• 

  Debate replay page with seek / scrub controls and chat logs.

• GIF

  Parallax logo GIF

• 

  Continuous deployment over private Tailscale VPN

• 

  Landing page

• 

  Inbox page

• 

  Outreach detail page

Parallax Agents Society

Inspiration

• Professors are drowning in AI-generated cold emails that look personalized but aren't. No tool helps them see through the noise from their side of the inbox.
  
• Every admissions tool serves the applicant. Parallax flips the vantage point: the professor's own publications and declared capacity become the yardstick every candidate is measured against.

What it does

• Parallax is a reactive triage tool for professors: inbound student outreach email is debated by a society of AI agents against the professor's own publications and lab capacity, returning a grounded decision with references, which I refer to as receipts.
  
• A cheap Gatekeeper pass (qwen-turbo) triages every email first, extracting a structured candidate profile and verifiable claims from the body and attached CV, and rejecting mass-mail spam before it costs a debate. Rejections land in a reviewable "Declined" queue, never a silent drop.
  
• Survivors go to a moderator-driven debate. The Gatekeeper opens by explaining why it let the candidate through; a Research-Fit Advocate, an Authenticity Auditor, and a Capacity & Funding Assessor then take sequential turns over one shared transcript, each rebutting or conceding what was just said, while a Moderator routes the floor. An Arbitrator judges the full transcript once at the end and drafts the reply.
  
• The debaters have names and personas and address each other and the professor by first name, so the transcript reads like a real admissions panel rather than role labels reciting findings:

• Free-text custom instructions set at onboarding steer the society (e.g., "I want only theory students, no pure-applied ML"). The professor reviews the decision and approves, overrides, or edits the drafted reply no outbound email sends without explicit human approval.
  
• Every debate is persisted as a replayable trace: turns, receipts, cross-references between agents, and every tool call each agent made.
  
• The replay plays with voice. After a debate, each turn gets a short spoken line, synthesized with a distinct Qwen3-TTS Flash voice per persona so the professor can watch and hear the society deliberate instead of reading a transcript.

How we built it

• The debate design is grounded in two papers, one of which is indexed in the demo corpus: ChatEval (Chan et al., 2024) a panel of debating LLMs beats any single LLM judge and Liang et al. (2024), which introduced the judge-driven Multi-Agent-Debate (MAD) framework and named the Degeneration-of-Thought (DoT) failure mode. Debaters argue the same candidate, rebutting or conceding; an Arbitrator judges once at the end. Never a supervisor splitting sub-tasks.
  
• The engine is a LangGraph StateGraph running MAD's judge-driven debate: a Moderator (qwen3.6-plus) routes the floor to whichever debater can best pressure-test the last turn, while debaters take sequential turns over one append-only transcript. Any debater may PASS, the Moderator can only end once every voice has been heard, and a hard turn cap guarantees termination of the structural defense against DoT.
  
• Qwen Cloud, end to end one API key, no non-Qwen model anywhere:

• Custom Agent Skills follow the open SKILL.md standard with progressive disclosure: claim-verification and capacity-math ship as skill folders whose full procedure loads only when an agent invokes them.
  
• Agents ground themselves through typed tools, each recorded on the turn as an AgentAction so the replay shows how each agent grounded itself: RAG over the corpus (RETRIEVAL), the two skills (SKILL), and a live student-visa/mobility lookup over an in-process MCP tool bus (MCP).
  
• The backend is hexagonal (ports & adapters): the debate contract lives in the pure domain layer; LangGraph and every Qwen call live in adapters. A compliance guard refuses to construct a client pointed at any non-Qwen-Cloud host, enforced by a test that greps the codebase for stray endpoints.
  
• Debate audio is generated off the critical path: a post-debate step reads the saved trace, compresses each turn into a spoken line, and synthesizes it so correctness never depends on TTS, and a synthesis failure degrades to a silent beat rather than breaking the run. The replay stays a deterministic function of one playhead clock, with the audio element slaved to it so scrub/step/speed keep working.

Challenges we ran into

• Keeping the domain layer framework-free meant a hard line: the NegotiationEngine contract depends only on domain models; LangGraph, the tool loop, and prompt rendering live behind it in an adapter.
  
• Naming mattered: a debate society (everyone argues the same question, a judge resolves) kept collapsing into task collaboration (a supervisor splitting sub-tasks). We wrote the distinction into the engine's contract.

Accomplishments that we're proud of

• Every agent claim carries a receipt traced to the professor's actual indexed publications the system cannot hallucinate alignment, only surface or refute it.
  
• Every mid-argument tool invocation is captured on its turn (kind, name, arguments, source), so the replay shows how each agent grounded itself, not just what it said.
  
• A rejected email is never silently dropped: triage rejections synthesize an overridable decline the professor can reverse in one click.
  
• We measured the society against a strong single-agent control same Qwen model, same RAG tool over the same corpus, explicit chain-of-thought, an equal per-call token budget, and the society at real production depth on 19 hand-labeled cases spanning fabricated citations, inflated alignment, capacity mismatches, and adversarial contradictions:

The "Efficiency Gain" of the Society: While the Society uses more tokens, it solves the "Trust Threshold" problem. An 84% accurate baseline requires the Professor to manually read every inbound email to find errors. By implementing Sequential Rebuttals and Conflict Resolution, Parallax achieved a 50% reduction in catastrophic False-Accepts (cases where a student with fabricated credentials would have been invited).

In a high-stakes academic environment, the "Efficiency" of Parallax is measured in the hours of human auditing saved, as the Society handles the "Critical Pressure Testing" that a single prompt cannot reliably sustain.

What we learned

• Multi-agent debate needs a hard termination condition from day one. Without a round cap and a decisive arbitrator, the Degeneration-of-Thought failure mode is a real production risk, not a theoretical one.
  
• We assumed cadence could stay fixed (simultaneous rounds) with emergence living in the argument content, but debaters speaking simultaneously are blind to each other within a round and never actually respond. Real rebuttal required a moderator-routed sequential debate where each agent sees the turn it is answering.
  
• Structured output and free-form tool-calling pull in opposite directions. Splitting them, tool-loop for debaters, structured output for Gatekeeper and Arbitrator verdicts kept both reliable.

What's next for Parallax

• Integrating with email providers
  
• Qwen-VL over the figures and diagrams in the professor's own papers, so the corpus includes what the text alone misses.
  
• Potentially partnering with professors from multiple universities.

Built With

• alibaba-cloud-ecs
• celery
• cloudflare-r2
• fastapi
• next.js
• postgresql
• python
• qwen
• redis
• tailscale