A matching layer that sits on top of a partner fund's existing deal flow. Listens to VC-founder calls, scores sentiment + fit, surfaces coaching notes for the founder and ratings for the VC, and turns high-confidence pairings into portable signal. Not another CRM — the engine that makes the next-best founder obvious.
The best founder–VC matches are obvious in the first 20 minutes of a call. MatchBox makes that signal legible, portable, and actionable — and, critically, durable and confidential. Memory is the cornerstone of the matching problem; the non-obvious connections between funds and founders only surface once a deep corpus exists. Transcripts are PII — so every analyzer call runs against a per-fund fine-tuned local Gemma on claws-mac-mini. Nothing sensitive crosses the Pi-harness boundary. See Runtime Choice.
Transcribes every VC-founder call and scores sentiment, fit, and red flags. Captures training data for v2 matching.
Pairs founders with the right fund based on call signal plus intake data, not just thesis keywords. (v2 deliverable.)
Returns founder-ready coaching notes and VC-ready call ratings after each conversation. Closes the feedback loop.
High-confidence deals get shared across a trusted VC network as the subscriber base grows. Network depth compounds.
| Key | Value |
|---|---|
| Repo | Organized-AI/MatchBox (private · MIT) |
| Default branch | claude/founder-fund-matching-urzts |
| Base project | Forked from ClawBox (still reports clawbox in package.json) |
| Shipping surface | Signal · src/components/SignalView.tsx |
| Founder intake | Airtable form |
| Strategy call | Granola notes |
Upgrade a working 7/10 VC workflow into the 10/10 industry benchmark by wiring MatchBox into a partner fund's existing process. No forced platform migration.
Jake meets Jon Forbes this Friday to lock compensation terms before broader rollout. Options on the table:
Three product versions, compounding value — signal captures the training data for matching, matching's ranking signals feed founder prep.
Turns each VC-founder call into structured signal. Transcribe → sentiment + fit + red flags → coaching notes for the founder + ratings for the VC. Captures training data for v2.
High-confidence founder ↔ fund pairings. Cross-network deal sharing between subscribed funds. Match quality improves fund-over-fund as the data pool grows.
AI avatar conducts the initial founder interview. Automated screening before the VC meeting. Due-diligence layer that cuts fraudulent claims and wasted partner time.
src/components/SignalView.tsx — the operator UI.src/types/signal.ts — FundProfile · CallSignal · SentimentScores · RedFlagHit · AirtableConfigView · AnalyzeCallInput.src/services/signal.ts — fetch client against the local backend at http://127.0.0.1:13000.src/store/signal.ts — Zustand state: funds, calls, Airtable config, loading/syncing/analyzing flags.internal/routes/signal.ts — Hono routes (Airtable config, sync, funds CRUD, analyzer).internal/signal/{index,airtable,analyzer}.ts — store + Airtable sync + LLM analyzer.Signal turns a call into a structured CallSignal: rating, sentiment scores, red flags with quoted evidence, fact-finding prompts, and coaching notes for the VC.
internal/routes/signal.ts| Method + path | Role |
|---|---|
| GET /api/signal/airtable | Current Airtable config (API key masked to last 4) |
| PUT /api/signal/airtable | Patch config · supports clearApiKey |
| POST /api/signal/airtable/sync | Pull FundProfile rows from Airtable |
| GET · POST · PUT · DELETE /api/signal/funds | Manual fund CRUD (coexists with Airtable-sourced rows) |
| GET · POST · DELETE /api/signal/calls | Call list + analyze + delete |
internal/signal/analyzer.tsBuilds the prompt from the FundProfile + transcript (truncated to MAX_EXCERPT_CHARS = 20_000), calls callOpenclawDefaultModelChatCompletion, validates the JSON response, and produces an AnalyzeResult that gets wrapped into a persisted CallSignal.
quote + reason.…[truncated N chars] so the model knows.MatchBox inherits the ClawBox three-process split: Tauri shell · React 18 SPA · Bun/Hono backend. Signal adds one route module, one store, and one analyzer over that base.
agents · channels · chat · config · cron · models · onboard · sessions · skills · soul · titles · tools).mock-gateway.mjs, smoke-backend, sync-version, Windows signing chain, macOS .dmg build.SignalView.tsx, signal.ts type/service/store trio, routes/signal.ts, internal/signal/*.MatchBox ships as an agent. Each VC gets a custom GPT trained on their own process, reachable where they already work — not as a forced migration to a new app.
Coaching notes + call ratings posted directly into the fund's working channel.
Same signal, delivered where the partner already texts with founders.
Follow-up + next-call prompts surfaced alongside the VC's existing thread.
Founder intake — the signup link. Rows sync into MatchBox's fund store.
Engine-identical to ClawBox. Same npm ci → npm run dev loop.
Default branch today: claude/founder-fund-matching-urzts. No main at repo root yet — signal + matching work land on topic branches.
Memory is the cornerstone of MatchBox. Matching funds to founders requires connections that only surface after the corpus is deep enough — unforeseen pattern matches across sectors, stages, and partner appetites. That means Hermes can't wait until v3. It lands alongside OpenClaw in v1, as the memory substrate.
MatchBox calls callOpenclawDefaultModelChatCompletion in internal/providers/openclaw-completions.ts. That's the sole tight link to OpenClaw. Everything else — the Tauri shell, React frontend, Bun backend, SignalView, the Airtable sync, the analyzer's validation layer — is provider-agnostic. A swap is a provider-module rewrite; an addition (what we're doing) is even cheaper.
Keeps doing what it's best at — one-shot structured-JSON analysis of a transcript. Fast, low-latency, mock-testable, Forbes-demo safe. No changes.
Writes every CallSignal into Honcho/Mem0 keyed by fund + founder + sector + stage. Queries return cross-call pattern matches. Same runtime also handles Slack/WhatsApp/email delivery.
claws-mac-mini. Reuse the Pi harness — inherit the launchd supervision, Codex OAuth primary, Gemma-4 self-heal fallback, repo-digest tool surface.SettingsView, PluginsView, GatewayRestartBanner, internal/compatibility.json all assume OpenClaw. Touches most of them.scripts/mock-gateway.mjs + smoke:backend only exist because OpenClaw has a defined minimum RPC surface. Hermes has no equivalent mock — smoke tests get more expensive.npm i -g openclaw@latest. Hermes: Python venv + service supervisor. If partner funds ever self-host, OpenClaw is a gentler ask.models.list, sessions.*, chat.*, config.* — testable against a mock, easy to reason about.Transcripts contain named founders, named funds, named partners, financial figures, and confidential thesis signal. None of it leaves the Pi harness. The analyzer runs against a per-fund fine-tuned local Gemma served by MLX (Apple's Metal-accelerated ML framework); the base Gemma-4 4-bit quant serves as the fallback for funds without a fine-tune yet, via mlx_lm.server on :8080.
claws-mac-mini is Apple silicon (M-series). MLX is native Metal — better memory efficiency than llama.cpp for fine-tuning, first-class LoRA tooling via mlx_lm.lora, single-framework stack for both inference and training. The mlx_lm.server exposes an OpenAI-compatible /v1/chat/completions endpoint so client code (local-gemma.ts) doesn't care that the runtime changed under it.
CallSignal history (analyzer inputs + outputs) is the training set. Stored locally under ~/.hermes/corpora/fund-<id>/ in JSONL. Never exported.mlx-community/gemma-4-it-4bit — Gemma 4 instruction-tuned, 4-bit MLX quantisation. Converted once via mlx_lm.convert and cached at ~/.hermes/models/gemma-base/.mlx_lm.lora --model gemma-base --train --data <fund-jsonl> --adapter-path <fund-dir>. Runs on-host; the M-series GPU handles it in minutes-to-hours depending on corpus size.matchbox finetune <fundId>.~/.hermes/models/fund-<id>/ containing the LoRA adapter + config. mlx_lm.server loads by path; optionally mlx_lm.fuse bakes the adapter into a standalone fund model for faster cold loads.fund-<id>.v<N>/ for rollback.analyzeTranscript may only call providers whose base URL is 127.0.0.1 or a Tailnet-scoped host. Lint-gated.internal/providers/openclaw-completions.ts is reserved for non-PII tasks (title suggestions, settings probes). An explicit allowExternal: false flag gates every call path.internal/delivery/sanitiser.ts) that strips transcriptExcerpt and the redFlags[].quote literal text. Only reason and structured scores travel.internal/providers/hermes-client.ts — thin client to the Hermes gateway on claws-mac-mini. Mirrors the shape of openclaw-completions.ts but targets ai.hermes.gateway's HTTP surface. Bearer-auth via a token stored alongside the existing gateway secrets.internal/memory/signal-memory.ts — the memory write layer. Every completed CallSignal emits a set of structured facts to Hermes Honcho, keyed by fund id, founder name, sector, stage, and call date.internal/delivery/hermes.ts — routes a CallSignal into Slack / WhatsApp / email via Hermes platform adapters. Reuses the same client.POST /api/signal/calls. After the analyzer returns a CallSignal, fire-and-forget a memory write with facts like:
:<id>/thesisAlignment/sector:<s>/stage:<t> = N:<name>/redFlagPattern/<pattern> = quote:<id>/coachingTone/<style>:<id>/summary · /rating · /model · /callDateinternal/logger; expose a health indicator in SettingsView so operators know when the memory substrate is stale.GET /api/signal/memory/suggestions?fundId=&founderName= — queries Hermes Honcho for cross-call patterns and returns ranked match candidates.SignalView: "What else does this fund care about?" powered by memory queries — visible insight even before v2 matching ships.SettingsView.MatchBox uses two hill-climb loops that never leave claws-mac-mini. AutoAgent evolves the MatchBox harness itself — prompt, validator, tool chain — against a held-out benchmark of real calls. AutoResearch evolves each fund's local fine-tune as its corpus grows. Both honour the PII invariant: no transcript crosses the Pi-harness boundary.
internal/signal/analyzer.ts + prompt + validators). AutoResearch mutates what gets run against (the per-fund Gemma fine-tune). Same closed loop (ρ · σ · ι · ε · κ), different target, different score function.
The MatchBox analyzer itself is an agent harness. It has a prompt, a tool surface, a validator chain. AutoAgent treats it as the mutation target and hill-climbs against a benchmark of held-out calls.
program.md at repo root — the only file a human edits regularly. Directive + constraints (no external inference, preserve JSON shape, keep transcript-excerpt truncation).tasks/ — held-out Harbor-format tasks. Each task: a transcript + an expected CallSignal shape (human-annotated rating, sentiment, red-flag categories).results.tsv — append-only round log (score · patch · kept/discarded). Gitignored.--agent-import-path matchbox-analyzer:analyze — never reaches out to an external LLM.allowExternal=false so it can't leak transcript fragments into an external provider's context.Each fund's LoRA adapter at ~/.hermes/models/fund-<id>/ is a living artefact. As the fund's corpus grows, the adapter gets retrained via mlx_lm.lora. AutoResearch mutates the fine-tune config (LoRA rank, target-modules, epochs, LR schedule, prompt template) via JSON Patch against a fund-scoped config, and hill-climbs against the fund's held-out eval suite.
The sibling GTM-autoresearch deployment uses Cloudflare KV as the version store. For MatchBox, the version store would contain fund-scoped behaviour signal derived from transcripts — still PII-adjacent. So the drift history lives on-host:
~/.hermes/versions/fund-<id>/ — one repo per fund, one commit per round. Each commit contains the applied JSON Patch plus the candidate config. Replayable.~/.hermes/versions/manifest.db — round log (round · fund · score · patch-sha · model-sha · kept/discarded). Queryable without checking out 47 fund repos.allowExternal gate is set to false on every meta-agent session.results.tsv. The sanitiser strips any literal transcript excerpt the meta-agent pasted into its reflection notes.mlx_lm.lora on the Mac mini's Metal GPU; no cloud training pipeline. Candidate evaluation and scoring also run via mlx_lm.server — the framework handles both inference and training in one Python environment.| Loop | Score | Threshold |
|---|---|---|
| AutoAgent (harness) | Mean F1 over sentiment dimensions + red-flag precision + coaching-quality + fit-accuracy · evaluated on the shared held-out task set | Ship when F1 ≥ 0.85 on the held-out set, with no regression > 0.03 on any dimension |
| AutoResearch (fund fine-tune) | Fund-scoped F1 · coaching-tone match · red-flag precision against the fund's own held-out corpus | Promote candidate when ≥ current by ≥ 0.02, with fund-specific partner sign-off |
results.tsv (AutoAgent), manifest.db + per-fund git repos (AutoResearch). Both exportable as bundles if partners audit.What's on the calendar, and what needs to land before Forbes rollout broadens.
| Owner | Item |
|---|---|
| Jake | Meet Jon Forbes Friday — define compensation structure |
| Jordan | Create technical guides and development assets (this guide + wiki are the first two) |
| Jordan | Scaffold Hermes dual runtime — memory writes + delivery peer · start accumulating corpus before v2 matching |
| Team | Follow-up meeting scheduled for Wednesday or sooner |
| Focus | Finalize the Forbes partnership terms before broader rollout |
package.json name + repo URL (still inherited from ClawBox fork).Who's involved and what they own.
Business lead. Owns the Forbes relationship and the Friday compensation meeting. Runs founder intake through Airtable.
Technical lead. Owns technical guides, development assets, and the Signal / Matching delivery surface.
Primary partner. Proven software funding pipeline; wants expansion into ambiguous industries. MatchBox plugs into his existing deal flow.
Seed node for network depth. Anchor fund for the case study.
Parent org. Ships MatchBox alongside ClawBox, OpenClaw, and the broader market projects.