FORG's intelligence layer runs entirely on statistical ML — no LLM in the real-time path. Every anomaly, classification, and recommendation fires in microseconds, not seconds.
Three layers, zero LLM latency. The Go agent collects signals and ships them encrypted to the Rule Engine. A statistical ML engine classifies each event — cost anomaly, model drift, off-hours usage — before any rule fires.
FORG Atlas lives entirely outside the real-time path. It runs as a batch process against your Supabase + pgvector index, surfacing natural-language answers in the dashboard only — never blocking a hook evaluation.
┌─────────────────────────────────────────┐
│ FORG Agent (Go binary) │
│ signal collector · zero intelligence │
└──────────────────────┬──────────────────┘
│ encrypted signals
▼
┌─────────────────────────────────────────┐
│ Statistical ML Engine │
│ · cost classifier · anomaly detector │
│ · model usage profiler │
│ · drift monitor │
│ < 2ms per evaluation │
└──────────────────────┬──────────────────┘
│ labelled events
▼
┌─────────────────────────────────────────┐
│ Rule Evaluation Layer │
│ · profile lookup · threshold check │
│ · alert dispatch · audit chain │
└──────────────────────┬──────────────────┘
│ (dashboard only)
▼
┌─────────────────────────────────────────┐
│ FORG Atlas (batch) │
│ · pgvector index · NL query engine │
│ · grounded answers · usage analytics │
│ NOT in real-time path │
└─────────────────────────────────────────┘
FORG Atlas answers questions about your historical usage data. It runs in the dashboard — never in the hook evaluation path.
Engineering led GPT-4o spend last week with $412.80 — 54% of org total. Breakdown by sub-team:
3 sessions found in the last 30 days:
Statistical anomaly detection fires in < 2ms per session evaluation.
user dan@acme.com consumed $84.20 in 23 minutes — 9.4× their 7-day average.
engineering team moved from Claude 3 Haiku → GPT-4o for code review tasks. Projected cost impact: +$320/mo.
3 automated sessions ran between 02:00–04:00 UTC with no rate limiting. No policy covers this window.
FORG profiles each session by task type and surfaces the cost-optimal model with evidence.
Highest pass@1 on HumanEval with 4× lower cost vs GPT-4o
Superior coherence over 8K+ tokens; avoids repetition artifacts
~1ms TTFT, 200× cheaper than GPT-4o for retrieval-only tasks
Statistical ML anomaly detection, FORG Atlas dashboard queries, and model recommendations — all included.