The ConglomerAIte™

Inspiration

Cloud agents often fail at exactly the wrong moment: a field device loses connectivity, latency spikes, or a long refinement loop exhausts edge memory. The ConglomerAIte™ treats those failures as first-class operating conditions. It combines QwenCloud reasoning with a bounded Jetson fallback so an agent can return the best safe result available instead of hanging, retrying forever, or crashing the edge node.

What it does

ConglomerAIte is an edge-native Generator/Critic system built primarily for Track 5: EdgeAgent, with Track 3: Agent Society mechanics. A Generator produces a candidate, an independent Critic returns a machine-parseable score and targeted corrections, and the system revises until a structurally valid 10/10 has no blocking issues. The loop also stops honestly on iteration, time, stagnation, provider, or hardware-safety boundaries and returns the best candidate seen.

When cloud use is allowed, the router targets QwenCloud's OpenAI-compatible endpoint. Retryable connectivity failures open a bounded circuit breaker and move subsequent work to the Jetson's loopback-only SYSOP Qwen Bridge. On edge, Generator and Critic are logical roles sharing one quantized Qwen-family model sequentially, avoiding duplicate weights and KV caches on the 8 GB Orin Nano. Metadata-only traces record routing, latency, tokens, scores, memory observations, and stop reasons without storing prompts or model output in the journal.

How it was built

The dependency-light Python core contains provider adapters, retry classification, a three-state failover router, strict Critic-schema validation, best-candidate retention, score-stagnation detection, and RAM guards using tegrastats, /proc/meminfo, finite cgroup limits, and optional NVML away from Jetson. Systemd units isolate the orchestrator and model service, while a staged SSH/rsync deployment supports atomic activation and rollback. An HMAC-verified Alibaba Function Compute receiver accepts only redacted scalar telemetry.

Evidence and results

One paid QwenCloud development run demonstrated an 8/10 first draft reaching structurally valid 10/10 consensus in two rounds. A three-task pilot observed lower latency and fewer output tokens than a budget-matched single-agent self-refiner on quality-noninferior cases, but it did not establish a general quality advantage: one task regressed by one evaluator point and the evaluator was not independent. On the deployed Jetson, a zero-paid forced-loss task made three route attempts, degraded twice to the local SYSOP bridge, retained its best 8/10 candidate, and stopped at the 125.7-second safety boundary with both services healthy and zero model restarts. The live Alibaba Function Compute receiver separately returned 200 for health, 401 for unsigned metadata, and 202 for one signed metadata-only event. Raw, sanitized summaries and claim limits are committed under evidence/.

Challenges and lessons

Jetson CPU and GPU share unified memory, so NVMe swap is not extra GPU memory and MemoryMax is not a CUDA reservation. This led to a layered safety design: application-level admission checks, one-slot inference, bounded context/output, systemd containment, and explicit terminal results. Live testing also exposed two correctness problems—rubric drift and hidden output truncation—which were fixed by propagating the immutable rubric on every round and refusing truncated generations as consensus candidates.

What's next

The production roadmap adds a durable round ledger and reboot resume, thermal/swap-growth guards, adaptive context shrinking, independent/blinded evaluation, and signed remote policy. The public evidence package now includes a 2:55 Andrew neural-voice demo with 14 seconds of physical Jetson network-link-loss footage and the verified GitHub release below. Cross-platform credential-free CI runs the complete offline validation on Windows and Ubuntu.

Public demo and evidence

Review the 2:55 Andrew neural-voice demo with physical Jetson footage, architecture diagram, and judge-safe Alibaba screenshots at https://github.com/Destr0yering/conglomeraite/releases/tag/hackathon-mvp-v1

Built With

• alibaba-cloud-function-compute
• hmac-sha256
• llama.cpp
• nvidia-jetson-orin-nano
• python
• qwen-plus
• qwencloud
• serverless-devs
• systemd
• tegrastats