Opportunities › Root cause analysis for LLM application failures › Proposed app
Threadpull
Find the root causes behind your AI agent's silent failures
- Score
- 35
- confidence 61
- Evidence
- 6
- items on the page
- Rivals
- 1
- named in sources
- First version
- 40 d
- estimate
- Type
- Desktop app
- PWA: runs in the browser, installs on Mac and Windows
- Market since
- 2023
- Emerging market · AI estimate
ScreensConcept
Desktop app: Engineers analyze failure clusters, review dense root-cause data, and examine session evidence for debugging and release decisions—sustained desk work. Sample screens on a computer, in the proposed brand. Their numbers are sample data.
Root causes
Last 7 days5 causes
About this appAI proposal
Threadpull finds out why your AI agent gives wrong answers. It collects signals from each session: thumbs-down, user edits, abandoned clicks, tone of replies and LLM-judge scores. It turns them into short facts about what went wrong.
It then groups similar failure hypotheses across hundreds or thousands of sessions. Instead of reading traces one by one, you see a short list of causes, each with example sessions and a suggested fix to review.
Built for engineers and teams running agents in production with high session volumes, where failures look random until they are grouped.
- Tagline
- Stop scrolling traces. See why your agent is wrong.
- Voice
- PlainPreciseCalm
- Why this name
- A thread is a session, and also the thing you pull to unravel a problem. Pull one and you see the cause behind many failures.
In the first version
Members say which features they would use. It tells PWAS what people want; it changes no score.
Signal collection
Ingest feedback, edits, clicks, sentiment and LLM-judge scores per session through a simple SDK or HTTP API.
No member yet
Session facts
Each flagged session is boiled down to a few plain facts: what the user wanted, what the agent did, where it went wrong.
No member yet
Failure clusters
Similar failure hypotheses are grouped across sessions, so scattered errors show up as a handful of patterns.
No member yet
Root cause cards
Each pattern gets a cause, the share of sessions affected, and example sessions you can open.
No member yet
Brand identityAI proposal
Check the name against trademark registers before using it.
Palette
#14231FDeep Ink · text#F6F3ECPaper · surface#1F5C4DSpruce · primary#E0902BSignal Amber · accent#B8432FBrick · data: failures#D9E2DCMist · secondary surface
Typography
Aa
Threadpull
PricingAI proposal
Kelet's pricing page (verified) lists a free plan, one paid plan at $400 a month, and prices on request for larger customers. These tiers follow that shape, with Team set below it, as a starting point to test.
Free
Free
- Up to 2,000 sessions analysed a month
- 1 project
- Root cause cards
Team
$299 a month
- Up to 50,000 sessions analysed a month
- Unlimited projects and seats
- Suggested fixes
- Release drift view
Scale
Contact us
- Millions of sessions a month
- Volume sampling controls
- Priority support
Would you pay Team · $299 a month?
Your answer tells PWAS which prices people accept. It changes no score.
No answers yet.
What it could be worthEstimate
What a software business making this much sells for today, from real deals. Pick a monthly revenue.
If it makes
Worth about
$190,000
Sale range $132,000 to $263,000
$120,000 a year in revenue
From 87 software businesses listed and sold on Empire Flippers in the last year, revenue vetted by the broker: they sell for about 19× their monthly revenue, most between 13× and 26×.
Measured on profit, the middle deal is 32× monthly net profit: about 2.7 years of profit.
An estimate, not an offer or financial advice. A real price depends on profit, growth, churn, how much of the work is the owner's, and the buyer.
How it would growAI proposal
How the product would reach new customers through its own use, and how the first ones would be found. A plan to test, not a result.
- Built-in loop
- No loop in the product. It reaches customers through content, developer communities and direct sales. Root cause reports could be shared inside a team, but they are internal and do not bring in outside users.
- Do they become customers?
- Test whether engineers who see a shared cause card become trial users. Until then, count only paid signups from posts, docs and outreach.
- First customers
- Engineering leads at small AI startups running a customer-facing support or sales agent with over 10,000 sessions a month and no tool for finding failure patterns.
- Where to reach them
- Hacker News and dev forums: Post worked examples of clustering failed sessions and invite readers to try the free tier on their own data.
- Direct outreach: Message engineers who write publicly about agent failures and offer a free analysis of one week of their sessions.
- First milestone
- Five teams in one vertical each upload a week of sessions, and at least two pay for Team after the trial.
Why people already pay for this
- VerifiedHacker News post received 47 points and 24 comments
- Public claimAI agents running in production can serve millions of sessions per day and fail silently by giving wrong answers rather than crashing.
- Public claimDebugging AI agent failures today requires manual investigation across hundreds of sessions.
- Public claimIndividual session failures appear random, but clustering failure hypotheses reveals patterns.
- Public claim
Tech stack
Every proposal builds on the same PWA base, installable from the browser on any phone or computer. Dashed boxes are what this app adds.
- AppNext.js and TypeScript
- PWAManifest and service worker
- UITailwind CSS
- Data and authSupabase Postgres
- PaymentsStripe
- EmailResend
- HostingVercel
- This app · Background jobsQueue and worker serviceSession analysis and clustering run asynchronously on large batches.
- This app · Vector searchpgvector in Supabase PostgresClusters similar failure hypotheses by embedding similarity.
RisksAI inference
- CommercialKelet already offers this with SDKs and a free beta. We must win on clearer causes or lower price, and teams may build it in-house.
- AIClusters or causes may be wrong or vague. Teams will drop it if suggested fixes do not improve results, so evidence must be shown beside each cause.
- DataSessions hold user conversations and may include personal data. Customers will need redaction, retention limits and clear handling terms.
- TechnicalAt high volume, analysing every session with an LLM is costly. Sampling must stay accurate while keeping unit costs below price.
- PlatformTeams use many tracing tools and formats. Without easy imports, setup effort may block trials.
Everything marked AI proposal or AI inference was written by a model from the evidence on the opportunity page. The proposed name must be checked against trademark registers before it is used.