The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
I built Vetra, an AI technical interviewer that makes interviews more structured, adaptive, and realistic.
Vetra conducts real-time voice interviews, explores a candidate’s experience, adapts its questions based on their responses, evaluates performance, and generates structured assessments.
I built the interview orchestration with LangGraph, added real-time voice interaction using Gemini Live, and created a separate evaluation layer for evidence-based scoring.
The goal was to create an AI interview experience that feels more like a real technical conversation than a chatbot.
Jonathan, keeping the evaluation layer separate from the interviewer is a smart choice. When the same model asks and scores, it tends to grade its own questions kindly. How are you handling silence and long pauses on the voice side? That was the hardest part for us on a voice AI build.
Thanks! Silence was definitely something I had to account for. I use calibrated VAD thresholds plus transcript buffering, so short thinking pauses don’t immediately trigger a new turn. The evaluator also runs asynchronously, so it never adds latency to the live conversation.
I can send you the demo if you want to try it out
Running the evaluator async is the right call, latency is the first thing callers notice. Yes please, send the demo over. Happy to give it a proper try and share what I find.
I reached out to you on linkedin
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Related posts
🐾 Paw & Polish
I built Paw & Polish, a pet grooming booking experience focused on making appointments easier for pet parents with multiple pets.
Instead of going through calls or messages, customers can add up to 5 pets, choose different services for each, see the total price and grooming time, and find a suitable slot—all in one flow.
I also added two booking options:
Book it myself — for customers who know what they need.
Let Paw & Polish help — a conversational AI flow that understands the pet’s needs, recommends services, builds the appointment, and checks availability.
I also included a lightweight admin portal so the business can manage bookings, pets, services, and availability.
The goal was to rethink the first step of the customer journey, rather than build just another booking calendar.
Built with Lovable.
#Lovable #LovableChallenge #VibeCoding #ProductDesign #AI #BuildInPublic
Designed and built an AI operations dashboard that brings projects, decisions, tasks, research, connected tools, and voice interaction into one clear workspace. The goal is to turn complex workflows into a simple, usable system that helps people see what matters and act faster.
James, putting decisions right next to tasks is the part I like most. Most ops tools track what got done but lose why it was decided, and that is exactly what people go looking for three weeks later. How are you using the voice side day to day?
Building AI agents is easy. Building one you can trust with a real business workflow is harder.
I built a Freelance Admin Agent to explore that problem using Strands Agents SDK and Claude Sonnet 4.5 on Amazon Bedrock.
The working MVP takes manually provided client emails and turns them into a structured payment workflow:
→ Classifies invoice and payment emails
→ Extracts client, invoice, amount, currency, dates and contextual details using structured output
→ Stores and matches records in SQLite
→ Uses deterministic logic to identify full/short and on-time/late payments
→ Uses a second agent to draft a contextual follow-up only when needed
→ Keeps the final decision with the human through Approve / Edit / Reject
One of the most valuable parts wasn't adding another AI feature. It was finding a real reliability issue during testing: reprocessing the same email could duplicate records and inflate payment totals.
I reproduced it through repeat testing, traced the database behavior, and fixed it.
That reinforced something I'm increasingly applying to my AI engineering work:
A useful AI product isn't just an LLM call. It needs clear boundaries between AI reasoning, structured data, deterministic business logic, persistence, testing, and human control.
The next iteration will focus on stronger production safeguards, including source-grounded financial extraction, precise currency handling, stricter trust boundaries for untrusted input, and eventually live inbox integrations.
I'm interested in building more systems like this for startups and teams with repetitive workflows that can genuinely benefit from AI agents and automation.
Stack: Python · Strands Agents SDK · Amazon Bedrock · Claude Sonnet 4.5 · Pydantic · SQLite · Streamlit
turning client emails into admin tasks is such a real freelancer pain, nice build 🔥