lighthouse.ai

slide 1
slide 2
slide 3
slide 4
slide 5
slide 6
slide 7
slide 8
slide 9

Inspiration

Dashboards told us what changed, never why; leaders spent weeks sleuthing while teams burned out. We wanted an “engineering lighthouse” that shines a beam to the root cause and suggests fixes.

What it does

Autonomously detects anomalies, investigates root causes with a ReAct loop, proposes ROI‑ranked interventions, routes approvals, executes, and monitors outcomes—learning via RLHF. Uses z‑scores ( z = \frac{x - \mu}{\sigma} ) to flag significant deviations.

How we built it

Agents in Python (agents/) orchestrated by orchestration/demo_full_workflow.py
ReAct investigator with one‑tool‑per‑step loop and saved CoT JSON (latest_investigation_cot.json)
Mock GitHub/Slack/Calendar tools for deterministic demos
Supervisor for validation/approvals/execution; Outcome Monitor + RLHF
FastAPI endpoint /api/investigation/latest/cot for the frontend

Challenges we ran into

Tool‑call protocol mismatches causing API errors
LLM compressing all reasoning into one step (fixed with loop + prompt rules)
Reliable JSON parsing/serialization and import ergonomics

Accomplishments that we're proud of

Transparent, multi‑step reasoning trace with evidence
Clean API for CoT visualization
Guardrailed supervisor flow and simulated outcome uplift
Judge‑friendly README and structure

What we learned

Strict interaction rules (one tool/step, explicit CONCLUDE) boost agentic quality
Referencing prior observations makes CoT auditable and useful
Even lightweight RLHF compounds value org‑by‑org