Lifemaxxers

Inspiration

University is where careers launch and bodies are built — but most guys navigate it alone. They bounce between generic YouTube fitness videos, random Reddit career threads, and contradictory advice from friends. No structure, no accountability, no personalization. Existing self-improvement apps treat everyone the same: the same workout, the same career advice, the same generic tips. We built Lifemaxx because a skinny computer science major targeting a FAANG internship needs a fundamentally different path than a muscular biology student aiming for med school. One-size-fits-all doesn't work — and it especially doesn't work for the high-stakes, high-pressure years of university.

What it does

Lifemaxx is a gamified self-improvement PWA that gives university-age males a personalized AI coach and a tailor-made weekly roadmap. Users complete a detailed onboarding covering body type, goals, career targets, GPA, and industry — then DeepSeek V4 Pro generates a fully structured five-week roadmap across two tracks (health: fitness, nutrition, grooming, and career: academics, professional growth). Every task gets a real calendar date, a YouTube tutorial link, and knowledge context pulled from a curated expert database via RAG. Users choose from four AI coach personalities — Drill Sergeant for tough accountability, Mentor for evidence-based guidance, Hype Man for relentless positivity, or Custom where they describe their ideal coach in plain English and the AI generates a unique persona on the spot. The coach sees their full profile, roadmap progress, yesterday's conversation, and relevant knowledge-base content in every interaction. The app tracks daily task completion, syncs to Google Calendar, maintains a daily log, and awards points with persona-specific affirmations — all wrapped in an installable PWA that works across any device.

How we built it

Frontend: Next.js 16 App Router deployed on Vercel as a PWA. shadcn/ui components with Radix primitives, Tailwind CSS, and next-themes for dark/light mode. AI: DeepSeek V4 Flash for chat responses, DeepSeek V4 Pro for roadmap and custom persona generation. BGE-M3 embeddings via DeepInfra API power the RAG pipeline, with Pinecone as the vector database and LangChain.js for orchestration. A curated knowledge base of fitness, nutrition, grooming, study skills, and career content is chunked, embedded, and indexed in Pinecone. Database: DynamoDB single-table design (LevelUpTable) with partition key PK: USER#{userId} and sort key prefixes for each entity type (PROFILE, STAGE, TASK, LOG, MSG, ROADMAP). Dashboard loads use BatchGetItem for single-call aggregation. Task completions use TransactWriteItems to atomically update the task, stage progress, and daily log together. Onboarding writes 60+ items across BatchWriteItem batches of 25. S3 backs up roadmap JSON artifacts for durability. Auth & Integrations: Google OAuth with offline access for Calendar sync. Users can connect their Google Calendar and all roadmap tasks automatically create calendar events on their scheduled dates.

Challenges we ran into

Roadmap JSON truncation: DeepSeek V4 Pro's output token limit wasn't enough for a single call generating both health and career tracks (10 stages, 50+ tasks). The model would cut off mid-JSON. We split it into two parallel per-track calls, each with its own full token budget, and added a bracket-closing fallback parser for truncated output. Markdown leaking into chat responses: Even with explicit "NEVER use markdown" rules in all three persona prompts, DeepSeek's training bias toward formatting would leak asterisks, hashes, and backticks into replies. We added a stripMarkdown() post-processing layer that sanitizes every response before it reaches the client — prompt prevention plus regex defense.

Accomplishments that we're proud of

• Fully personalized roadmaps generated in under 2 minutes — five weeks of daily tasks across two tracks, each task enriched with RAG knowledge, YouTube links, and real calendar dates
• Four AI personas including a custom persona generator that transforms plain English descriptions into full system prompts with names, taglines, and distinct coaching personalities
• Single-table DynamoDB that handles every access pattern without a single GSI — profile, stages, tasks, logs, and chat history all under one partition key
• Atomic task completions via TransactWriteItems — task, stage, and daily log update together or not at all
• Installable PWA with full manifest, cross-platform icons, and Google Calendar sync — zero service worker needed for basic install
• TypeScript across the entire stack — frontend, API routes, and server libraries all share types from a single source of truth

What we learned

• Split before you scale: Two parallel LLM calls with focused prompts are more reliable than one mega-prompt with a high token limit. Truncation isn't solved by budget — it's solved by scope.
• LLMs need defense-in-depth for formatting. Prompt instructions alone aren't enough to control output format — a post-processing stripping layer catches what the prompt misses.
• Single-table DynamoDB design forces clarity. Every access pattern must be defined up front, and when it works, it works beautifully — one table, no joins, predictable performance.
• Batch writes overwrite completely. A PutRequest in BatchWriteItem replaces the entire item, not merges fields. Timing writes carefully or using UpdateItem avoids silent data loss.

What's next for Lifemaxx

• Streak system — track consecutive days of task completion with visual rewards and escalating multipliers
• Stage completion gating — hybrid progression that requires both the calendar date AND the previous stage reaching a completion threshold before unlocking the next stage
• Push notifications — daily reminders for scheduled tasks, coach check-ins, and streak warnings
• Offline support — service worker caching for dashboard, task completion, and chat history without connectivity
• Mobile native feel — haptic feedback on task completion, gesture navigation, and PWA shortcuts
• Expanded knowledge base — more domains (personal finance, mental health, relationships) indexed in Pinecone for richer RAG context
• Roadmap regeneration — periodic re-evaluation where the AI coach analyzes progress and adjusts upcoming tasks based on what's working and what isn't

Built With

• amazon-web-services
• dynamodb
• next.js
• rag
• react-native
• s3
• v0
• vercel