Marketing intelligence, agent-run.
5 briefs · 2 modules · 1 platform — P2 retail + P5 P6 P7 P11 gaming
One platform that listens to your customers, learns which creative makes money, makes the next round, and proves it works — with a human hand on every trigger.
Nikhil Mani · built with Claude
LISTEN — every review & CS transcript, one enriched stream · WATCH — Sentinel hunts spikes, routes issues to teams · COMPARE — same pipeline pointed at competitors.
LEARN — theme × pLTV insights, budget reallocation · MAKE — variants graded before spend · PROVE — A/B designed & read out with statistics in code.
Shared spine: one design system, one agent runner — every run step-logged and Langfuse-traced — and workspace config as the only place brand names live. Modular per brief, integrated as one story: customer signal → creative decision → proven test → back to signal.
796 signals, 4 source types — site reviews, CS chat/call/email, competitor reviews — one enriched feed: sentiment, topic, intent, severity, incentivized-flag. Vietnamese native, accuracy eval-gated.
Detects spikes vs baseline in code → investigates the cluster → writes root cause citing actual customers → routes to the brief's named teams, human ack/resolve. In hours, not weeks.
Hasaki and Watsons — the brief names them. 592 real scraped reviews through the identical pipeline: share of voice & sentiment by topic. Steal their post-mortems.
Replaces the monthly deck. Top movers, emerging issues, competitor deltas — evidence linked.
"Top delivery complaints this month vs competitors?" — answers with charts and cited reviews.
Vision on review photos — damaged products spotted automatically.
$2.1M spend, 60 creatives, 8 themes — every theme carries two returns: week-one ROAS and lifetime (installs × predicted LTV), per channel, per segment.
Event-offer: 1.11 week-one ROAS — 2.78 on whales. Week-one says kill it; lifetime says it prints. That's the brief's "high-value players versus tourists" on one card.
Flags winners, fatigue, value-divergence and payer-depth with cited creative IDs. Thresholds are code; the model explains, never invents math.
Feature-list on Facebook: ROAS −37.6% in 7 days while spend held at ~$109k/month. Caught in week one, not quarter's end.
Approved shifts pushed through Meta/TikTok APIs — human click required.
"Lift in pLTV-weighted ROAS on agent-recommended themes."
Campaigns that look cheap on install cost but acquire low-pLTV cohorts get called out — and the inverse: expensive-looking themes quietly acquiring whales.
"Move $X/day from theme A on Facebook to theme B on TikTok" — projected pLTV-ROAS delta attached, capped at ≤30% of daily spend so ops can actually execute it.
Whale / payer / new / lapsed ROAS skew surfaced per theme — reach is not value.
The copilot recommends with receipts; nothing moves money without a click.
"ROAS improvement from agent-recommended reallocations."
Winning theme → concept briefs + copy variants per channel and market (VN + EN). Hard length limits enforced in code: FB 125 · TikTok 40 · Google 30 · in-app 60.
Every variant pre-flight graded against winning evidence: SHIP / REVISE / KILL + an honest ROAS band from historical quartiles.
Tone rules and banned claims come from Workspace config — the generator and evaluator obey them, not suggest them.
"Ready for A/B" means a human clicked approve. The agent proposes; your team disposes.
Static and video creative generation against the same brand kit and evidence.
Market-specific variants with back-translation QA.
"Time-to-first-variant; win rate of generated versus baseline creatives."
Hypothesis in → metric, sample size per arm (two-proportion power calc, α .05 / power .8), duration at your real traffic. All math in code, not in the model.
Two-proportion z-test in code decides the verdict — significant win → ship, significant loss → kill, ambiguous → iterate with exactly how many more days to run. The LLM only narrates.
Three seeded experiments return exactly ship / kill / iterate — 3/3 closed with a statistically-sound decision. Verify the z-scores by hand.
Institutional memory: every test, every decision, searchable.
"% of experiments closed with a statistically-sound decision; cycle time per test."
Ad exports (Meta, TikTok, Google), MMP/pLTV feed, marketplace & Google reviews, app stores, CS webhooks. API connectors COMING SOON
Tone rules, banned claims, format constraints — enforced by the generator and evaluator, not suggested.
Owner, Analyst, Creative, Reviewer. Agents recommend; humans click.
Every agent step traced in Langfuse — inputs, decisions, cost. Your analysts can check the math.
Slack, Zalo, email routing per team.
Retailer today, game studio tonight — same platform, different workspace. No code changes.
of manual reporting effort gone — the brief's own target, hit by the one-click exec summary
not quarters, to kill fatigued creative — with the burn in dollars
budget follows winning themes — proven by statistically-sound tests, not gut
Issue spikes caught in hours protect retention; competitor benchmarks show commercial where to attack.
Listen. Learn. Make. Prove. Repeat.
Five briefs — P2 · P5 · P6 · P7 · P11 — chained exactly as written.
Nikhil Mani · AABW 2026 · live demo + source available