Ben Nisevich

Speko

Founding engineer (YC S26). A router for voice models: one key, every model.

What I built: I build products on Speko's voice platform. My main one is Ring: you type what you need, and it calls up to five businesses at the same time, asks your questions, and compares the answers.
Hard part: Phone calls are slow and messy. I had to run parallel calls, let the voice agent call tools mid-call, and keep it strictly to gathering info (it never books or pays for anything).
What I learned: How voice AI fits together: speech-to-text, the LLM, text-to-speech, and failover between vendors.
Results: Ring works end to end. I also ship PRs to the platform and make launch content.

Todd

Former software engineer intern. AI for agriculture.

What I built: I worked on Todd's web platform, the marketing site and the client dashboard that farms and researchers use.
Hard part: It's a monorepo with a real team, real reviews, and real customers, so every change had to fit existing standards instead of my own habits.
What I learned: How a research company turns science into software people use, and how to write code other engineers can pick up after me.
Results: I shipped features and fixes to the dashboard and the marketing site.

Cyntral.AI

CTO & co-founder. Bootstrapped, $5k MRR.

What I built: I own the codebase for ConvoyLeads, a widget car dealers embed on their vehicle pages. A shopper clicks "Unlock Instant Price", verifies by SMS or email, and becomes a lead in the dealer's CRM.
Hard part: The widget runs on dealer websites we don't control (Dealer.com, Dealer Inspire, custom sites), so it has to survive any page. I also built the CRM delivery, including a CDK Fortellis integration and certification.
What I learned: Dealers care about leads in their CRM, not features, and every bug costs a sale.
Results: Bootstrapped to $5k MRR with live dealer customers, and working with automotive industry leaders.

Nonprofit hackathon

Organizer. 100 participants, $80k+ in sponsorship value.

What I built: I co-organized HackStorm, an AI + IoT hackathon for makers in Silicon Valley. I built the event website, worked on sponsors, and helped run the weekend.
Hard part: Getting sponsors to trust a student-run event. We had to pitch companies like Bambu Lab, Lovable and Alibaba AI Models with nothing but a plan.
What I learned: Running an event is a product launch with a hard deadline. Logistics, communication and small details matter as much as the big idea.
Results: 100 participants, $80k+ in sponsorship value, and 45+ hours of organizing.

Contract development

$15k+ in revenue. Websites serving 50k+ monthly visits.

What I built: I build fast, SEO-friendly websites for small businesses and local groups, like a site for The Graystones, a Bay Area rock band. I handle design, build, hosting and upkeep.
Hard part: Clients don't speak in specs. I had to turn vague asks into a site they love, and keep it fast and cheap to host on Cloudflare and AWS.
What I learned: How to scope work, price it, and deliver on time. Also that performance and SEO are what actually bring visitors.
Results: $15k+ in revenue, with sites serving 50k+ monthly visits.

Bay Builders Hackathon

1st place, Best Use of Nimbleway.

What I built: I built Scout, an "AI SDR floor": a team of agents that runs outbound sales end to end. You give it one goal, like "find used-car dealerships in the Bay Area worth reaching out to", and agents find, research and draft outreach.
Hard part: Making agents trustworthy. Every "why now" signal needs a real citation, the writer agent has no send ability, and a compliance agent vetoes made-up claims.
What I learned: Agents work best as a team with narrow jobs and clear limits, not one big prompt. I also learned to design so the demo never breaks, even with no API keys.
Results: 1st place and Best Use of Nimbleway at the Bay Builders Hackathon at AWS.

Agents You Love Hackathon

2nd place at Frontier Tower.

What I built: I built Witness with my team. It finds the invisible work engineers do, like unblocking a teammate or catching a bug in someone else's PR, and returns cited evidence instead of a score.
Hard part: Credit attribution without guessing. I made it a time-based join over Slack, tickets and PRs in Postgres (InsForge), not an LLM call, so every finding can be checked.
What I learned: The best AI products know when not to use AI. A plain, checkable rule was more convincing to judges than a clever model.
Results: 2nd place at the Agents You Love Hackathon at Frontier Tower.

Hackstorm

2nd place, sponsored by Tuya Smart, Stanford & Berkeley. $3,000 prize.

What I built: I built PracticePod, an AI music coach: a 3D-printed device on a Tuya T5AI board, a Flutter app, and a cloud backend that listens to you practice and gives feedback.
Hard part: No reliable sheet-music API exists, so I automated SoundSlice with Playwright to extract MusicXML. Then I matched it to pitch tracking (PYIN) and had Gemini write the coaching.
What I learned: How to ship across hardware, firmware, mobile and cloud in one weekend, and how much faster things go when you design the data flow first.
Results: 2nd place and a $3,000 prize at the hackathon sponsored by Tuya Smart, Stanford and Berkeley.

Buzzle

Live multiplayer classroom games, gamifying learning.

Live site

What I built: I co-build Buzzle, live multiplayer games for classrooms. A teacher types a topic, and the whole class is playing in under a minute. I work on the games, the realtime server and the teacher host screen.
Hard part: Real-time multiplayer on school Chromebooks and bad Wi-Fi. For games like Dash and Climb, I moved server and client physics into one shared module so prediction holds up under lag.
What I learned: Netcode is hard, and tests are the only way to trust it. We run network simulation tests on every push.
Results: Live at playbuzzle.com and used in real classrooms.

Summurcal

A quiet calendar for summer plans.

Live site

What I built: I built Summurcal by myself: a quiet three-month calendar for summer. You can add events by hand or just describe them, and Gemini places them on the right days.
Hard part: Turning messy natural language ("beach trip the second week of July") into correct dates, and keeping the UI calm and simple.
What I learned: Small, focused tools are easier to finish and nicer to use. I also got better at Supabase and prompt design for structured output.
Results: Live at summurcal.com.

Lifeboard

A vision-board canvas for media, links & notes.

What I built: I built Lifeboard, a vision-board canvas where you pin images, links and notes. It's React and TypeScript on the frontend, an Express API on Cloud Run, Firebase Auth and Firestore.
Hard part: Uploading media safely. I kept the S3 bucket private and used presigned upload and download URLs, with Firebase tokens checked on the backend.
What I learned: This was my project to learn AWS and real backend security, not just make a UI.
Results: Live at lifeboard.studio.

Keen

Seven micro-games that measure how precise your senses really are.

What I built: I built Keen, seven quick browser games that test how precise your senses are: timing, color, pitch and more.
Hard part: Making each test fair and measurable in a browser, where screens, audio and input lag are all different.
What I learned: Small games need a lot of polish. Feel and feedback matter more than features.
Results: Live and free to play on the web.

Spotted

AI car-spotting social app on iOS.

App Store

What I built: I built Spotted, an iOS app for car spotters. Take a photo, the AI identifies the car with specs and rarity, and you share it with other spotters.
Hard part: Accurate recognition of rare exotic cars, plus privacy. I added automatic license-plate blurring before anything is shared.
What I learned: How to take an app all the way from idea to the App Store, including review, onboarding and marketing.
Results: Live on the App Store. Part of my 1,000+ downloads across 70+ countries.

Wudly

A practice tracker to watch your progress pay off.

What I built: I built Wudly, a practice tracker for iOS and Android. You record short videos of your practice, and it makes before-and-after comparisons so you can see your progress.
Hard part: Video on mobile: recording, trimming, and making side-by-side comparisons on the phone. It's local-first with optional cloud sync, and streaks recalculate when you delete a clip.
What I learned: Local-first design is harder but much better for users. I also learned to ship one Flutter app to two stores.
Results: Live on the App Store and Google Play.

Snap

AI outfit-rating app on iOS.

App Store

What I built: I built Snap, an iOS app that rates your outfits. It scores fit, color, style, occasion and overall look using Gemini.
Hard part: Getting AI feedback that is consistent and useful, not random. I had to structure the prompts and scores carefully.
What I learned: How to scope a simple, fun app and ship it fast.
Results: Live on the App Store.

Synqit

One shared main, no branches to review.

What I built: I built Synqit, a coordination layer on top of Git. You run "synqit push" with a feature, and it integrates the change into the latest main and syncs everyone. No branches or PRs to manage.
Hard part: Merging parallel work automatically without breaking main, especially when many AI agents write code at the same time.
What I learned: Git is great at history but hard for coordination. Designing around intent instead of branches changed how I think about dev tools.
Results: Working CLI and demo, open source on GitHub.

Witness

Your engineering dashboard is lying to you.

Live site

What I built: I built Witness with my team at a hackathon. It finds work that helped someone else's ticket or PR and gives managers cited evidence, never a score.
Hard part: Making the output honest when data is missing. If you remove a source, the tool says so instead of giving a confident, wrong review.
What I learned: Trust comes from evidence you can check. That changed how I design AI tools.
Results: 2nd place at the Agents You Love Hackathon.