Syed Mohammad Rizwan

Seven AI agents play Werewolf; people watch and bet, or take a seat.

• A LangGraph state machine runs setup → night → morning → discussion → voting, with a win check each round.
• 7 players per game (2 wolves, 1 seer, 4 villagers) drawn from 12 characters with personas.
• Each agent keeps private reasoning separate from its public speech; the UI shows both side by side.
• Model chain: Groq llama-3.3-70b → llama-3.1-8b → OpenAI gpt-4o-mini → deterministic mock, so a game never stalls.
• The server filters hidden roles out of the bettor view.
• At the hackathon bets settled on-chain on Monad; the current build uses play money.

Answers California regulation questions, citing a section for every claim.

• Async breadth-first crawl collected 74,975 of 74,997 discovered regulation sections (99.97%).
• Extraction runs 20 at a time and checkpoints every 100 records, so a crash resumes instead of restarting.
• Sections are embedded (all-MiniLM-L6-v2) into Pinecone; the agent retrieves the top 8.
• The model answers only from retrieved sections, cites one per claim, and asks a follow-up when context is insufficient.
• Configurable model with a fallback; visitors see a plain message on failure, never a stack trace.

Offline-first net-worth tracker for India, with an assistant that asks before it acts.

• The assistant turns speech or text into one validated action; nothing is written until the user confirms.
• The model has no database access.
• Offline-first PWA with an IndexedDB mutation queue; Supabase with row-level security.
• 22 asset classes, 7 calculators, Excel/CSV import, PDF statements.
• Blog pipeline falls back across Gemini, Groq and OpenAI.
• A React Native (Expo) mobile app reuses the same Clerk and Supabase backend, with an on-device SQLite cache, a mutation queue and a sync engine.

Nine voice agents of his own in Murf AI's ten-day challenge, one a day after the starter.

• Each agent runs on LiveKit Agents with Deepgram nova-3 speech-to-text, Gemini 2.5 Flash and Murf Falcon text-to-speech.
• Day 1 brings up Murf's starter agent; days 2–10 are nine agents of his own: a coffee barista with a live order receipt, a wellness companion, an active-recall tutor, a Razorpay SDR, a fraud-alert desk with a strict security question, Instamart ordering, a D&D game master, an agentic-commerce shop and an improv game, one branch per day.
• Each agent keeps its own structured state (order, lead, fraud case) that the UI renders live.

A live player auction for a college league, with AI commentary in Telugu.

• Bids and player releases run as locked Postgres functions, so purse, points and roster always move together.
• Supabase Realtime pushes every bid to the projector and to phones.
• Commentary picks a meme template locally, then Groq personalises it: qwen3.8-27b, then gpt-oss-120b, each with an 8s timeout; if both fail, the browser shows a local template line.
• 87 players, 7 franchises, 124 Telugu meme lines.

A 4 ft hardware rehab companion robot that turns exercise into play for people with limited mobility.

• Built over two days at Impact Lab Bengaluru, a “hacking human ability” hackathon presented by Claude Community Events.
• A phone is the face (expressive eyes, voice and games) on a PVC neck over a LEGO base; the demo base was remote-driven.
• The design wires ultrasonic and motion sensors through an ESP32 to the base and head motors, with the phone linked over Wi-Fi or Bluetooth.
• Rehab as play: follow-me walks, air piano, a boxing partner, and reaction and memory games.

Learn Malayalam and Kannada through Tamil, with pronunciation scored by voice.

• Whisper transcribes the learner; a normalised edit-distance score turns it into 0–100.
• An AI tutor and quiz generation run on Groq Llama 3.3 70B.
• 8 modules and 140 vocabulary entries with native script, transliteration and the Tamil cognate.

An emotional companion that answers in verse, the way he would, after chatbots never felt like company.

• Three model steps: classify the message into one of 13 emotional modes, write a verse with that mode's own prompt, then add an English translation.
• An unclear or unknown mode falls back to Safeguard mode, and a flat “I’m fine” is read as hidden hurt, not as fine.
• Nine South Asian languages and scripts (Hindi-Urdu by default, Tamil, Telugu, Bengali and more) with a tone picker; signed-in visitors keep a verse history.
• He built it on Lovable, an AI app builder; it is live at your-souls-echo.lovable.app.

Team productivity dashboard that forecasts task trends, built in 12 hours at his first hackathon.

• Built in 12 hours at Terralogic's Hackathon 2025 in Nellore, his first hackathon.
• FastAPI backend on Neon Postgres; a cron job simulates an organisation at work, generating synthetic tasks and activity.
• An in-house model forecasts daily task trends and predicts the probability a task closes within 24 hours.
• Live KPIs, team performance and an AI summary, plus natural-language questions about the team's productivity.

Donors find verified NGOs nearby and pledge exactly what they need.

• Radius search uses a haversine query; pledges expire after 48 hours via scheduled jobs.
• NGO trust score from 0 to 100; donor, NGO and admin roles verified with Firebase tokens.
• 71 Java source files and 15 test files on Spring Boot 3.