Samdesk Demo by Aryan!
LIVE

SamdeskDemo

An autonomous agent that sifts through noisy social chatter and works out what's actually going on in a crisis, with the sources to back it up, all in a few seconds.

Just a heads up, this is a demo. I wanted to build something like this for my own interest, so it's super simplified and put together over a weekend.

GitHubLinkedInby Aryan Kumar

How it works

Why it matters

Speed

The agent investigates and decides in seconds. Watch it work in the live demo. Tool calls happen in real time: corroboration search, source reliability check, contradiction detection, asset exposure assessment.

Trust

Every claim in the brief is tied to a retrieved signal ID. The grounding guard blocks publication of any unverifiable claim. When sources conflict, the agent escalates rather than guessing.

Measured

Clustering precision/recall, classification F1 per event type, agent decision accuracy, false-verify rate, mean tool calls, grounding faithfulness, and p50/p95 latency, all computed over the labeled corpus.

Live Demo

Real Anthropic tool-calling running server-side. Select a scenario and watch the agent investigate.

01
Choose a scenario
Pick from three pre-loaded incidents on the left: a verified fire, planted misinformation, or an ambiguous flood.
02
Run Pipeline, then Investigate
Click Run pipeline first to cluster and classify signals, then click Investigate to launch the live agent.
03
Reset before switching
Hit Reset before selecting a different scenario, otherwise the previous run state carries over.
04
Run Evaluation after escalation
The eval rail scores decision accuracy, false-verify rate, and grounding. It runs 3 live agent calls so it costs a few cents and takes ~30s.
1 · Ingest
2 · Cluster
3 · Classify
4 · Investigate
5 · Decide
Choose Scenario
Signal stream · preview
@porthalworth_fd14:03:11ZEN

Structure fire confirmed at Docklands Warehouse 7. Hazmat team en route. Shelter in place advisory issued for 500m radius.

AQI_SENSOR_PH_0414:04:22ZEN

PM2.5 spike detected: 284 µg/m³. CO elevated at 38 ppm. Sensor reading consistent with combustion event.

Reuters_Wire14:07:55ZEN

FLASH: Fire breaks out at chemical storage facility in Port Halworth docklands district. Emergency services on scene.

@dockwatch_intel14:09:30ZEN

Confirmed visible smoke column from Docklands Quay area. Multiple independent eyewitness accounts. Fire visible from 3km.

Select a scenario and run the pipeline

The agent will investigate in real-time

In Production

A proposed reference architecture, illustrative integration grounded in the public role description. Not the company's actual internal systems.

I wanted to build this out to see how something like this would look in production, since I found it super interesting.

IngestionIntelligenceAgentHuman-in-LoopDeliveryPlatform

What changes from demo → prod

ConcernIn this demoIn production
Data sourceSynthetic fixturesLive streaming feeds (Kafka/SQS)
CorroborationToken/entity similarityEmbeddings + vector DB (RAG)
ClassificationFeature lexicon scorerFine-tuned model on analyst-verified data (GPU)
Agent toolsRead in-repo fixture dataCall internal services and databases
DeliveryDemo UI onlySlack/Teams/Everbridge/Esri/API
EvaluationOn-demandContinuous eval + drift monitoring + model-version gating
ModelsSingle modelProvider abstraction + model routing (cost vs frontier)

Model Strategy

Provider-agnostic abstraction over Anthropic and OpenAI. Route triage work to a cheaper model, high-stakes briefs to a frontier model. Every model upgrade is gated behind the eval harness: each new version must measurably beat the last on decision accuracy and false-verify rate. Drift monitoring and latency dashboards watch production continuously.