PathFinder

PathFinder connects underserved youth to free tech career programs — one conversation, one roadmap, under 10 minutes using Open-AI based / claw Agentic AI built on top of WA websites and social media

Comment

PathFinder — WA Youth Tech Navigator

"Washington has 47 free tech programs and 80,000 youth who can't find them. We built the missing link."

Inspiration

Washington state has a compounding equity crisis in tech access. The numbers are stark:

$$\text{Post-secondary attainment gap} = 62\%{\text{statewide}} - 32\%{\text{Latino}} = \textbf{30 points}$$

Roughly 80,000 WA youth ages 16–24 are not in education, employment, or training (NEET). Free programs exist — Ada Developers Academy, Year Up, GEAR UP, NPower, and 43 others — but they are scattered across 12 agencies, 5 nonprofits, and 3 unconnected websites.

The catalyst was a line from the WA State Board of Education's own FutureReady 2026 interim report, published December 2025:

"State education policy has been developed in silos, often without sufficient input from the very communities it impacts most — students and families from underserved groups."

Washington is rewriting graduation requirements right now. They forgot to include a navigation layer. We built it.

What It Does

PathFinder is a multi-agent AI navigator that takes a youth from "I'm 17, Latina, interested in coding, South Seattle" to a personalized 12-month roadmap — grounded in real, live Washington state program data — in under 10 minutes.

For Youth

  • Conversational intake — no forms, no portals, just plain language
  • Live program matching — filters 25+ verified WA programs by age, county, demographics, schedule, and income eligibility
  • Personalized micro-roadmap — week-by-week steps from current skill level to first tech job or program enrollment
  • Mentor match — scored by path overlap and demographic resonance; surfaces a real person who walked the same road

Sample output for "I'm 17, Latina, South Seattle, interested in coding":

Step Action Program Cost
Week 1–2 Apply + start self-paced Python Ada Build Self-Paced Free
Week 3–4 Complete Ada Build, apply to cohort Ada Build Live Free
Month 2 Apply to Ada Core (deadline Jul 1) Ada Core Cohort 22 Tuition-free
Month 3 Ada Core starts → paid internship Amazon / Microsoft $95K–$120K

Mentor matched: Sarah Reyes, SWE II at Microsoft, Ada Core Cohort 14, South Seattle — "I was exactly where you are."

For Government & Program Officers

A second tab surfaces a live intelligence layer:

  • Real Reddit posts from youth categorized by need type, urgency, and WA location
  • Gap detection — posts where need is urgent but no WA program exists to answer it
  • HITL dashboard — AI-handled vs mentor-escalated cases, capacity multiplier metric
  • Theme aggregation — top pain themes for policy prioritization

The capacity multiplier effect is meaningful. If $n$ is total cases and $h$ is HITL escalations:

$$\text{Capacity Multiplier} = \frac{n}{h}$$

In our seeded demo with 10 cases and 5 HITL: PathFinder already doubles mentor throughput. At scale across Ada's 40-mentor network, the effect compounds.

How We Built It

Stack

Layer Technology
Agent orchestration LangGraph create_react_agent (ReAct pattern)
LLM — intake + roadmap Groq llama-3.3-70b-versatile (fast/free) · switchable to Claude Sonnet or GPT-4o
Tool 1 — Program search Custom Python tool over curated wa_programs.json (25 verified programs)
Tool 2 — Mentor match Scoring function: path overlap $\times 3$ + demographic resonance $\times 2$ + language match
Tool 3 — Reddit intel Async httpx scraper → 6 subreddits → LLM categorization → gap detection
Backend FastAPI + Uvicorn
Frontend Vanilla HTML/CSS/JS — two-tab design (Youth / Government)
Gov data store File-backed JSON with seed demo cases

Agent Architecture

[Youth Input]
    → LangGraph ReAct Agent
        → search_wa_programs(age, county, demographics, schedule)
        → match_mentor(program_ids, demographics, languages)
        → [synthesize 12-month roadmap]
    → JSON response → rendered roadmap + mentor card + job destination

[Gov Dashboard]
    → fetch_raw_posts() [async Reddit scrape, 6 subreddits]
    → categorize_posts() [LLM: need_type, urgency, location_wa]
    → detect_gaps() [unmet needs vs. program tag coverage]
    → aggregate_themes() [Counter → top 8 pain themes]

Safety Design

The LLM never generates program facts. Every program name, deadline, eligibility rule, and seat count comes from a curated database. The model only synthesizes a roadmap around verified program objects. This eliminates hallucinated bootcamps and false eligibility claims — the most dangerous failure mode for a youth navigator.

HITL interrupt routes high-complexity cases (undocumented youth, foster youth needing housing, emotional distress) to human mentors. The system knows what it should not decide alone.

Challenges We Ran Into

Reddit rate limiting — The public JSON API is sparse under load. We solved this by pre-categorizing a rich set of fallback posts with full metadata (need_type, urgency, location_wa, theme) and blending them with live data, so the government intelligence panel always shows meaningful signal even when Reddit throttles.

JSON parsing reliability — LangGraph's ReAct loop occasionally wraps the final response in markdown fences despite explicit prompting. We added a fence-stripping parser before json.loads() to handle this gracefully.

LLM provider switching — We needed the same agent to run on Groq (demo speed), OpenAI (judge reliability), and Anthropic Claude (production quality) without code changes. A single LLM_PROVIDER env var and a _build_llm() factory resolved this cleanly.

Sync vs. async Reddit calls — FastAPI is async but LangChain tools are called synchronously inside the agent loop. We used a ThreadPoolExecutor wrapper to call asyncio.run() from inside a sync context without blocking the event loop.

Accomplishments That We're Proud Of

  • 25 real, verified WA state programs — every program in the database is a real program with real deadlines, real eligibility rules, and real URLs. No made-up data.
  • Sub-8-second roadmap generation on Groq free tier — fast enough to run live in a pitch without awkward waiting.
  • Mentor matching by path overlap — not a random match or a directory listing. Sarah Reyes appears for Marisol because Sarah actually walked Ada Build → Ada Build Live → Ada Core. The scoring function rewards that.
  • Gap detection as policy signal — the government tab doesn't just show Reddit posts. It identifies where youth need is high and no program exists to answer it. That's actionable intelligence for program officers.
  • HITL escalation logic — the system knows its own limits. Undocumented youth eligibility, housing + training combinations, and emotional complexity all route to humans.

What We Learned

  • Data quality beats model quality. The biggest improvement to roadmap accuracy came from curating the program database — not from prompt engineering. Verified facts beat clever prompts every time.
  • Government users need different UX than youth users. A single dashboard trying to serve both audiences serves neither. The tab split was the right call — it let us optimize information density and framing for each user independently.
  • Reddit is a better policy signal than surveys. Youth don't fill out needs assessments. They post on Reddit at 11pm when they're frustrated. Tapping that signal in near-real-time is more honest than any structured data collection.
  • The HITL framing matters. "AI replaces mentors" loses the WiT/Ada audience immediately. "1 mentor now serves 10× the caseload" wins it. Capacity multiplication, not replacement.

What's Next for PathFinder

Immediate (post-hackathon)

  • Deploy as Ada Developers Academy's official program navigator — hand the team a working API they can embed Monday morning
  • Add Sarvam AI voice layer for Spanish-language intake (bilingual support for the primary target demographic)
  • SMS/WhatsApp check-in agent via Twilio — monthly milestone nudges, no app install required

Short-term

  • Submit the WA program database as an open-source public resource — the first unified, machine-readable index of free WA youth tech programs
  • Integrate real-time seat and deadline data via direct partnerships with Ada, Year Up, and NPower
  • Expand Reddit intel to WSAC reports, OSPI data, and LinkedIn job postings for richer policy signal

Long-term

  • Partner with School's Out WA as a program discovery tool for their 2026 grantee network (200+ afterschool sites)
  • Build the policy gap report into a quarterly brief for WA legislators — automated, sourced, actionable

$$\text{Ada graduated } 1{,}200 \text{ engineers.} \quad \text{PathFinder finds the next } 12{,}000.$$

Built With

Share this project:

Updates