Agentic AI · Python · LangGraph
Live

Noiseless

AI agent that finds high-signal conversations on X and learns from your feedback to surface better results over time.

Slack — Direct MessagesDIRECT MESSAGESNNoiseless3SlackbotTeam UpdatesCOMMANDS/discover/stats/run-statsNNoiselessNNoiseless12:30 PMDiscovery Digest — Jul 28, 2026🧭 Based on your profile, we searched for posts about AI agents,distributed systems, and production engineering...Found 8 items from 12 queries. Review each item in the thread below.Run #47 | Jul 28 12:30 PM | /discover💬 8 repliesThread8 repliesNNoiseless12:30 PM@swyx(Topic Discussion) | Score:8/10Building LangGraph agents in production is tricky.Here's what I learned about state management...💡 Topic Discussion | Jul 28 · 3h agoLikeDislikeSkipNNoiseless12:30 PM@karpathy(Expertise Match) | Score:9/10Cost control in autonomous agents is underrated.Per-request budgets prevent runaway spending...🔬 Expertise Match | Jul 28 · 5h ago✓ LikedNNoiseless12:30 PM@dan_abramov(Topic Discussion) | Score:6/10React Server Components mental model for...💡 Topic Discussion | Jul 27 · 1d agoLikeDislikeSkip
01 — Overview

Finds signal in the noise. Gets smarter every time you use it.

Most discovery tools either flood you with noise or require hours of manual searching. Noiseless sits in between. It profiles your expertise from multiple sources (your posts, your code, websites you provide), searches X daily for conversations where your perspective adds value, and delivers ranked results to Slack. You like, dislike, or skip each result, and that feedback shapes what surfaces next.

The system runs as a scheduled daily pipeline: search, classify, rank, deliver. Each user gets an isolated profile, budget, and learning history. After delivery, feedback flows back asynchronously through Slack as users react to results over hours. The learning loop was the architectural starting point, not something grafted on after the fact.

The agent gets better the more you use it. Every like, dislike, and skip feeds into preference profiles that reshape future ranking. Query strategies shift toward topics producing higher-signal results. The profile itself is built from cross-referenced evidence across sources, not self-reported claims.

02 — The Hard Parts
01

Scheduling Discovery Across Users

The pipeline runs on a daily schedule per user, across timezones. Each run executes search, classification, ranking, and delivery autonomously, then feedback trickles back through Slack over hours as users react to results. State survives process restarts. The pipeline owns scheduling and orchestration, Slack owns the feedback interface, and the database is the only shared state between them. Getting that boundary wrong means lost feedback or results that never learn from user behavior.

02

Spending Real Money on Every Request

The agent makes API calls that cost real money: searches to find posts and LLM calls to classify, rank, and generate digests. Without hard limits, a misconfigured run could burn through a budget in minutes. Cost enforcement runs before any work begins. Per-user daily and monthly limits are checked independently. Every external action gets a durable audit entry before the call is made, so if the process crashes mid-request, there is always a record of what was attempted. This is the only responsible way to run a system that spends money on behalf of other people.

03

Building Profiles from Noisy Sources

The system needs to know what you are actually an expert in to decide which conversations are worth surfacing. But the inputs are noisy: social posts mix signal with performance, repositories vary in relevance, and self-reported interests are unreliable on their own. Profile building merges signals from multiple sources with confidence scoring: your posts carry more weight for interests, your code carries more weight for technical depth. Overlapping mentions across sources reinforce each other; isolated claims get discounted. If a topic falls outside your demonstrated expertise, the system skips it entirely rather than surfacing noise.

03 — Outcomes
Minutes

Spent approving, instead of hours hunting for conversations worth joining.

Learns

Every like, dislike, and skip makes the next digest sharper than the last.

Capped

Daily and monthly spend limits per user, with cost visible the whole way.

Zero

Config to write. A Slack conversation builds the profile, no forms or keyword lists.

04 — Stack
PythonFastAPILangGraphClaudeOpenAI EmbeddingsPostgreSQLSupabaseSlack Socket ModeViteRailway
05 — What's Next

An agent that compounds.

Richer feedback dimensions: Move beyond like and dislike to capture why a conversation was relevant or not, enabling more precise preference modeling over time.
Profile evolution: Expertise profiles that adapt as your interests and skills naturally shift, instead of staying locked to the initial onboarding snapshot.
Broader discovery: Expanding beyond X to find relevant conversations wherever your audience participates.

Try it, follow the project, or reach out to discuss the engineering.

© 2026 Tanish Nahata