Pulse

Inspiration

We love building in public. Kush does growth at a YC-backed startup and has seen Large Language Model Optimization (LLMO) quietly become one of the most underrated marketing levers the industry doesn’t build for. When someone asks ChatGPT or Perplexity to recommend a product, the product that shows up, or gets recommended as part of an API call, receives an undeniable market edge – this is the new SEO. Many of us also understand the social side from experience, growing brands on Instagram and learning firsthand how frustrating it is to post content with no real signal on what consumers actually think or react across the internet.

Our core belief behind Pulse lies in the idea that the consumers of today live online. Social media is the new environment where attention thrives, and Large Language Models (LLMs) are where future attention is headed. The difficulty of customer relationship management has never been about the lack of dedicated resources, but instead the difficulty of acting on feedback fast enough to satisfy customers. Agents fix that. Autonomous AI agents are what turn a monitoring dashboard into a teammate that scrapes the open web, audits every major LLM, classifies thousands of signals, detects anomalies, and executes responses, all in a loop that never stops. What used to require a full team now runs every 25 seconds to address any and all problems, surfing Reddit, X, Youtube comments, Instagram, and more to maintain the perfect company brand.

Agents are only as good as the infrastructure underneath them, and we made heavy use of the important tools sponsoring this hackathon. Nimble gave us real-time web data at agent speed with zero scraping infrastructure to maintain. Senso gave us something that has never existed before: a continuous structured audit of how every major LLM actually describes your brand against verified ground truth. ClickHouse gave us a database fast enough to query anomalies in single-digit milliseconds on every loop iteration. Grok gave us the model speed and token efficiency that made running this continuously economically viable at all. Without any one of them, Pulse does not work.

To restore a brand's ownership of its own narrative across both social and AI, we built Pulse. What Pulse Does Pulse is an autonomous brand intelligence platform that saves companies hours of manual monitoring while catching reputation risks far ahead of time, across both the open web and major AI models trusted by the general population.

Pulse continuously tracks brand mentions across social platforms and the open web using Nimble, which returns structured, accurate data that works resiliently for sites throughout the web. Using parallel agents, it probes ChatGPT, Claude, Gemini, Grok, and other major LLMs on a set schedule, asking them about company products and catching any hallucinations, outdated facts, and competitor-favoring narratives as soon as they appear in model answers.

As mentions arrive, Pulse generates embeddings with Groq and uses pgvector to sort semantically similar feedback into classified themes. Pulse uses a five-factor severity score: Volume (post count × 10) + Engagement (likes × 2 + shares × 5 + comments × 3) + Sentiment (negative percentage × 3.5) + Velocity (past 2-hour posts/total posts) all multiplied by Influence. The score weights issues with real business impact and classifies each problem using Groq, producing a severity metric that encompasses the core complaint, the customer base affected, and how widespread the problem is.

Clickhouse processes massive streams of data analytics and outputs sentiment trends in milliseconds even processing hundreds of thousands of posts per day.

When a cluster crosses the threshold, Pulse decides the action to take and drafts it: anything from a customer reply written in the brand’s voice to send through direct message or platform comments, a structured support ticket with steps and affected user estimates, an option to escalate problems to higher levels of management, an update to the FAQ, or a correction for hallucinating models. Drafts use Groq while never making unauthorized promises, and every action has a risk rating — low-risk responses fire automatically, while sensitive responses require human approval with one-click buttons for approval, editorial, or rejection. Nothing posts without clearing human approval, which is made as simple as possible.

The entire pipeline implements Datadog tracing from end-to-end, with an average target of < 2 minutes from draft to post. Pulse executes approved actions and tracks the outcome, feeding it back into a scoring model so in the future, it learns to prioritize issues better and better. Pulse is the solution for brand risk that never sleeps.

How We Built It Pulse runs a six-step agent loop every 25 seconds: scrape, analyze, store, score, decide, act. Three parallel scrape calls fire on every iteration via Promise.allSettled so no single failure kills the loop. Nimble pulls fresh open-web mentions across Reddit, HackerNews, and SERP with rotating queries. Senso audits one rotating LLM per loop so all four major models get freshly scored every hundred seconds. A health ping checks brand uptime in parallel.

Every mention gets batched into a single classification call rather than one per mention, returning sentiment scores, topic labels, confidence ratings, and a disambiguation filter that drops anything not genuinely about the brand. Results are written into two ClickHouse tables, one for open-web signals and one for LLM audit snapshots, both ordered by brand and timestamp so anomaly queries return in milliseconds at any scale.

The policy evaluator is the heart of the autonomy story and it contains zero LLM logic. It is pure deterministic code that checks hard gates: confidence threshold, sentiment delta, risk keywords, source whitelist, blocked terms. Groq drafts the response. The policy decides whether it fires. If a mention contains "lawsuit" or "breach," the agent escalates to human review regardless of everything else. Every gate check is logged, giving a complete audit trail for every decision the system has ever made. Autopilot actions post to Slack with full gate results attached. GEO corrections push back into Senso so LLMs ingest accurate ground truth on their next crawl. Anything blocked lands in a human approval queue. Challenges Our first architecture passed a large JSON context file into every model call. It worked until the token costs made continuous operation impossible. We pivoted to a leaner TOML-based format and redesigned batching to send all mentions in a single call per iteration, cutting cost and latency by an order of magnitude. We also replaced setInterval with chained setTimeout so the loop never overlaps itself when model calls run slow. Disambiguation was a constant tuning challenge since brand names are rarely unique, and getting false positives low enough to trust autonomous execution required aggressive confidence floors and brand-specific context injection on every classification call. Tech Stack Next.js, TypeScript, FastAPI, Python, Nimble, Senso, ClickHouse, Grok, Gemini, VADER, Slack Webhooks. What's Next LLMO is the defining marketing discipline of the next decade and almost no company has visibility into it yet. Pulse is the infrastructure layer that changes that. Multi-brand support, Twitter and LinkedIn monitoring, and a self-improving ground truth pipeline are next. The long-term vision is simple: every company deserves an agent that knows exactly how the world sees them and has the autonomy to fix it before the damage is done.

Our RAM was messed up so we couldn't upload a video demonstrating the reddit or X post, and we decided to live demo this instead.