Teliport S4 · Case 5
offline
CAN · COVESA VSS
private code

VehicleGPT

An offline assistant for a car's own logged signals. A small language model turns each question into one verified tool call on a read-only store, so the model never reads raw data.

my role: team lead · team of 3
stack: Python · llama.cpp · CAN · COVESA VSS
when: oct 2026

cat problem.md

Team entry to Tata Elxsi Teliport Season 4, Case 5. The proposal is an in-vehicle assistant that answers natural-language questions about the vehicle's own data, with no cloud connection.

Deterministic code ingests and stores the signals and detects events. A small language model, 2B parameters or fewer, only turns a question into one tool call, constrained to a flat JSON schema. Code checks the call against the signal catalogue, runs parameterised SQL, and checks that every number in the reply matches a query result.

tree pipeline/

  CAN tap (SocketCAN, listen-only)
  VSS catalogue
        │
        ▼
  read-only store
        │
        ▼
  small LM routes the question
        │
        ▼
  verified tool calls
        │
        ▼
  answer

A CAN tap in listen-only mode and the VSS catalogue feed a read-only store. A small language model routes each question to one tool call, which is verified before it runs. The reply is written from the rows the call returns.

cat results.tsv

model benchmark, CPU

metric	result	setup
quantized model variants	12	2,400 checksum-audited results
MiniCPM5-1B Q8_0 accuracy	82%	first 50 questions, hybrid pipeline
median latency	1.74 s	same run
peak memory	1,437 MiB	same run

x86-64 CPU (AMD Ryzen AI 7 350), 2 cores, 2 threads, 2 GiB container memory, no swap. The hybrid pipeline is the rule router plus the model.

system, hybrid rules and model

metric	result	setup
rules only	29.5%	200 questions, GPU
rules and small model	up to 95.5%	200 questions, GPU; 72.0% to 95.5% across seven models
schema-valid replies	all	every model in the hybrid run

The 200 questions run against a synthetic 21-day vehicle store, with gold answers from the gold tool calls.

cat limits.md

The CPU accuracy, latency and memory figures come from the first 50 questions, not all 200. Each configuration was run once, deterministically.
The data is synthetic. The CPU runs used an x86 proxy, not the ARM target board.
The rule router misroutes 4 of 50 questions, which caps tool choice at 92% for every model in the hybrid arm.
The in-car architecture (read-only tap, safety partition, privacy controls) is a design. The measured parts are the model routing and the tool calls.

Code is private while the competition runs.

$ cat problem.md

$ tree pipeline/

$ cat results.tsv