Work03AI & computer vision

AeroSentry

Multi-agent AI operations layer for autonomous drone fleets, with a human in the loop.

Role
Solo — design, engineering, deployment
Language
Python 99.5% · Dockerfile 0.5%
Timeline
Sep 2026 → Sep 2026
Last push
12 days ago · 1 commits
Source
specialist agents + supervisor
3 + 1
offline tests in CI
17
approval gate on every high-risk action
HITL
latency tracked in golden-set evals
p50 / p95

A LangGraph supervisor routes operator requests to planner, analyst and knowledge agents that gather evidence with validated tools, ground themselves through hybrid RAG, analyse thermal frames — and hand every high-risk action to a human approval gate.

Problem

Agentic systems that only suggest are safe but useless; agents that act are useful but dangerous. Drone operations need an AI that can schedule, inspect and alert on its own, yet can never launch, abort or escalate without deterministic policy checks and an operator’s explicit approval.

Approach

  1. 01

    A LangGraph supervisor loads persistent SQLite memory and routes to specialists; agents run a ReAct tool loop over Pydantic-validated Python functions (telemetry, weather, geofence, schedule, abort).

  2. 02

    The knowledge agent answers from operator manuals and incident logs via hybrid retrieval — BM25 plus dense embeddings fused with Reciprocal Rank Fusion — with citations.

  3. 03

    The analyst reads aerial thermal frames with a vision-language model and falls back to a deterministic numpy hotspot detector; alert severity must be backed by imagery evidence.

  4. 04

    A safety gate applies hard policy checks (battery floor, wind limit, geofence, altitude ceiling) and interrupts the graph before high-risk actions until a human approves via API, console or CLI. Every LLM call, tool call, routing decision and verdict is traced to JSONL; a golden-set eval harness scores routing, tool accuracy, HITL gating and latency in CI.

Results

specialist agents + supervisor
3 + 1
offline tests in CI
17
approval gate on every high-risk action
HITL
latency tracked in golden-set evals
p50 / p95
Created
3 Sep 2026
Commits
1
Default branch
main
License
MIT

Stack

  • Python
  • LangGraph
  • Pydantic
  • FastAPI
  • Hybrid RAG (BM25 + dense, RRF)
  • SQLite
  • Docker
  • GitHub Actions

#agentic-ai #drones #fastapi #human-in-the-loop #langgraph #multi-agent #python #rag

From the notes