AI runtime optimization layer for agentic systems: compress noisy infrastructure logs into a compact machine-readable state, ground it with web search, then reason over the state instead of the raw logs.
Status: backend and frontend complete (MVP).
flashmesh/
├── backend/ FastAPI service (Python 3.11)
├── frontend/ Next.js 15 + TypeScript + Tailwind + shadcn/ui dashboard
├── data/ sample logs + generator, runtime uploads/results
├── docs/ ARCHITECTURE.md
└── README.md
cd backend
python3.11 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env # add NEBIUS_API_KEY and TAVILY_API_KEY, or set MOCK_MODE=true
uvicorn app.main:app --reload --port 8000Docs: http://localhost:8000/docs
curl -X POST localhost:8000/api/samples/k8s_redis_oom.log # -> upload_id
curl -X POST localhost:8000/api/analyze/<upload_id> | jq .metricsRegenerate the ~28k-token demo log: python data/generate_sample_log.py (run from data/).
Requires Node 20+. Start the backend first.
cd frontend
cp .env.example .env.local # NEXT_PUBLIC_API_URL=http://localhost:8000
npm install
npm run dev # http://localhost:3000Click "Analyze the demo incident" on the homepage, or drop in your own .log file.
On Windows PowerShell use copy .env.example .env.local.
/title, before/after compression visual, upload area, pipeline workflow/analysis/[id]log preview and token count, state graph JSON, dependency chain, Tavily grounding, root cause + remediation + fix + confidence, metrics dashboard with per-stage latency
Results are cached by id on the backend, so refreshing an analysis page does not re-bill the model.
Four simulated tools (restart_service, increase_memory_limit, rollback_release, scale_deployment);
the agent reads the real FlashMesh state and must pick the right one for three incident types.
cd data && python generate_validation_logs.py # creates 2 synthetic logs; swap in real incident logs of the same names
cd ../backend
python -m scripts.validate_agent --baseline # --baseline also runs the agent on the raw logs as a controlNeeds real Nebius credentials (MOCK_MODE=false). Prints a table and writes data/validation/report.json.
Tool arguments are range-checked; malformed or out-of-range calls count as failures.
See backend/.env.example. Set NEBIUS_MODEL to the exact Nemotron model ID shown in your Nebius Token Factory catalog.
docker build -t flashmesh-api backend
docker run -p 8000:8000 --env-file backend/.env -v $PWD/data:/data flashmesh-apiWorks on any container host (Fly.io, Render, Cloud Run). Set CORS_ORIGINS to your frontend URL.
Deploy frontend/ to Vercel (set root directory to frontend) and add the env var
NEXT_PUBLIC_API_URL=https://<your-backend-host>. Then set the backend's CORS_ORIGINS to the Vercel URL.
Self-host alternative: npm run build && npm start.