🔁HelpLoop

HelpLoop

Find help. Match help. Move help.

People struggling to get basic things shouldn't have to search ten outdated websites and then coordinate help themselves. HelpLoop researches what's actually open near you, uses AI to pick the one that fits your situation, and connects a volunteer in realtime on a live map.

How it works

One request, five stages, four sponsor tracks.

  1. 1

    Someone asks

    Four answers: what they need (food, clothes, shelter), where they are, who they are (student, parent, senior…), and how they'll get there. No account, no ID.

  2. 2

    Linkup researches the open web

    Linkup · Deep Research

    One structured search across food-bank directories, 211 and city pages. Every result is stored as a finding. Then the pipeline looks at what each finding does NOT say — hours, walk-in policy, eligibility — and searches again for exactly that. Contradictions between sources are flagged and a verification query settles them. Everything is cited.

  3. 3

    Nebius ranks for this person

    Nebius · Applied AI

    The verified list plus the person's situation go to one model on Token Factory with one job: score each option and say why, in plain words. A pantry that closes before dinner, requires a student ID they don't have, or can't be walked to loses — and the reason is shown.

  4. 4

    Convex connects a volunteer, live

    Convex · Multiplayer

    The request becomes a shared row both screens subscribe to. A volunteer taps 'I can help'; the requester's screen changes instantly. Every status — picked up, on the way, delivered — flows the same way. The volunteer's position is tracked but only ever shared rounded to a quarter mile.

  5. 5

    The map makes it visible

    NERDCONF · Fun Build

    A 3D city with real buildings and terrain. The person waves until help comes; riders travel the route; MISSION ACCEPTED and HELP DELIVERED flash on both screens. Help becomes something you can watch arrive.

Who it's for

People in a hard week

Students between paychecks, a parent whose fridge is empty on a Thursday, someone newly unhoused. They don't need a directory; they need one answer they can act on tonight, and a hand getting there.

Volunteers with an hour

People who would help if it were as easy as accepting a ride request. No training, no shifts, no background paperwork for the MVP — just a live map and one button.

Cities, campuses and 211s

They already publish the data. What they lack is the last mile: turning scattered listings into a verified answer and a person who shows up. HelpLoop is that layer on top of what they have.

The business

The people using it never pay. The institutions already spending money on this problem do.

Campus and city licensing

A university food-security office or a city human-services department pays a flat annual fee for a branded instance: their resources prioritised, their volunteers, their dashboard of unmet need by neighbourhood. Sales cycle is slow but retention is high.

211 and food-bank partnerships

211 networks and regional food banks run helplines staffed by humans reading the same directories. HelpLoop's research pipeline is the tool their navigators would use — sold per seat, with the verification trail as the audit record.

Outcome data nobody has

Today no one knows whether the person who called 211 actually got fed. Every HelpLoop request closes with a delivered/not-delivered outcome and a timestamp. That dataset — where help succeeds and where it stalls — is what funders and cities pay for.

Honest state of things

Working today, on real data

  • • Live web research with follow-ups, conflict checks and citations
  • • Model ranking measured at 19/20, ~1.6 s, $0.0003 per request
  • • Realtime coordination across devices, race-safe accept
  • • Privacy-rounded live tracking; simulated volunteers for demos
  • • 3D map with terrain, avatars, light and dark themes

What we'd build next

  • • Real GPS on the volunteer side. The privacy rounding is built; the phone app that feeds it is next.
  • • Organisations posting directly. Let a pantry claim its listing and correct hours in one tap — the research trail becomes the fallback, not the source.
  • • Bigger evaluation set. 20 scenarios proves the pipeline; 200 with real outcomes proves the model.
  • • Volunteer trust. Ratings, ID verification, and a safety check-in for both sides before this leaves a hackathon.
  • • Clothes and shelter, hardened. The plumbing works today (shelter finds real beds); the gap checks and labels need the same care food got.
  • • Languages. Spanish, Chinese and Vietnamese first — the people who need this most often aren't searching in English.

Built at Burning Token with Linkup, Nebius Token Factory, Convex and MapLibre.