Conversational Agents Hack Day

BillRosetta is an AI-powered medical billing audit tool for self-insured employers. Upload a batch of EOBs (Explanations of Benefits) and BillRosetta audits every claim in seconds — cross-referencing CPT procedure codes against CMS Medicare benchmark rates to flag overcharges, upcoded procedures, and duplicate billing. The output isn’t a spreadsheet. It’s a dispute-ready document your HR or benefits team can send directly to the insurer or TPA. Self-insured employers (200–2,000 employees) pay their own health claims. They’re legally entitled to audit every one — but almost none do, because the tooling doesn’t exist. BillRosetta is that tool. One upload. Sixty seconds. A result you can understand. Built with Flask (Python), OCR via pytesseract, CPT/Medicare rate cross-referencing, and Stripe for billing. React dashboard for claim review and dispute tracking.

HackerSquadMossBright DataPython, Flask, pytesseract (OCR), Pillow, SQLite, Stripe, React, fpdf2

armani

A modern-day problem that our society faces is that we aren't able to converse with real-time people who work in customer support, especially. The problem with customer support is that they have thousands of calls and limited infrastructure, and now we see in the lives of AI agents. However, they're not really seamlessly being integrated, and they're not able to pull real-time data, and that's what our project is here to address.

HackerSquadMoss

Checkpoint is a local knowledge enrichment graph program that captures your screen locally, creates privacy-safe query expansions for BrightData, and retrieves local data with Moss.

HackerSquadMossBright DataSwift / Local LLMs

Prateek

Cutting platform hopping cost bringing BrightData into transient apps in Codex Chat and Moss low latency semantic search to iterate on the data. The argument is iterations in existing control surface buys intuition and helps develop better. The leverage is existing session memory serves context for Brightdata and propelled by semantic memory, the agent in chat can argue better for the user while user observability sits righ tat the front door.

HackerSquadMossBright Data

Slot Sniper

A voice AI agent that watches scarce, no-notification booking pages (starting with US visa appointment slots in India) and alerts the moment a slot opens — a hackathon project on the Bright Data + Moss + LiveKit + OpenAI stack.

HackerSquadMossBright Data

repomate

A voice-based co-pilot that helps a new developer understand a repository via a guided conversation. This aims at helping reduce onboarding time in enterprise teams.

i love space games

Bright Data

An AI newsroom apparatus for a single reporter: one sentence of beat prose in — a grounded, self-filtering, self-remembering, self-writing radar out. Sonnet 5 plans the radar with LIVE web grounding (Bright Data SERP as its search tool, reasoning streamed in the UI), then four Bright Data jobs scan in parallel — X profiles + Reddit discovery (Web Scraper API, async trigger→poll), news (SERP API), tracked sites (Web Unlocker). Every scraped item is relevance-judged against the beat with a written reason for each cut, fully auditable in the UI. Survivors cluster into stories carrying two independent judgments: editorial urgency (breaking/developing/context) and memory novelty (new/update/seen-before) — the latter checked against everything ever scanned, held in a Moss semantic index running in-process at ~3ms per query. The kicker is voice drafting: RAG on yourself. The reporter's own posts are scraped and indexed into Moss tagged isOwn; each story retrieves their 3 most-similar past posts as style exemplars, so drafts carry the reporter's actual cadence — reference posts displayed as proof. A persistent memory bar searches everything ever scanned with a live ms readout and a keyword↔semantic blend slider (Moss hybrid alpha). Slow sources fold in late and stories re-cluster + draft themselves live. Everything the AI decides — searches, filters, clusters, drafts — is visible as it happens.

HackerSquadMossBright DataVercel, Claude Code

EventGenius

WhoGo is a voice-first conversational agent that finds you the right Bay Area tech event and gets you registered — hands-free. The problem: the best AI/tech events are scattered across a dozen JavaScript-rendered, bot-protected platforms (Luma, Eventbrite, AGI House, Deep Tech Week, Frontier Tower, curated email digests), and none tell you which of tonight's 14 options is worth your time. Builders lose ~1.5–2 hrs/week to this search. How it works: You talk to WhoGo. Speech-to-text → an LLM "brain" (gpt-5.4-mini, structured JSON output) plans a search → the widget ranks a live pool of 220+ curated events on-device by a 1–5 "Builder's Scorecard" rating → it speaks the answer plus tappable refinement chips ("show-then-ask": results first, one smart follow-up second). Bright Data is WhoGo's eyes: the nightly ingestion pipeline uses Bright Data's Web Unlocker, SERP API, and Scraping Browser to extract events from those bot-protected, JS-rendered sources that plain HTTP can't reach. And the WhoGo Concierge closes the loop — say "sign me up" and WhoGo drives a Bright Data Scraping Browser (Playwright/CDP) session to register you on Luma. Moss is WhoGo's memory: every voice turn and 👍/👎 is written to a Moss semantic index, and each new search recalls your history to personalize the answer ("last week you loved small invite-only dinners — here are three this week"). Stack: Next.js 16 / React 19 / TypeScript on Vercel, OpenAI voice, Bright Data (web data), Moss (memory), Supabase, Python ingestion pipeline.

HackerSquadMossBright Data

AccessTeam

https://accessibility-defender-ni4ka0q0g-agents-v3.vercel.app/ Accessibility Defender **A Moss × Bright Data hackathon submission that helps small-business owners find common website accessibility barriers—and helps their developers turn those findings into scoped, paid remediation work.** Accessibility Defender turns a public business website into a focused accessibility report. Bright Data discovers the business and retrieves rendered HTML. Five deterministic DOM checks identify common problems. Moss connects each finding to pre-authored WCAG guidance and supports natural-language search across the report. The result is a plain-language Scorecard, a three-item Fix Roadmap, and exports a business owner can hand to a developer—or a developer can use to propose and deliver a focused remediation engagement. > Accessibility Defender is an early-warning and remediation-planning tool—not a legal opinion, certification, complete accessibility audit, or promise that a demand letter or lawsuit will not occur.

HackerSquadMossBright Datahttps://github.com/itsajchan/KOL-Copilot

go-gators

Deal Hunter is a voice shopping copilot. Ask for a product out loud and it searches the live web for current listings, ranks them cheapest-first, reads the best options back to you, and remembers the items you want to watch.

MossBright Data

Minions

MyMinion is a voice-first personal agent that completes real-world tasks using persistent memory, live web research, specialist agents, and multi-step mission execution and the end goal is to have it integrated with always on listening device which could be your daily journal and live recommendation device

HackerSquadMossBright Data

Karl

Acts as a local, reactive travel guide. Uses Moss so that it skips the dreaded 30-60 second wait, helping you really understand where you are.

MossBright DataOpenRouter

Every founder struggles with decisions and advice. We built an advisory board for the founder to help with different areas: market, fundraising, equity distribution, etc. Here, we demo 3 opinionated AI advisors debate pricing live competitor evidence — and disagree on purpose. The chair will take all the different persona opinions and provide a synthesis, not an average of the results. we use FLASK as a framework for decisioning. We use Bright Data for live access to competitor data and Moss for retrievel. Gemini and Gemini TTS deliver the LLM and audio capabilities

HackerSquadMossBright Data

Guardline is a real-time scam-call defense that listens to a phone conversation live and warns you the moment manipulation starts — while the scammer is still talking. Instead of matching keywords, it uses Moss as a semantic engine: every spoken sentence is queried against a vector index of known scam tactics, so Guardline recognizes a fraudster who's improvising or paraphrasing, not just repeating a script. When it detects a threat, it surfaces the truth (e.g. "the real IRS never demands gift cards"), tells you exactly what to do, and escalates a live threat meter as the pressure builds. And because scammers invent new scripts constantly, Guardline stays current with Bright Data — one click scrapes the latest scam alerts from sources like AARP and BBB, extracts a fresh pattern, and indexes it straight into Moss, so the system keeps learning new scams on its own.

HackerSquadMossBright DataRender

AidBay

AidBay is a voice-first service navigator for people experiencing homelessness and the outreach workers, neighbors, or caregivers supporting them. A user can say, “I’m helping a woman near 7th and Howard who needs somewhere safe tonight.” AidBay asks one clear question at a time, checks urgency, location, eligibility, and personal preferences, then recommends direct service providers with contact details, hours, availability evidence, and an explanation of why each option may fit. Users can open service cards, call an organization directly, return to AidBay after the call, report what happened, and continue searching. AidBay never promises that a bed or appointment is available. It distinguishes verified information from uncertain or temporary signals and encourages confirmation before traveling. Moss provides semantic retrieval across a curated index of San Francisco services. Bright Data supports collecting and refreshing fragmented public service information. ElevenLabs Conversational AI provides the natural voice conversation. AidBay’s evidence layer considers source quality and freshness so weak or stale reports are not presented as confirmed availability. AidBay is designed to be calm, concise, nonjudgmental, accessible to people with limited literacy or technology experience, and transparent about uncertainty.

HackerSquadMossBright DataElevenlabs, Codex, Figma

Edge Desk

Edge Desk is a voice-first brokerage desk for real-time support and trading. Callers dial in, verify with SMS 2FA at the start of the call, then talk to an agent that uses Moss for sub-10ms local retrieval over support and trading playbooks (no cloud vector-DB round-trip). Market context comes from Yahoo Finance (quotes, analyst ratings) with Bright Data Discover for company news (yfinance fallback). Orders, withdrawals to nicknamed bank methods, stats, and order-dispute tickets run against a sandbox ledger. Live demo at https://demo.kawuc.uk with an ops dashboard showing Moss latency and agent activity. Dial-in: +1 (515) 303-4334 via ElevenLabs + Twilio. # Edge Desk — Phone Demo (Judging) **Dial:** +1 (515) 303-4334 **Backup OTP:** `123456` **Dashboard:** https://demo.kawuc.uk/dashboard **App:** https://demo.kawuc.uk --- ## Before you start (30s) 1. Open the dashboard on a laptop. 2. Have the phone ready. 3. One person dials; one watches the dashboard (Moss ms + tools). --- ## Script (~2–3 minutes) ### 1. Answer + verify - Call → **“Edge Desk Trading.”** / code prompt - Enter SMS code, or say **`123456`** - Expect: **“Verified.”** ### 2. Moss pitch (to judges) > Retrieval runs locally with Moss — sub-10ms, no cloud vector-DB round-trip. Watch the dashboard for Moss latency. ### 3. Market analysis - **“Give me current market analysis on Apple.”** → short lean/buy + mean target + headline + “say more for depth” - **“More.”** → deeper distribution / targets / headlines ### 4. Quote + paper buy - **“Buy 5 Apple.”** → **“AAPL is $X. Say confirm.”** - **“Confirm.”** → **“Bought 5 AAPL at $X.”** - Should **not** ask to verify again ### 5. Account / support (pick 1–2) - **“What’s my volume?”** → stats - **“Recent orders.”** → last fills - **“What’s the PDT rule?”** → Moss support retrieve (point at dashboard) - Optional: **“Withdraw 100 to Chase checking.”** → **“Confirm.”** ### 6. Close Hang up. Point at dashboard + https://demo.kawuc.uk --- ## If something breaks | Issue | Fix | | --- | --- | | No SMS | Say **`123456`** | | Asks to verify again on buy | Hang up, redial, verify once, buy again | | Weak analysis | **“Analyst ratings for AAPL”** or **“Latest AAPL news”** | | Line silent | Short clear phrases after the agent finishes |

HackerSquadMossBright DataElevenLabs Agents

GreedyTrip is an anti-itinerary voice agent that applies greedy search to spontaneous travel. Bright Data discovers nearby candidates, Moss retrieves evolving user preferences, and GreedyTrip selects the highest-value next move. It recommends one place at a time, learns from every reaction, and recalculates whenever the user’s context changes.

MossBright Datagemini API, google map embed API

S

Checky

Checky is a new kind of offline shopping experience. People love offline shopping not because it's efficient, but because it's FUN. Checky keeps that joy and makes it smart, with a real-time AI agent built into your smart glasses.

HackerSquadMossBright Data

Greedy Trip

GreedyTrip is an anti-itinerary voice agent that applies greedy search to spontaneous travel. Bright Data discovers nearby candidates, Moss retrieves evolving user preferences, and GreedyTrip selects the highest-value next move. It recommends one place at a time, learns from every reaction, and recalculates whenever the user’s context changes.

MossBright Datagemini API, google map embed API

Woodshedding

this is an AI-powered music teacher capable of understanding the texture, tone, and voicing of your playing. Imagine having a personal coach with the knowledge and experience of legendary musicians from the 1940s, 1950s, and 1960s, providing personalized recommendations based on how you sound. Using sound frequency analysis, the system can listen to your performance, identify patterns, and generate feedback or execute complementary solos in near real time.

HackerSquadMossBright DataClaude, Lovable, Python, Typescript

Adina Finance

An interactive, real-time voice and text assistant that helps consumers maximize their credit card rewards and avoid hidden fee traps while protecting their overall cash flow. Demo video link: https://youtube.com/shorts/YVTwMpxstAI?is=M9tVA30DgAqIJII2

HackerSquadMossBright Data

M

nff_

We are building the deployment platform for embedded/robotics developers. Think Vercel, for hardware. Teams can now iterate with coding agents on firmware running live on robots in production. Talk to nff_ as if it was a firmware engineer.

HackerSquadMossClaude code, Typescript, FastMCP

Best project description

MossBright DataClaude Code, Codex, lovable, whatever have you.