Gemini and Gemma
What I love and hate about building with Gemini
I have used Gemini and Gemma across most of my hackathon projects. These are the experiences that stuck with me.
36 projects 34 competitions Gemini API, Gemini Live native audio, structured output, Gemma
What I loved
-
Native audio was fast enough that interrupting it felt natural
Room Sense at Gemini 3 Hackathon, Jan 2026
- What I was trying to do
- Room Sense is a rehearsal room for stand-up comedians. You upload a set, then talk it through out loud with four coaching voices, one each for premise, structure, cutting words and delivery. Comedy runs on timing, so the voices had to answer at the speed of a conversation, not a chatbot.
- What happened
- Instead of the usual chain of transcribing speech, sending text to a model and turning the reply back into audio, Room Sense streams raw microphone audio from the browser straight into Gemini's native-audio model over a live connection, and gets audio back. The replies came back fast enough that I could interrupt it and the exchange still felt like a conversation. The same model also picked up on hesitation and a lack of conviction in a performer's delivery in a way a text model never could.
- Why I loved it
- The transcribe-then-generate-then-speak pipeline is too slow for this and kills the joke. Without that speed there is no jam session, only a delay. Native audio is what made the live half of Room Sense possible at all.
What I hated
-
It fabricated citations routinely, even inside a valid JSON schema
Signal Packet API at Brainwave 2026 – X402 Blockchain Track, Jul 2026
- What I was trying to do
- Signal Packet API reads a hackathon submission, the description, the README and the docs, and reports which claims those documents actually back up, quoting the exact passage as proof. Judges get about ten minutes per entry, and sorting "asserted" from "documented" is mechanical work a machine should do for them. I asked Gemini 3.5 Flash for structured output: a JSON answer in a fixed shape, one entry per claim, each with a verbatim excerpt from the source.
- What happened
- The JSON came back in the right shape every time, and the excerpts inside it were regularly invented. Quotes that appeared nowhere in the README or docs I had fetched, presented as evidence. Not occasionally; routinely.
- Why I hated it
- A schema guarantees the shape of an answer, not its truth. I ended up treating the model's output as a proposal rather than a finding: the server checks every cited excerpt word for word against the fetched source, drops the ones that fail, downgrades the claim to "unclear", and reports the corrections to the caller instead of hiding them. Builder Signal, Judge Ops and Signal Review Reel all needed the same check. This is not a claim that Gemini "hallucinates" in general; it is what happened every time I asked it to quote evidence.
Everything I built with Gemini and Gemma
36 projects from the participation archive, newest first.
- Clarity
Clarity finds the one vague claim worth asking about, then uses a short adaptive phone call to turn it into a clearer fact, backed by the candidate’s own words.
- Audition Agent
Audition Agent turns an AI-generated character into a human casting workflow, creating auditions, organizing applicants, and handling follow-up while keeping casting decisions human.
- Payoff
Payoff helps creators design what an audience should feel, then uses real audience reactions to discover whether the story actually got them there.
- Builder Signal
Turns a technical builder’s real experience into evidence-backed, editable applications tailored to hackathons, fellowships, accelerators, and other selective programs.
- Judge Ops
Helps judges, recruiters, and program reviewers evaluate submissions faster with structured, evidence-backed analysis while keeping scores and final decisions human.
- Beatboard
Beatboard turns a comedy idea into a shoot-ready storyboard with timed beats, generated art, voiced dialogue, direction, and dependency-aware revisions in minutes.
- Signal Packet API
Send in a project write-up and get back a report on which of its claims the README and docs actually back up, with quotes as proof. Costs 5 cents per check, paid automatically with x402.
- Signal Review Reel
Give it a project description and its demo video. It watches the video and makes a short narrated clip showing which claims the demo actually proves, which it skips, and which it contradicts.
- Signal Evidence Graph
It maps a project's files to the claims they support in DataHub, then an AI agent reads that map and reports which claims the evidence proves. When it can't tell, it says so instead of guessing.
- First Session
First Session helps NYC parents find a Medicaid teen therapist by surfacing the highest-leverage contact path first. One email can reach 15 therapists. Fail fast, not by chasing the perfect match.
- TradeCast
Self-improving AI trader: Bull, Bear & Risk agents debate, paper-trade via Alpaca, auto-publish explainer Shorts to YouTube, and rewrite their own prompts by querying Arize Phoenix traces over MCP.
- AlreadyAsked
Semantic duplicate-post detection for Reddit mods. Catches paraphrased reposts that keyword bots miss, using sentence embeddings plus an LLM verification step to keep false positives low.
- Fan Meaning Map
Discover what kind of Team USA fan you are. Gemini decodes 4 quick answers into your fan archetype, your athlete matches, and the surprising reasons others see sport differently.
- Memory Moment
Helps caregivers find the right words to say to someone they love with memory loss, in the moment it matters.
- First Impression
See how the same face gets judged completely differently at first glance.
- Clinical Decision Room
A virtual tumor board for everyday clinical decisions. Four reasoning frameworks analyze the same patient in parallel, surface where they disagree, and show exactly what would flip the decision.
- Sacrifice Equation
A way to see how different perspectives judge your idea, and what tradeoffs make it succeed.
- Agent Identity Manager
A control center for AI agent identity. Assign roles, enforce permission boundaries via Auth0 Token Vault, and make every credential exchange and access decision visible in real time.
- Persona Fork
Take one idea, cook it different ways, and see what people actually like.
- Room Sense Live
Room Sense Live lets comedians test jokes with a live AI audience that laughs, reacts, and shows how different crowds interpret the same material.
- Drift Guard
Drift Guard turns every merge request into a lightweight architecture review, detecting structural drift before it becomes technical debt.
- Audience Mind Split
An interactive audience simulator that shows how different personas interpret the same line and lets you talk back to them in real time.
- Backlash Predictor
A multimodal AI tool that scans campaign materials to simulate audience reactions, surface backlash risks before launch, and suggest targeted rewrites without flattening the creative voice.
- Cook It
Cook It identifies difficult ingredients or tools in the video and offers tiered substitutions based on the user's reality.
- Learn It
Learn It turns coaching and tutorial videos into guidance that fits your time. Paste a link and get a deep dive, a quick overview, or the one thing that matters right now.
- Tiny Crossword
A tiny 5×5 crossword generated by AI using daily or player-chosen themes. Quick to solve, easy to play, and fun to race against puzzles created by the community.
- Play Console
A physical console for orchestrating live AI improv scenes. Shape energy, pacing, and relationships in real time, and jump into the scene yourself without typing prompts.
- Snipper
Transforming raw, long-form comedy footage into high-engagement social media content by leveraging the multimodal reasoning of Gemini 3 Flash.
- Laugh Spikes
Laugh Spikes treats stand up comedy like sports game tape by aligning audience laughter to transcript lines, visualizing joke impact, and generating data grounded AI feedback to improve a set.
- Room Sense
An AI native performance lab for comedians where you workshop your set in real time with four specialized voice agents powered by Gemini 3 and 2.5.
- Timing Coach
AI coach that teaches natural speech timing in 15 languages through real-time feedback and rhythm visualization, so you sound human, not textbook, across borders.
- SixView
SixView helps lender teams analyze complex loan agreements from six perspectives, surfacing risks, flexibility, and negotiation hotspots in minutes instead of weeks.
- YesAnd
AI improv partner you can talk to in real time. Practice scenes with natural voice interaction and get instant coaching powered by Gemini. Built with ElevenLabs Conversational AI + Google Cloud.
- Business Opportunity Finder
Spot high potential businesses for any neighborhood by analyzing local competitors, customer pain points, and unmet demand using Yelp data enhanced with Gemini AI.
- Roastify
An AI that roasts your photos and turns them into scroll-stopping, laugh-worthy captions in seconds.
- Meme Roaster
Compete with AI in a meme roast challenge. Create your funniest line, and let Reddit decide who delivers the better
Gemini and Gemma on the participation page All the platforms I built with