Dealflow Intelligence

A venture fund’s memory lives in its CRM. If a meeting you took last week never made it into Affinity, or a deal sits at “Term Sheet” in your head but “3rd Meeting” in the database, then the Monday dealflow review runs on fiction — and the fund makes decisions on a pipeline that quietly disagrees with reality.

Dealflow Intelligence is the three-stage pipeline that keeps the disagreement from happening: it pulls the deal list out of Affinity, enriches each active deal with its real meeting history, and hands Mundi — the financial-intelligence Jedi — a short brief of what actually needs a human. In the end, it is a machine for making sure the most expensive meetings of the week left a trace.

The pipeline

Three scripts under ~/.openclaw/scripts/, each owning one boundary:

Affinity ─▶ v_deal ─▶ prep ─▶ Mundi ─▶ Signal
  v2 REST    DuckDB   v1 API   4h    Force Flow

Stage	Script	What it does
Sync	affinity-sync.py	Pulls the deal lists from the Affinity v2 REST API into workspace.duckdb (the v_deal view). Daily, com.sanctum.affinity-sync.
Enrich	dealflow-meeting-prep.py	Reads the active pipeline from v_deal, fetches each deal’s real meeting/email recency, and buckets it.
Brief	mundi-briefing.py	Every four hours (com.mundi.briefing), folds the enrichment into a terse brief and pages it through Force Flow to Signal.

The active pipeline is the ~60 deals in a decision status — Meeting Pending, To be Reviewed, 2nd/3rd Meeting, Due Diligence, Term Sheet, Refer to LP — not the nine thousand discarded first-screens, and not the standing “Weekly Review” list that fills with deals nobody has touched since 2022.

What it surfaces

Every healthy active deal lands in one of four buckets (a fifth, OUTAGE, is covered just below), and the brief leads with the only one that matters at 8 a.m.:

• NEEDS YOU — you met or emailed them and nothing is booked next. The ball is in your court.
• SCHEDULED — a future meeting is on the calendar. (A meeting whose date has passed does not count as scheduled; that was a real bug, and it hid deals that needed action.)
• COLD — an active deal with no contact in three weeks.
• UNRESOLVED — the deal carries a stale CRM id that no longer resolves. Cleanup, not crisis.

A failure is never a quiet day

The hard rule of the pipeline: an outage must never look like good news. If the Affinity API throws auth/rate-limit/5xx errors, if the sweep can’t reach the database, or if the sync has gone stale, the brief says so loudly — it never degrades into a cheerful empty “nothing to report.” dealflow-meeting-prep.py even exits 3 on a systemic outage so the briefing layer can tell a real failure apart from a genuinely quiet pipeline.

Boundaries that bite

Most of the engineering here is at the seams between layers, where the obvious thing is wrong.

Meeting intel comes from the v1 API, not v2. Affinity’s v2 relationship-intelligence field returns zero fields for most companies — one well-connected deal happened to populate, which masked the gap until a wider sweep exposed it. The reliable source is the v1 /organizations/{id}?with_interaction_dates=true endpoint. The lesson is the standing one: a contract that passes on one happy row is not tested.

The sync is atomic. affinity-sync.py writes both lists inside a single transaction. A partial failure rolls back rather than leaving v_deal half-applied yet stamped fresh — which would blind the freshness guard into reporting health over a broken pipeline.

workspace.duckdb is single-writer

The central store allows one writer at a time. When a sibling ingest job (mail, calendar, messages) holds the write lock, the sync waits and retries with backoff before alerting — it does not crash, and a read that genuinely can’t get through is reported as an outage, never as an empty pipeline. If you add a new writer, honor the doctrine: hold the lock for seconds, not minutes.

The freshness guard exists because the last sync died silently for two months. The previous Rube/Composio integration failed quietly and nobody noticed until the pipeline was a quarter stale. Mundi now checks how long it has been since v_deal was refreshed and warns past thirty hours — and treats a failed freshness read itself as a symptom, not as “fresh.”

Why meeting notes aren't auto-imported

Granola holds the richest meeting notes, but its local store is encrypted and its API is interactive-only — unreachable from a scheduled job. So the pipeline reconciles meeting recency automatically and leaves note-writing to a human-in-the-loop step, rather than pretend a cron can read an encrypted database.

Configuration

# Cadence (launchd)
com.sanctum.affinity-sync:  daily            # Affinity -> v_deal
com.mundi.briefing:         every 4 hours    # enrich -> Force Flow -> Signal

# Tunables (in-script constants)
ACTIVE_STATUSES:  decision-stage statuses only
cold-days:        21        # quiet-deal threshold
STALE_HOURS:      30        # sync-freshness alert (daily sync + grace)
cache:            ~20h      # meeting-intel cache, so the 4h run sweeps once/day

# See the active pipeline on demand
~/.openclaw/scripts/dealflow-meeting-prep.py            # human-readable
~/.openclaw/scripts/dealflow-meeting-prep.py --json     # for the briefing

Tested end to end

The pipeline ships with 49 unit tests over the pure logic — bucket classification, status dedup, Affinity field extraction, HTTP-error categorization, the briefing’s outage-vs-quiet partitioning — plus an end-to-end harness, tests/e2e_dealflow.py.

The harness crosses every real boundary in order — Keychain, Affinity v2, the DuckDB transaction, the v_deal read, the v1 interactions call, Mundi’s consumer contract, the freshness guard, and a live Force Flow page. Structural assertions prove a field exists; the harness proves a real artifact crossed all of them before it ever reached your phone — which is the only proof a Monday brief is worth trusting.