Siphyy

Open-source agentic framework for fleet telematics analytics — fuel anomaly detection, predictive maintenance, and driver behavior intelligence, with a provider-agnostic canonical schema and structured LLM reasoning.

What it does

Fleet operations data lives in dozens of telematics platforms (Samsara, Geotab, Trakzee, Verizon Connect, regional providers, raw OBD-II streams) — each with its own schema, polling cadence, and quirks. Siphyy normalises all of them into a single canonical event stream, runs cheap deterministic detectors over that stream, and uses an LLM agent grounded in a curated case base to interpret only the events that matter.

The output is structured, auditable, and ready to be wired into fleet management systems — not free-text mechanic advice.

Architecture in 30 seconds

   [Telematics Provider]    ←  Samsara, Geotab, Trakzee, OBD-II, ...
            │
            ▼
   [Adapter]                ←  Translates provider → canonical schema
            │
            ▼
   [Tier 1: Detector]       ←  Cheap rule-based detection. No LLMs.
            │
            ▼ (interesting events only)
   [Tier 2: LLM Agent]      ←  Agentic RAG over case base + OEM manuals
            │
            ▼
   [Structured Output]      ←  FuelAnomalyReport, MaintenanceRiskReport, ...

Quickstart

The fastest way to see the framework in action:

git clone https://github.com/siphyy/siphyy-core.git
cd siphyy-core
uv sync                              # installs everything incl. dev tools
uv run python examples/quickstart.py

The quickstart loads examples/data/sample_trakzee.json (17 anonymised, synthetic rows), runs the full pipeline, and prints structured fuel-anomaly reports. It auto-detects which LLM provider to use:

• OPENAI_API_KEY in env → uses OpenAILLMClient (real call against gpt-4o-mini)
• ANTHROPIC_API_KEY in env → uses AnthropicLLMClient (real call against claude-haiku-4-5)
• Neither → uses MockLLMClient with realistic canned verdicts, so the demo works with zero setup

Drop a .env at the repo root with your key(s) and the quickstart picks them up automatically (python-dotenv is in the dev extras).

Install just the library

pip install "siphyy[trakzee,llm]"
# or
uv pip install "siphyy[trakzee,llm]"

The minimal in-code version of what the quickstart runs:

from siphyy.adapters import TrakzeeAdapter
from siphyy.agents import FuelAnomalyAgent, OpenAILLMClient
from siphyy.detectors import FuelSiphonageDetector
from siphyy.knowledge import SEED_CASES
from siphyy.schema import CaseBase

# 1. Translate raw provider rows into canonical TelemetryReading events.
adapter = TrakzeeAdapter()

# 2. Tier 1: cheap, deterministic rules surface candidates.
detector = FuelSiphonageDetector()

# 3. Tier 2: an LLM-grounded agent interprets each candidate against
#    historical cases. The LLMClient is pluggable — swap OpenAILLMClient
#    for an Anthropic / Gemini / Ollama implementation without touching
#    agent code (see "Bring your own LLM" below).
agent = FuelAnomalyAgent(
    llm_client=OpenAILLMClient(),  # reads OPENAI_API_KEY from env
    case_base=CaseBase(SEED_CASES),
)

for event in adapter.adapt(trakzee_rows):
    if (interesting := detector.process(event)) and (report := agent.process(interesting)):
        print(report.assessment, report.confidence, report.summary)

Bring your own LLM

LLMClient is a Protocol with a single method. To use a different provider, implement it in ~20 lines:

from pydantic import BaseModel
from siphyy.agents import LLMClient

class MyAnthropicClient:  # satisfies LLMClient structurally
    def complete[T: BaseModel](self, *, system, user, response_model: type[T]) -> T:
        ...  # call Anthropic's tool-use endpoint, return response_model(**parsed)

The agent itself doesn't import anthropic or openai — only the client implementation does, so optional extras stay optional.

Why open source

The framework is Apache 2.0 and always will be. The plumbing — canonical schema, adapters, detectors, agent code — is the commodity. What we sell separately is the Siphyy Knowledge Pack: curated case bases, indexed OEM manuals, calibrated detection thresholds, and the enterprise dashboard. You can use the OSS framework without ever paying us a cent.

Writing your own adapter

The whole point of the canonical schema is that you don't have to wait for us to support your telematics provider. Subclass TelematicsAdapter in ~50 lines and you're integrated. See docs/writing-adapters.md — TrakzeeAdapter is the reference implementation.

Project status

Alpha. The canonical schema, the Trakzee adapter, the FuelSiphonageDetector, and the FuelAnomalyAgent are stable today. Additional provider adapters, additional detectors and agents, the CLI, and the storage backend are still in active development — track CHANGELOG.md for what lands when. Pin to a specific version until 1.0.

License

Apache 2.0. See LICENSE.