composition, ZIP code, goals, and tier. Updated as you complete your assessment and take actions.
Location Intelligence
HUD Fair Market Rents, ALICE county thresholds, local job density, available programs, transit access. Filtered to your exact ZIP — not national averages.
Research Knowledge
Opportunity Insights mobility rates, Social Capital Atlas, policy documents, expert interview library. Retrieved with semantic search, not keyword matching.
Not a paragraph of advice. A structured decision.
Given real inputs, here’s what the system produces — alerts, goals, and recommendations with sources attached.
Input
Output
ALERT: Benefits cliff detected
At $39,520/yr, SNAP eligibility ends. Annual SNAP loss ≈ $4,200. Net household gain: ~$1,120 — not the expected $10,400. Do not accept without running the subsidy check below.
GOAL: Apply for childcare subsidy before job start date
Tennessee Child Care Certificate Program (85% SMI). Saves $340/mo. Changes net monthly gain from +$93 to +$433.
RECOMMENDATION: Negotiate schedule before accepting
School pickup window conflicts with shift end time. A 3:30 pm release on school days eliminates $214/mo aftercare cost.
Sources
ALICE Shelby County 2023 · HUD FMR 2024 · TN DHS SNAP table · Tennessee Child Care Certificate Program
Six validation checks on every recommendation
A recommendation that fails any check is blocked and replaced with a clear explanation.
Benefits cliff check
Would this income change leave the household worse off after accounting for lost program benefits?
Structural check
Does this resource or program actually exist and operate in the user's county?
Household check
Is this recommendation valid for the user's specific family composition?
Data freshness
Is the underlying ALICE data no more than two years old for this county?
Capital completeness
Have all four capitals been assessed before generating a mobility plan?
Citation present
Can every factual claim in this recommendation be traced to a specific source document?
What we never do
Privacy and data integrity are architectural constraints — not policy promises.
NEVER send client case data to third-party AI APIs
All sensitive data stays on our self-hosted infrastructure, never leaves the VPC.
NEVER store PII in the vector database
Qdrant holds only research knowledge and public documents — never personal records.
NEVER display a metric for cohorts under 50 people
Population analytics enforce a minimum cohort of 50 before any metric is shown.
NEVER generate or modify EMS scores using an LLM
The Economic Mobility Score is computed deterministically — no AI in the calculation.
The AI advisor uses a local Mistral 7B model hosted on our own infrastructure — your data never leaves our servers for AI inference. The EMS score is computed deterministically using the lib/scoring/ engine — no AI involved in the calculation, ever.
Ready to see the engine in action?
Schedule a walkthrough with our team or create a free account to explore the platform.