HackStack

Project Info

Quad

Devpost

Inspiration

Every company knows more than its public surface says. The team remembers what got decided in a meeting; the website still shows last year's pricing. Support answers a question one way; the docs say another. AI search engines then read the stale version and repeat it with total confidence. We kept hitting the same failure. AI tools are great at generating answers and terrible at proving them. A chatbot will happily tell a customer something that contradicts your own internal truth, with zero evidence and zero paper trail. For anything customer-facing, "sounds right" isn't good enough. So we built Quad: an AI employee that proves its work. Not a chatbot. An employee with a company brain, real browser-grounded evidence, live execution, and an approval-first contract.

What it does

Quad connects to your systems of record and continuously audits them. It finds conflicting claims, stale facts, and broken links across every tool your company runs on, and traces every answer back to the exact source document that produced it. Every fix is drafted in your brand voice and queued for one-click approval before anything ships. Every AI-generated conversation gets evaluated for quality in production. Quad also joins meetings. It listens live, in real time, and turns conversation into structured memory the moment it's said. Ask ASI:One "what did we decide on the Q3 budget" and Quad retrieves it, sourced back to the exact meeting and moment. Quad takes action. It operates a real browser, executes the fix, and orchestrates other agents through Agentverse to get the work done. Quad gets better from doing the work. Every meeting it sits in, every gap it finds and fixes, every correction a human makes, trains the system that runs the next one. The system running next month answers better than the system running today. Every claim Quad makes carries proof. This is Quad Chain. When Quad says your refund window is 30 days, it shows the exact source. When it compresses a long meeting trace into a short memory, it proves what was kept and what was dropped, so an agent or a human downstream can verify the summary before acting on it. Quad traces a claim all the way back to the moment it was true, a document or a sentence someone said in a call.

How we built it

Connect: integrates with the systems of record knowledge actually lives in. Audit: a continuous pass across every connected source finds conflicts, stale facts, and gaps. Live meeting capture: Quad joins meetings and listens in real time, converting conversation into structured, sourced memory as it happens. Memory: a persistent, embeddings-backed company brain, queryable retroactively through ASI:One, accumulating from documents and conversations. Quad Chain: verifiable receipts on every memory and every claim. Compression with proof of what was preserved. Tamper-evident, anchorable on-chain, private data stays off-chain. Learning loop: outcomes from real meetings and real corrections feed back into the system, so accuracy and usefulness improve with usage. Trace: every answer resolves back to its source, with drafted fixes in brand voice queued for approval. Action: real browser execution via Browserbase, real multi-agent orchestration via Agentverse, registered and callable as a live agent through the Fetch.ai Agent Chat Protocol and ASI:One. Reasoning: a multi-call Anthropic pipeline does analysis, drafting, independent verification of each draft against its evidence, and synthesis. Observability: Arize traces and evals on every cognitive step, scoring groundedness and hallucination risk before anything ships. Reliability: Sentry-instrumented end to end so a failed step degrades instead of taking down the run. Retrieval: OpenAI embeddings power recall across the memory graph. Quadchain Modern agent systems do not just need shorter prompts; they need compressed memory that can be routed, audited, repaired, and rejected when declared obligations are missing. We introduce QuadChain, an obligation-verifiable context compression architecture for multiagent LLM systems. QuadChain couples extractive compression with explicit evidence obligations, answer-concept checks, role-aware routing, omission manifests, handoff integrity metadata, and verified selective rehydration. The output is not a proof of semantic faithfulness; it is a verifiable memory packet that tells downstream agents which declared facts were preserved, which spans were omitted, what was repaired, and whether the handoff should be accepted. In controlled coding-agent fixtures and public-benchmark-style local adapters, the measured 4-agent workflow drops from 9,000 raw tokens to 2,283 routed tokens (74.63% reduction). Verified selective rehydration reaches 0.9390 deterministic task score with 88.89% mean token reduction and 210/240 accepted packets under matched budgets. These results do not establish state-of-the-art prompt compression; they support a narrower systems claim: compressed agent memory should be accountable, rejectable, and selectively repairable rather than blindly summarized.

Challenges we ran into

Reinforcement learning from real conversations had to improve answers, not just add to the model indiscriminately. We gated the loop so only verified outcomes feed it. Live meeting capture had to stay reliable enough to demo. We built a fallback path that keeps the session running if a piece of the pipeline hiccups. The action layer and the knowledge layer had to work as one system. We built them on the same backend from the start. Accomplishments we're proud of A knowledge platform that connects, audits, traces, drafts, and evaluates, and also captures knowledge that never existed in a document. A verifiable memory protocol, Quad Chain, that's infrastructure-grade. Real agent orchestration through Agentverse with real execution. An agent that's genuinely discoverable and callable through ASI:One. A reinforcement loop where the system improves with usage.

What we learned

Auditing documents is the easy part of this problem. The hard part is the knowledge that was never written down, the meeting where the real decision got made, and proving that what an agent remembers is still true by the time it acts on it.

What's next

Extend live capture into every channel knowledge moves through. Deepen the learning loop so improvement compounds faster with usage. Open Quad Chain as the verifiable memory layer other agents ground themselves against.

Analysis

Compare with all teams

View

Metric

Figures cover GitHub contributors during the hackathon window. A co-authored commit counts in full for each author, so per-member totals add up to more than the whole-team figures.

Technology

Found in codeClaimed only
  • AnthropicIn code
  • CSSIn code
  • HTMLIn code
  • JavaScriptIn code
  • Next.jsIn code
  • PostgreSQLIn code
  • PythonIn code
  • ReactIn code
  • SQLIn code
  • SupabaseIn code
  • Tailwind CSSIn code
  • TypeScriptIn code
  • RedisClaimed

12 of 13 appear in the indexed code. 1 claimed on Devpost could not be matched to code, which may simply mean the tool leaves no trace in the repository.

AI coding agents

  • Claude CodeConfig · Commits
  • CodexConfig

Detected from committed agent config files and commit authorship. Absence of a signal is not proof an agent was unused.

Codebase size

Source size

2.2 MB

Source files

395

Counts recognized source files only; vendored directories, binaries and lockfiles are excluded, so this is smaller than the repository on disk.

0 stars