01 · Thesis
The people you need are one or two hops away. Nobody can see them.
In 1973 Mark Granovetter showed that most people find jobs through acquaintances, not close friends. Weak ties carry most of the opportunity, and they're the first ties we forget. Human memory holds roughly 150 meaningful relationships (Dunbar); a community of 3,000 people is invisible to everyone inside it.
Communities have always tried to solve this on their own: neighborhood associations, alumni networks, the one friend who knows everyone. Who has a spare room, who is hiring, who knows a good lawyer? Today that knowledge is scattered across WhatsApp groups nobody can search and nobody remembers.
Topels is the first mutual-aid network with a mind.
| Product | Core unit | How it works |
|---|---|---|
| Social feeds | Friendship | Shows content; wins the longer you stay. |
| Profile | Waits for you to search and apply. | |
| AI assistants | One-on-one chat | Helps one person; knows no one else in the network. |
| Topels | Needs × offers × time | Reasons over the whole network and connects the right people before anyone asks. |
02 · An intro, step by step
From one sentence to the right person, in seven steps.
A typical story: someone who just moved to town needs a car, and a member is selling one next month. Neither knows the other exists.
- Step 1/7
Elif
Just moved to Queens. I need a used car in the next three weeks, budget $12k max.
Yuva
Got it, I'll keep an eye on the network. I won't share anything with anyone without asking you first.03 · Architecture
Knowledge lives in the database. Intelligence lives in the models.
We don't train language models on private conversations. Facts about people sit in a permissioned, deletable store, and models read them only for the task at hand. What Topels learns over time is judgment: which introductions work.
Governance
- Consent ledger
- Audit log
- Deletion pipeline
Frontier language models
Conversation, extraction, nightly synthesis
Our own models
Matching · trust · foresight · immune
Runtime
- Job queues
- Nightly run
- Notifications
→ Every model call goes through one router. The language layer is swappable; the matching intelligence is ours.
04 · Agent society
Eight Topels, one society of mind.
Marvin Minsky described the mind as a society of small agents, each doing one simple job. Topels works the same way: every agent is a model call with its own tools and rules, and they all read and write one shared memory. Members see eight Topels; behind them, ten agents do the work.
YuvaOracle
Learns the typical order of someone's first months in a new city (housing, paperwork, bank, car) and gets ahead of the next need.
SelamMatchmaker · Envoy
Finds and weighs pairwise matches, then picks the right moment and the right words for each ask.
MantıGuardian
The consent gate and the immune system: strips sensitive data, enforces the clean room, and spots fraud patterns across the network.
BilgeScribe · Archivist
Extracts facts from conversation, decides what to remember and what to forget, and answers "ask the network" from collective memory.
SomunMatchmaker (work)
Referrals, jobs and collaborators. Fair by design; no score is ever shared with an employer.
DemliArchitect
Many-to-many matching: dinner tables, teams, chains and swap cycles across cities.
HayalDreamer · Explorer
Runs the nightly synthesis and takes random walks on the graph to find connections nobody searched for.
VefaPulse
Measures network health without personal data: who's isolated, who's overloaded, where help flows back.
05 · Memory
Four layers of memory, and the right to forget.
01
Personal
Only in your Topel
Who you are, what you need, what you can offer, your preferences. Encrypted per member.
02
Relationship
Between two people
Who introduced whom, who vouched for whom, which introductions worked.
03
Collective
Anonymous and aggregate
The network's know-how: which neighborhoods are great to live in, which accountant people trust. No one's personal data is visible.
04
Temporal
The future
Upcoming dates, standing intents and "someday" dreams, matched when the conditions come true.
Forgetting is the default. Stale needs fade, sensitive details are never written down, and members can erase everything at any time. That's also why private data never gets baked into model weights.
Temporal memory · example
In 2026 you said, "Someday I want to open my own bakery in Boston." This week Zehra joined: a pastry chef looking for a partner. Want to meet?
06 · Matching engine
Not a feed. A short list of the right people.
Simplified score
score(a, b) = fit × timing × trust + serendipity − fatigue
- fit
- need ↔ offer similarity
- timing
- urgency and time window
- trust
- behavior-based, never identity-based
- serendipity
- unexpected value
- fatigue
- attention budget
01
Candidate generation
Hybrid retrieval: embeddings pair needs with offers, graph walks find trust paths two or three hops out, and hard constraints (place, time window, language) prune the rest.
02
Scoring
Each pair is scored on fit, timing, trust, reciprocity and novelty, minus fatigue. A learned ranker does the math; an LLM judge writes the reason a person will actually read.
03
Attention budget
Only the most valuable opportunity reaches a person's attention. Everyone's time is protected.
04
Many-to-many
Chains and cycles (three people in three cities swapping rooms), teams and tables: where pairwise matchers never look.
05
Exploration quota
A share of intros always goes to new members, so the system never recreates the room where nobody knows anybody.
06
Explainability
Every suggestion comes with a reason that both people can see.
07 · Privacy by architecture
Topels knows everyone and exposes no one.
Personal vaults
Each member's memory lives in its own encrypted partition. Other agents see only what that member allowed.
Clean-room matching
Two personal agents compare permitted fields and log every comparison. Raw conversations never leave the vault.
Double opt-in
Names are revealed only after both people say yes.
Consent ledger
Every disclosure is recorded and visible to the member.
No training on private chats
Private conversations are never used to train language models, so deleting your data actually deletes it.
People aren't ranked
Topels never ranks people by who they are, only by fit for a specific need. Sensitive personal information is never used in matching.
Always an AI
Topels always says it's an AI and never impersonates a person.
Neutral
No political profiling and no voluntary data sharing with any government. Legal demands are challenged and published in a transparency report.
08 · Learning
A mind that gets sharper with every introduction.
- 1
Daily
Every intro's outcome becomes a training example; the matching model retrains every week.
- 2
Weekly
Every bad suggestion joins the eval set. No prompt or rule ships unless it passes every past set.
- 3
Monthly
Collective know-how and life-journey patterns are recomputed.
- 4
At scale
Repetitive work (extraction, classification) moves to small open-weight models: cheaper, faster, less dependent on any one provider.
| Phase | Language layer | Our own models |
|---|---|---|
| Phase 0–1 | Frontier models plus rules | None yet; every outcome is logged with consent |
| Phase 2 | Frontier models | Matching and trust (after the first ~1,000 outcomes) |
| Phase 3 | Frontier models, on-device personal agent, fine-tuned Scribe | Foresight and immune |
| Phase 4 | The best model for each job | Full learning loop |
The moat: the outcome graph
Profiles are everywhere. A record of which introductions actually changed something exists only in Topels, and it grows with every intro. A competitor starting today can't buy that history.
09 · Network design
Density before breadth.
A matching network is only as good as its local density. We grow one tight community at a time, by invitation, so every new member meets someone useful in their first week.
01
Founding 100
New York
02
New York communities
Events and community partners
03
US cities
Boston, Chicago, San Francisco
04
Europe
Berlin, London, Amsterdam
05
Every community
Same playbook
Every member gets two invitations. Invitations form a trust genealogy: we always know who vouched for whom.
North star
Weekly valuable introductions
Intros that both sides call useful. Not member count.
Health metrics for the mind
Reflex time
< 72 h
Time from a stated need to meeting the right person.
Precision
↑
Share of intros both sides found useful.
Isolation
≈ 0
Share of members with no new connection in 30 days.
Reciprocity
↑
Share of helped members who later help someone else.
Serendipity yield
↑
Share of valuable intros nobody asked for.
Trust
< 1%
Share of messages flagged as unwelcome.
10 · Stack
Boring where it should be, ambitious where it matters.
- Clients
- Expo (React Native) and Expo Router: iOS, Android and web from one codebase · Next.js
- Data
- Postgres (Supabase) · pgvector · auth · realtime chat · storage
- Brain runtime
- Job queues · scheduled nightly run · push notifications
- Models
- Provider-agnostic router · Claude Opus 5 for conversation and extraction · Claude Fable 5.1 for nightly synthesis · our own small models
- Characters
- Rive animations, one file for app and web
- Infrastructure
- Vercel · Supabase · AWS
Unit economics
At list prices, model spend is about $0.30 a month for a passive member and $4–5 for an active one, and it falls as repetitive work moves to our own models.
11 · Roadmap
The mind grows up in four stages.
Phase 0–1 · Sep 2026 – Jan 2027
Reflex
Pairwise intros with your yes. Personal memory, the person one step ahead, teams, tables, trusted services, no one left alone.
Phase 2 · Feb – May 2027
Memory
A mind that remembers time: standing intents, nightly dreams, seasonal chains, dormant ties, solidarity mode.
Phase 3 · Jun – Sep 2027
Reasoning
Groups and chains, foresight, serendipity engineering, ask-the-network, the immune system, on-device personal agents.
Phase 4 · Oct 2027 onward
Wisdom
A network aware of itself: community pulse, collective bargaining power, portable trust, bridges between networks.
12 · Constitution
The constitution of the mind.
- 01The network thinks; people decide.
- 02Topels knows everyone and exposes no one.
- 03Consent is the architecture.
- 04People are never ranked; only fit for a need is weighed.
- 05Every suggestion has a reason the person can see.
- 06Success is value created in people's lives. No infinite feed, no addiction by design.
- 07Topels knows how to forget.
- 08Topels always says it's an AI.
- 09If the rules change, the people of the network are asked first.
We're building a network with a mind, starting in New York.
Join the waitlist. If you're an investor, or you want to help build this mind, write to us.