The model proposes, code decides
The model only fills a typed facts object. Everything with consequences (trip state, ETA, alerts) is plain code reading those facts after the output guard has checked them against what the driver actually said.
FleetVoice
A WhatsApp agent for truck drivers that turns Tamil, Hindi, or English voice notes into trip updates, where the model proposes facts and plain code decides what happens.
- Role
- Designed and built
- Where
- DevFest Chennai 2026 hackathon
- When
- Oct 2026
- Access
- Private repo
Problem
Truck drivers report delays and problems by voice, in Tamil, Hindi, or English, often on a bad connection. Dispatchers and customers need structured updates fast, and a model that mishears a number or invents a place can send the wrong alert.
01
Acknowledge fast, never twice
The webhook verifies the signature, returns immediately, and processes the message in the background. A message ID it has already seen is ignored.
02
No model when none is needed
Greetings, button taps, location pins, and commands are handled entirely in code. Only free text, voice, and photos go further.
03
An input guard before any model
Control characters are stripped, text is capped, and injection phrases are caught by rules. A flagged message never reaches a model and can never change a trip.
04
Cheap first, larger only when unsure
A TypeSafe Jev decision model extracts typed facts in one call and is accepted only above a learned confidence bar. Below it, the message goes to Gemini, then NVIDIA NIM, then plain rules.
05
Circuit breakers per provider
Three failures open a provider's breaker, the cool-down doubles up to ten minutes, and a half-open probe brings it back. Background health checks run every two minutes.
06
The output guard
Facts are clamped to the schema, delay numbers must come from the driver's own words, places the driver never mentioned are dropped, and safety words force a human. The model has no tools.
07
Code acts, templates speak
Trip status, ETA, and flags are updated in plain code, and only fixed templates go to customers and dispatchers. Urgent or unsafe reports go straight to a person.
08
A bar that tunes itself
Drivers' correct and wrong taps become few-shot examples and move the confidence bar: up when accuracy drops below 85%, down when it passes 95%, always between 0.60 and 0.98.
Measured
- Model paths plus a rules fallback
- 4
- Failures before a provider's breaker opens
- 3
- Deadline before a guaranteed reply
- 25 s
- Tools given to the model
- 0
How it's verified
- Harness
- A fixed cascade (Jev, Gemini, NVIDIA NIM, rules) with a circuit breaker per provider, a 12-second per-call and 25-second per-message deadline, and SQLite state.
- Evals
- Online only: driver feedback on each answer is tracked per path and tunes the confidence bar; there is no offline eval set yet.
- Guardrails
- Rule-based injection blocking before any model, schema-clamped typed output, checks that numbers, places, and quantities came from the driver, no model tools, template-only outbound messages, and human escalation.
- Testing
- 6 offline smoke tests; built in a hackathon, with no CI yet.
In depth
Every message gets an answer
A 25-second safety net replies to the driver even if every model is slow or down, and the message is replayed later once a provider recovers.
Decisions
- 01A cheap typed classifier first and larger models only when it is unsure.
- 02No tools for the model: it can describe the world, not change it.
- 03An instant webhook acknowledgement plus a guaranteed reply, because drivers will not wait.
Stack
Python · FastAPI · SQLite · WhatsApp Cloud API · TypeSafe Jev · Gemini · Groq Whisper · NVIDIA NIM · OpenRouteService · Docker