Tool concept · event intelligence

Event Research Orchestrator

A repeatable workflow for turning a messy Luma / Financial Club attendee list into a practical networking brief: who is in the room, what they actually do, which companies matter, who Sasha should approach first, and what to ask.

Built from the AI Agents event run 30 micro-agents 165 profiles QA-first, not hype-first

What the tool answers

Not “who registered?” as a flat list. The useful question is whether the room contains buyers, partners, investors, serious peers, or just vendor noise.

Who is this person?

Identity, current role, company, confidence, public profile links, and enough career context to avoid awkward shallow networking.

identityexperienceconfidence

What do they actually do?

Company/product stripped of marketing language: inputs, process, outputs, buyer/user, and where money, risk, compliance, or AI enters the flow.

plain productworkflowtraction

Should Sasha talk to them?

Score 0–5, category, recommended action, conversation hooks, and a practical follow-up path for CashQ / Finori / regulated AI workflows.

buyerpartnerinvestorpeer

Generated outputs

The detailed research stays usable later; the live brief stays small enough to open on a phone at the event.

Full dossierAll profile notes, evidence, caveats, and raw chunk outputs.

Ranked targetsA/B/C/skip, score, category, identity confidence, QA flags.

Live cheatsheet10–15 people max: why talk, opener, one useful question, follow-up route.

Follow-up noteOnly real conversations: what was discussed, next step, draft message.

Reusable TODOManual QA queue, process improvements, export/sync tasks.

Coverage reportWhich source requirements were met, partially met, or blocked by runtime/search limits.

Quality gate

The orchestrator’s job is not just to spawn agents. It checks that outputs exist, parse, cover every attendee, and do not hide uncertainty.

All source rosters fetched and deduped.

Every attendee has a profile, even if sparse.

Each profile has evidence URLs or explicit “not found”.

JSON outputs parse before merge.

Schema fields are normalized across agents.

Sasha / internal people excluded from target ranking.

Weak identity matches are flagged, not papered over.

High-priority people get manual QA before outreach.

Known limits from this run

This is the honest part: the run was useful, but expensive and not perfectly executed against the original “max reasoning across models + Exa everywhere” ideal.

Area

What happened

Product implication

Model routing

GPT-5.5 max reasoning / GPT-5.5-2 / GPT-5.4pro were requested, but subagent routes with xhigh/reasoning rejected or timed out.

Default future run should test model route first, then fan out only on verified routes.

EXA was included in instructions, but not uniformly used by every subagent; some agents hit search/LinkedIn/tool limits and used public fallbacks.

Add an explicit orchestrator-side Exa/person-search stage before spawning agents.

Cost

Visible minimum was about 3.6M tokens; realistic total closer to 3.7–4.0M.

Use cheaper prefiltering, smaller target set, and “deep research only for likely A/B” mode.

Confidence

109 profiles carried QA flags due to sparse public data, blocked LinkedIn, or weak identity match.

The tool should separate “event decision quality” from “CRM/outreach quality”.

On mobile, rows stack: area → what happened → implication.

Next product version

Turn this from an ad-hoc swarm into a buttoned-up event-intelligence tool.

Fast mode: should I go?

Fetch attendee list.
Classify obvious founders/operators/investors/buyers.
Return a 5-minute verdict: go / maybe / skip.
Deep research only for top 20.

Deep mode: I am going

Run verified search stack: Exa + normal web + LinkedIn enrichment when quota exists.
Produce mobile cheatsheet.
Create follow-up workspace.
Sync final artifacts to ClavaObsid event folder.