SellYourLogic · Expert Finder (Logi)
One AI engine, two very different jobs
A company describes a need, "a data analyst for 2 weeks, under $100/h", and gets a shortlist of real experts. Same agent pattern as level 4, running on the engine SellYourLogic already had for the Interview Coach.
The business need
Companies describe work in their own words, rarely in our skill names. They need the closest real experts, honestly labelled, and a quick path to a hire request they send themselves.
12 / 12
evaluation cases passed on the live site
2.7 s
median answer time in that run
10×
faster after fixing one wrongly named setting
How one request flows
From the visitor's words to what they see, step by step.
-
1
Guard
Emails and phone numbers are refused before any model call.
-
2
Same providers
The same ordered provider list, bench and "Pause all AI" switch as the Interview Coach.
-
3
Search experts
Filters in SQL, then ranking by meaning against each profile's vector.
-
4
Relax if empty
Nobody fits every filter? Search again without them and say these are the closest.
-
5
Draft
A hire request is prepared as a pre-filled form, never sent.
-
6
Cards
Expert cards from tool results; demo profiles labelled on every card.
The key ideas
Reuse the engine, not just the idea
No new model or account: the finder reads the same provider list and fallback as the Interview Coach. Only its tools, prompt and limits are new.
Tool calling on any provider
A library translates the same tool definitions for a local model, Google Gemini, OpenAI or Anthropic, so the identical agent runs wherever an AI is available.
Search by meaning over people
Every profile (headline, field, skills, bio, problems solved) is embedded. "A chatbot for our support team" finds language-model experts even if no profile says "chatbot".
Re-embed only what changed
A hash of each profile's text means only edited profiles get new vectors.
Graceful relaxation
Strict filters returning nobody is a bad answer. The search relaxes, marks the result "not an exact match", and the model says so.
Real people first, demos labelled
Real experts always rank above sample profiles, and every sample is labelled on its card.
Honest about third parties
When a hosted AI is switched on, the chat names the company that may process the message.
No gender guessing
The model once wrote "his rate". The prompt now forbids guessing anyone's gender, and names are used instead.
Evaluate what the agent did, not its words
Twelve fixed questions check behaviour: which tools were called, with which filters, which cards and drafts appeared, and what was refused. Every answer is also checked for links, HTML, pretend cards and gendered pronouns. It runs before any model change.
A trace for every question
An admin can ask a question and see each tool call with its arguments and result, which model answered and how long it took. That is how the bugs below were found.
What testing taught us
Found by running the real model, not by guessing. This is the most useful part.
It drew the cards itself
Our code shows the expert cards, but the model sometimes copied them into its reply: once as "[Expert card for …]", once as a block of raw HTML. Visitors would have seen that as text. A longer prompt did not stop it, because a small model keeps finding new formats. The fix is code: the reply is cut where a pretend card or any HTML begins, and the evaluation now fails any answer that contains one.
It said "I have prepared a draft", and had not
Asked to prepare a hire request, the model often wrote that the draft was ready without ever calling the tool that makes it. The visitor would have looked for a draft that did not exist. Now our code checks after every answer: if the visitor asked for a hire request, or the model claims to have made one, and no draft exists, code drafts it for the best match, reading the budget from the visitor's own words with a rule. In five tries afterwards, five real drafts.
One wrongly named setting made it ten times slower
The local model can "think" before it answers. We turned that off, but under the wrong option name, so it was silently ignored. The model spent its whole answer budget thinking: answers took about 10 seconds, and some came back empty. With the right name, the same answers take about 1 second. A trace showing time and tokens per answer made it visible.
"No experts found" was a bug, not an answer
The model treated "part-time" as availability and filtered everyone out. The tool description now says part-time is not availability, and an empty result relaxes the filters.
Tests must never reach a real model
An early test quietly used the real AI through a leftover provider. Tests now pin a fake driver and remove providers when testing the offline path.
The same rule, one more time
The model talks; our code decides. Experts, rates and the hire request all come from our data.
What you learn at this level
- Reusing an AI engine
- Provider-independent tool calling
- Semantic search over people
- Graceful relaxation
- Evaluating an agent
- Verify, don't trust
- Honest labelling
Where it stops
The evaluation set has only twelve questions, written by us, so it is easier than real visitors. It grows whenever a real conversation goes wrong. Reranking by rating and vetting is next.
AI Playground
Change chunk size, similarity threshold, model and prompt on our case studies, and see what was retrieved, the exact prompt, the answer, the speed and the cost.
Open the Playground →Live workshops
Hands-on sessions that build each level step by step, hosted by our founder and by experts on SellYourLogic.