Engineering

The math that runs the ads.

EgoBoson is a decision system, not a dashboard. It ingests spend, revenue and pacing every minute, evaluates against a journaled rules engine, and mutates Meta + Google state through native APIs. No screen-scraping. No human in the hot loop. Manual override always wins.

60s

Loss-detection latency

High

Mutation throughput

99%+

Apply success rate

Native

API integrations

See architecture Sales overview

Architecture

Four layers. One operator surface.

The platform is intentionally thin. Data from four authoritative sources, policy in a journaled rules engine, mutations through native APIs — never UI scraping. The operator sees one hierarchy.

Layer 01 · Ingress

Authoritative sources only

Every signal has a canonical source. Your tracker for revenue and lead attribution. Meta Marketing API for live spend and structural state. Google Ads API through Manager Account. Your asset store for creative. Zero UI scraping, zero brittle selectors.

60s refreshNative APIsSchema-typed

Layer 02 · Policy

Rules engine, journaled

Stop guards fire every 1 minute. Bid dial and bump ladder run every 15. Horizontal scatter rebalances winners across accounts every 1 hour. Every decision recorded with confidence, reason, and rollback handle.

1min stop15min bid1h scatter

Layer 03 · Egress

Native API, circuit-broken

All mutations through Meta Marketing API v22, Google Ads API v18+, and FB Pages for dark-post reuse, branded content, partnership ads. Upstream rate limits respected via shared circuit-breaker — we cool down before FB tells us to.

v22 / v18Circuit-brokenIdempotent

Layer 04 · Surface

One tree. Everything mutable.

A unified live tree spanning your full portfolio. Budgets, bids, statuses — all inline, all mutable. Rules editor, terminal, Telegram bridge on the same surface. Operator overrides flow through the same journaled rules path as automated ones.

Live treeRules editorTG override

Decision loop

How a decision happens.

Every 60 seconds the platform re-reads the portfolio. Every 15 minutes it checks for bumps and bid dial moves. Every decision has 5 stages — and every one is journaled.

01 · Read

PFM tree + Meta v22 spend, structure

~2s · full sweep

02 · Classify

Lifecycle state · zone · tier

in-mem · 0ms

03 · Decide

Rules engine · band-aware

deterministic

04 · Mutate

FB API v22 · circuit-broken

~700ms · idempotent

05 · Journal

Decision · reason · rollback

JSONL append

No-lead stop v7

if leads = 0 AND spend ≥ $20 AND age ≥ 30min
then PAUSE adset · 24h dedup · HV only

Force-bump v6

if budget < $50 AND leads ≥ 1 → bump $50
if budget ≥ $50 AND leads ≥ 2 → +100% · 1h cooldown

Data flow

Sources in. Mutations out.

Five inputs, one policy layer, two mutation channels. Every signal carries provenance and a refresh timestamp. Nothing is cached beyond what policy allows.

→ PFM lag typically 30–60s. Critical decisions verify against live FB spend before acting.

→ Token routing is per-account — the right system-user is picked automatically, never hardcoded.

→ Mutations are journaled with confidence + reason + rollback handle. Every change is reversible.

PFM→GraphQL · stats per ad set (spend, revenue, leads)60s

META→v22 · live spend, structure, bid strategyon-demand

GOOGLE→v18 · GAQL · campaign, keyword, asset stats5min

DBOX→chunked video · thumb extraction · dedupon-launch

POLICY→rules engine · confidence + reason recordeddeterministic

MUTATE→Meta / Google · journaled · circuit-broken~700ms

API surface

No magic. Open APIs.

Every signal we ingest and every mutation we apply is on the public side of Meta and Google. Below: the actual queries and shapes the platform uses.

// Meta Marketing API — pause an ad set. Idempotent + journaled.
POST https://graph.facebook.com/{version}/{adset_id}
Content-Type: application/x-www-form-urlencoded

status=PAUSED&access_token={token}

// Response
{ "success": true }

// Token routing — per-account, never hardcoded
const tok = tokenFor(accountId);
// → routes to the right system-user automatically

// Google Ads API v18 · GAQL · search_term_view · last 7d
SELECT
  search_term_view.search_term,
  metrics.cost_micros,
  metrics.conversions,
  metrics.cost_per_conversion
FROM search_term_view
WHERE segments.date DURING LAST_7_DAYS
  AND metrics.cost_micros > 10000000
ORDER BY metrics.conversions DESC

// Enhanced Conversions for Leads — push hashed offline conversions
{ "conversion_action": "customers/123/conversionActions/456",
  "user_identifiers": [{ "hashed_email": "sha256(...)" }],
  "conversion_value": 42.50 }

// Every mutation produces a journal entry. Append-only JSONL.
{
  "ts": "2026-05-03T17:36:44.965Z",
  "adset_id": "{adset_id}",
  "action": "park",
  "from_bid": 1400,
  "to_bid": 1080,
  "reason": "park-as-halve: consecutive lowers",
  "confidence": 0.85,
  "rule_version": "bid-autopick-v2",
  "t0": { "roi": -58, "spend": 66.4, "leads": 1 },
  "t60_due": 1714756004965
}

// Effectiveness tracking — T+0 / T+60 delta per decision
// Health endpoint exposes 4h / 24h aggregate

Production metrics

Real numbers. From real ops.

Last-30-day operating numbers from the production deployment. Updated nightly.

Loss-detection latency · p95

54_seconds

95th percentile time from PFM signal to FB pause call. p99 = 78s.

↓ 12% vs prior 30d

Bid-dial cycles / day

Cron tick every 15 min. Avg 8.4 mutations applied per cycle.

stable

Mutation success rate

99.7_%

FB + Google API success. Failures auto-retry with exponential backoff.

+0.3pp vs prior 30d

Active ad sets

847_avg

Concurrent active across the full portfolio.

stable

Circuit-breaker trips

2_/30d

FB rate-limit cooldowns. 30-min auto-reset.

↓ from 8 in prior 30d

Effectiveness · ΔROI within 60min

+11.4_{pp avg}

Average ROI improvement on adsets touched by autopick after T+60.

measured

Local LLM · on-prem

A model that thinks like an operator.

Trained on real decisions — not the open web. Runs on dedicated GPU inside your perimeter. Zero data egress to external LLM providers. Cross-reviewed by an external model only on critical cases — and only on the evidence, never raw account data.

Always-on context. The model reads a fresh snapshot of your portfolio on every query — never stale memory.

Tool-using. Reads tracker state, rules history, lifecycle records — answers grounded in real data.

Detector sweep. Flags rule violations and dangerous patterns on a tight cadence; deduplicated to never spam.

Pair-reviewed. Critical findings get an independent second opinion before reaching humans.

Advisory by default. Any mutation it proposes routes through human approval. No autonomous writes.

tuned model · listening

Watching every account. Reading every signal.
Reasoning where it counts. Quiet where it doesn't.

Partners & stack

We don't reinvent what works.

EgoBoson composes best-in-class third-party systems with a proprietary decision layer. Our partners carry the infrastructure we'd never build in-house.

Your tracker (GraphQL)

Canonical source of tracker-attributed performance — spend, revenue, lead counts. We integrate with any tracker that exposes a GraphQL or REST surface. Drives every monitor and bid decision.

any GraphQL · any REST

Meta Marketing API

Direct integration with Facebook & Instagram advertising — campaigns, ad sets, creatives, branded content, insights. No UI scraping, ever.

developers.facebook.com →

Google Ads API

MCC-level Search / PMax / Demand Gen — GAQL reporting, portfolio bidding, offline conversion upload for Smart Bidding.

developers.google.com →

Dropbox

Creative asset storage, automated ingest pipeline. Chunked video upload, thumbnail extraction, deduplication across your full portfolio.

dropbox.com/developers →

Security & data

Your data. Your tenant. Your rules.

Tenant isolation

Each customer runs on isolated infrastructure. No shared databases. Tokens scoped per account. Enterprise tier offers VPC peering.

No password handoffs

Meta system-user tokens, Google service-account keys, Dropbox OAuth refresh. You authorize once; we never see passwords.

Audit log export

Every mutation, override, login is journaled. Export to your SIEM in standard JSONL. Enterprise tier available.