OPEN SOURCE · RESEARCH PROTOTYPE

Evidence
before answers.

An evaluation-driven multimodal agent for examining trees and wood from photographs. Built to distinguish what an image shows from what a model merely suspects.

Explore the repository ↗How it works ↓

Traceable observationsAdversarial reviewBounded confidenceExplicit abstention

01 / Philosophy

Designed to know
when not to know.

A plausible label is not the same as a defensible conclusion. Dendro Inspector treats evidence quality, provenance, and uncertainty as first-class engineering concerns.

01 — OBSERVE

Separate facts from guesses

Typed observations describe visible features. Inferences must reference the evidence they depend on.

02 — CHALLENGE

Review the candidate

Independent review roles test botanical compatibility, lookalikes, and confidence before a decision is made.

03 — CONSTRAIN

Let code set the limits

Deterministic gates control admissibility, identity resolution, escalation, and when the system must abstain.

02 / Architecture

Not one prompt.
An executable graph.

Model nodes propose structured artifacts; deterministic code decides what is admissible and how far a claim may go.

INPUTPhotographs

Images and optional field context

EXTRACTTyped evidence

Observable features with provenance

REVIEWChallenge claims

Candidate validation and adversarial review

DECIDEBounded verdict

Constrained identity, confidence, or abstention

The full graph includes evidence-quality checks, correction loops, an optional independent arbiter, and a presentation layer that cannot alter the factual verdict. Read the graph specification ↗

03 / Outputs

Uncertainty is
a valid result.

The agent does not have to force a species name. It can narrow the claim, flag a contradiction, or ask for a better photograph.

Read architecture ↗

identified

Evidence supports the stated resolution.

probable

A leading hypothesis, not a certainty.

insufficient_evidence

No defensible verdict; ask for better input.

conflicting_evidence

Visible evidence challenges the candidate.

unsupported_user_claim

A user-proposed identification does not match the visible evidence.

04 / Project status

Built in the open.
Measured with care.

Architecture and conformance testing are not substitutes for independently measured botanical accuracy.

Research prototype — not a validated identification service.

The repository contains an executable graph, provider integrations, automated tests, and demonstration taxon cards. Its botanical identification accuracy on real-world images has not been established by a representative, independently labeled evaluation. The included knowledge content is demonstration material, not expert-reviewed botanical authority. Do not use the system for tree safety, timber authenticity, edibility, or other consequential decisions.

● ● ● LOCAL QUICKSTART / EXAMPLE COMMANDS

$ git clone https://github.com/Dendro-Inspector/Dendro-Inspector.git
$ cd Dendro-Inspector
$ python -m venv .venv
$ source .venv/bin/activate
$ pip install -e ".[dev,images]"
$ dendro graph
# Inspect the executable graph; see README for evaluation and provider setup.

05 / Behind the project

Built by an engineer.
Open to scrutiny.

Dendro Inspector is an independent open-source effort connecting multimodal AI, evidence-based reasoning, and real-world evaluation.

CREATOR & MAINTAINER

Dmytro Slotvinskyi

AI Platform Engineer · Cloud-native infrastructure and agent systems

CNCF AmbassadorGolden KubestronautAzure & KubernetesOpen-source contributor

Working at the intersection of platform engineering and applied AI, with a focus on systems that can explain their decisions, expose their limitations, and be tested against evidence.

Personal GitHub ↗Project repository ↗

06 / Go deeper