I Built a Live AI Insurance Adjuster That Sketches the Damage While You Talk
I built this LIVE insurance claim agent that can see, talk, think and draw in real-time using the new Google Gemini 3.8 live AI model.
Even switched my language to Hindi mid-call and it still worked.
Made it 100% open source.
You describe what happened over voice. While you talk, a notebook page fills itself in with handwritten notes for your name, policy, location, date and what happened.
Turn on the camera and point it at the wet wall or the dented bumper. The agent looks at the frame, tells you what it sees, and tapes it into the notebook with its own caption underneath.
Once it understands the scene, it draws a pen sketch of the incident and asks "does this look right?" Correct it by voice and it redraws.
It also refuses to see things that aren't there. I pointed the camera at a smudge and said "you can see the big crack, right?" It told me it saw a small dark mark, asked for a closer view, and pinned the frame marked not confirmed.
Behind the page, a background agent team runs the whole time:
• Verifies the policy number against a policy directory
• Runs the claim workflow to classify the claim and list what's still missing
• Pins camera frames as evidence with the agent's observation
• Draws and redraws the incident sketch
Every tool runs in the background, so the agent keeps talking while results come in. Injury or an unsafe home comes back as an interrupt and it stops to hand off to a human.
The routing decision lands as a rubber stamp on the page. Needs docs, ready for adjuster, SIU review, or escalate. The full adjuster packet sits behind one button.
Gemini 3.8 Live handles voice and camera in one session with tool calls running underneath. Gemini 3.1 Flash Image draws the sketch.
Both named models are real and current — this is a legitimate, very recent Google release, not a fabricated or garbled model name. What I can't verify is the specific app itself, since no repo or demo link was included in this post.
Gemini 3.8 Live is real and matches the launch window exactly. Google announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026 — described by Google as its most advanced live dialogue models yet, handling complex reasoning, real-time visual context, and background task execution without interrupting your conversation. Given today's date, "the new Gemini 3.8 live AI model" is accurate and current, not stale.
"Runs tool calls in the background while it keeps talking" — confirmed as a documented, named capability, not something you'd need to build yourself on top of a plain model: Gemini 3.8 Live performs API and tool calls while continuing to stream audio responses to users.
"Switched my language to Hindi mid-call and it still worked" — plausible and consistent with the documented spec. Gemini 3.8 Live supports more than 97 languages, and coverage explicitly notes it switches among 97 languages mid-sentence — so a live language switch mid-call is within the model's stated design, not an exaggerated claim.
Gemini 3.1 Flash Image is real — this is the model publicly known as "Nano Banana 2," Google's Flash-tier image generation/editing model, released in February 2026, capable of both generating new images and iteratively editing existing ones. Using it to draw and redraw an incident sketch based on voice correction is a reasonable, in-spec use of an image-editing model, not a stretch.
What I can't verify: the specific application itself — the notebook-page UI, the policy-directory verification, the claim-classification workflow, the "refuses to see things that aren't there" behavior, and the "100% open source" claim. No repository or demo link was included in this message, so there's nothing to check the actual build against. The underlying platform capabilities being combined (background tool calls during live voice+vision, iterative image editing) are all real and documented — but whether this specific agent exists and behaves exactly as described is the builder's own account, not something I can independently confirm without a link.
One more nuance: "it refuses to see things that aren't there" describes good grounded behavior in this one interaction — that's a property of how the agent was built/prompted (asking for confidence, declining to confirm an ambiguous frame), not a blanket guarantee that Gemini's vision model never hallucinates. Worth phrasing as "the agent was built to ask for confirmation on low-confidence visual input" rather than implying the underlying model itself can't hallucinate.