Recommended Free Tools
The pipeline worked at the data level and failed at the business level. According to its builder, Vivian Oliveres, the Dwelverson product scanned about 1.75 million posts a day across Reddit, Bluesky, X and Hacker News for people actively looking for a tool, then passed potential leads to vendors. Inference ran at roughly €10 a day. The author reports one trial from 800 cold emails, and the project reportedly shut down on September 15 after development that began in February. In the author’s own retrospective, the core mistake was sequencing: the full system was built before anyone confirmed that a buyer wanted the output.
Every figure below comes from Vivian Oliveres’s own project post and post-mortem. They are one founder’s measurements, not independent audits, and they should not be read as typical conversion rates, costs or model benchmarks.
As an Amazon Associate I earn from qualifying purchases.
What Dwelverson was meant to do
Dwelverson was a B2B lead-intent product. Its premise was that people who post about a problem on social platforms, such as asking for a recommendation or describing a workflow they want to fix, are close to a purchase decision. The system watched those posts, classified them for buying intent, and sent the likely leads to software vendors who could help.
Free tools Windows power users keep installed
One-click scans. No signup required.
The product was built solo. The author describes the four sources as Reddit, Bluesky, X and Hacker News, and reports a volume of about 1.75 million posts per day. The post does not state the exact publication date, and the post-mortem’s date is also not established in the available material, so the timeline of the build and shutdown should be read as the author’s own account.
#1 Best Overall
The reported figures
The author published the following numbers. Each one is attributed to Vivian Oliveres and carries the qualification noted beside it.
| Figure | Reported value | Qualification |
|---|---|---|
| Posts processed | About 1.75 million per day | Author-reported in the project post; the post’s exact date is not established. |
| Inference spending | About €10 per day, down from an earlier estimate of about €100,000 per month | Author-reported in the post and post-mortem; publication date not established. |
| Trial yield | One trial from 800 cold emails | Author-reported in the project post; a single outreach campaign, not a benchmark for cold email. |
| Reddit share | 0.2% of collected posts, but about 80% of true leads | Author-reported; “true leads” is the author’s own label. |
| F1 ceiling | 0.27 for the task | Author-reported in the post-mortem; the author attributed the low ceiling to task difficulty. |
| Inference cost after distillation | About 30 times lower, with the same recall | Author-reported in the post-mortem; the comparison is the author’s own, not an outside test. |
How the pipeline was built
The technical design is where the project is most instructive, because it shows how much engineering effort went into a system that still could not sell.
Rank #2
A 14-step cascade
The author describes a 14-step cascade that filtered posts in stages. Cheap methods ran first, including regular expressions and embeddings, followed by smaller encoder models, then a local LLM and finally an API-hosted LLM. The idea was that each stage handled what it could and passed only the uncertain cases onward, keeping the expensive models for the hardest posts.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Distilling LLM labels into smaller encoders
To cut inference costs, the author used LLM outputs as labels and trained smaller encoder models to reproduce them. The post-mortem reports roughly 30 times lower inference cost while retaining the same recall. That is a claim about this project’s models and data, and the post does not provide a public dataset or independent replication.
Measurement and evaluation
The author describes a frozen dataset, conditional recall, confidence intervals, nightly audits and a “golden cohort” of verified examples used to check the system over time. The F1 ceiling of 0.27 is the author’s explanation of how well the task could be solved at all. In plain terms, F1 combines precision and recall into one score, so a low ceiling means that even a carefully built classifier struggled to separate real buying intent from noise on this task.
Hardware and operational instability
The system reportedly ran on an NVIDIA GeForce RTX 5090 kept on the author’s desk. The post-mortem describes months of GPU instability, including Xid 109 errors, driver segmentation faults and a swap livelock, before the nightly run became reliable. This is an account of one personal setup, not a tested hardware review or a recommendation.
Why the business stalled
The author’s account points to distribution and reach, not classification quality, as the binding constraint.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Reddit: most of the leads, very little of the volume
Reddit made up only 0.2% of collected posts but roughly 80% of the true leads, according to the author. The problem was on the response side. The author says clients could not readily reply to Reddit leads from accounts without enough platform history, including karma and Contributor Quality Score. This describes the author’s experience with the platform at the time, not a complete or current statement of Reddit’s rules, so anyone building on Reddit data should check the platform’s current policies directly.
Cold email: one trial from 800 messages
The outbound effort produced one trial from 800 cold emails. The author attributes this to sender reputation and spam filtering. That is the author’s interpretation rather than a measured diagnosis: the post does not separate deliverability problems from weak offers or poor targeting.
A second pipeline for Reddit self-promoters
The author also built a second outreach pipeline aimed at Reddit users who promote their own products. It sent 60 emails and received no replies.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the author would do differently
The author’s closing lesson is short: test the market before building the system, using one lead found by hand to check whether a buyer wants the result. In the author’s words: “What I’d do differently fits in one line: test the market in February, with one lead found by hand, before writing the pipeline.” (Vivian O., project author, in the project post.)
That lesson is more specific than “validate your idea.” The pipeline could find intent at scale, but the project had no early proof that a vendor would pay for a lead, could contact the person who wrote the post, and would get a reply. Each of those steps was a separate risk, and the engineering work only tested the first one.
What the evidence does and does not establish
- Established by the author’s account: Dwelverson was a lead-intent product that scanned social posts for buying intent, and the project shut down.
- Reported by the author but not independently verified: the post volume, the inference costs, the trial count, the Reddit split, the F1 ceiling and the distillation savings.
- Not established in the available sources: the exact publication dates of the post and post-mortem, and the year of the September 15 shutdown.
- Interpretation, not measurement: the causes the author gives for the cold-email and Reddit outcomes.
- Not generalizable: these results cannot tell a reader what conversion, reply or cost to expect from a similar product.
For a builder, the useful takeaway is the sequence the author describes. Find one buyer, get that person to act on one lead, and only then decide how much of the pipeline is worth building.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
