Inspiration
The idea for DevPulse came from a frustration we've seen play out too many times in engineering teams: a talented developer goes quiet, their commits slow down, their messages get shorter — and nobody notices until they hand in their resignation.Burnout doesn't announce itself. It creeps in gradually, hidden behind closed laptops and late-night commit timestamps. By the time a manager sees it, the cost is already enormous — not just financially ($50,000–$85,000 to replace a mid-level developer), but in team morale, knowledge loss, and delivery delays.We kept asking: all this behavioral data already exists inside Git platforms. Every commit, every review, every timestamp is a signal. Why isn't anyone listening to it? That question became DevPulse.
What it does
DevPulse is an automated developer wellbeing intelligence agent that monitors behavioral signals across Git platforms — GitHub, Gitee, GitLab, CODING, and Alibaba Codeup — and surfaces a private, empathetic nudge to the team lead when patterns suggest someone may need a check-in.When a developer makes a commit, the Git platform fires a webhook to DevPulse instantly. The app logs the event — timestamp, hour of day, commit message length, platform — and builds up a behavioral picture over time. When the team lead runs an analysis, the AI compares each developer against their own personal baseline, looks for clusters of signals like late-night commits, declining activity, and shorter messages, and generates a risk assessment with a suggested check-in message.Alerts go only to the team lead — never to HR, executives, or other developers. Developers can self-report context like vacation or a focused sprint to suppress alerts for any period. Every alert explicitly states it is a signal for human review, not a diagnosis.No surveillance. No scoring. No ranking. Just a timely, informed heads-up and a suggested message to start a conversation.
How we built it
We built DevPulse entirely through MeDo's multi-turn conversation interface across five focused sessions — no code written manually. Session 1 — UI: Described all five screens and the dark navy design language. MeDo generated the complete React frontend with sidebar navigation, developer cards, and status badges. Session 2 — Database: Described a two-table Airtable schema and asked MeDo to build the integration. It generated a full Supabase Edge Function with complete CRUD operations, HMAC signature verification, and filterByFormula queries. Session 3 — AI Engine: Described the burnout detection logic in plain language — personal baseline calculation, cluster detection, suppression logic. MeDo wired this to Google Gemini and built the complete analysis pipeline returning structured JSON with risk level, signals, wellbeing tips, and a personalized check-in message. Session 4 — Webhooks: Described the real-time ingestion requirements. MeDo built the webhook receiver handling push and PR events from all five platforms, with Chinese-language setup instructions included for Gitee and CODING. Session 5 — Polish: Asked for a full production pass. MeDo updated 82 files with zero lint or TypeScript errors — onboarding persistence, activity bar charts, IST time formatting, email alerts via Resend, self-report suppression, animations, and a #BuiltWithMeDo footer.
Challenges we ran into
Webhook receiving: Our first question was whether a MeDo-deployed app could receive incoming POST requests from external services. We tested this before committing to the architecture. It worked — and that confirmation unlocked the entire real-time pipeline. Persistent storage: MeDo's native storage handles form submissions, not time-series behavioral data. We solved this by building an Airtable custom plugin via MeDo's API integration, giving DevPulse a proper database without leaving the platform. API debugging: Connecting Airtable required fixing a 404 (needed Table IDs not names), a 401 (token scope missing), and a 422 (POST vs PATCH logic). Each was resolved with a single precise MeDo prompt. LLM provider access: We designed the analysis engine to use Baidu ERNIE to align with the hackathon sponsor, but Qianfan API console access proved difficult outside mainland China. We pivoted to Google Gemini and architected the app to swap LLM providers via a single environment variable. Credit efficiency: Building a production app on a credit budget forced discipline — one precise comprehensive prompt instead of back-and-forth. That constraint actually made the build cleaner and faster.
Accomplishments that we're proud of
The privacy and ethics framework. It would have been easy to build a surveillance tool dressed up as a wellness feature. We went the other direction — every design decision was made to protect developers, not expose them. Alerts go only to team leads. Every message frames signals as prompts for human conversation, never as verdicts. Developers can suppress alerts themselves. The system compares each developer against their own baseline — never against teammates. Data stays within the team's environment. We're also proud that the entire production-grade app — webhook ingestion, persistent storage, AI reasoning, email delivery, real-time dashboard — was built through conversation alone. No code written manually. That's what MeDo makes possible.
What we learned
Describing logic precisely is a superpower. The more clearly we described what we wanted — field names, API formats, conditional logic, error handling — the better MeDo's output. Vague prompts produced vague results. Precise prompts produced production-ready code. Architecture decisions matter more than individual prompts. Choosing webhooks over polling, Airtable over a custom database, personal baselines over team averages — these decisions shaped the entire app. MeDo executes the architecture you give it, so thinking it through first paid off enormously. Privacy is a product decision, not a legal footnote. The most meaningful feedback we got during testing was from people who said "I'd actually trust this." That trust came from deliberate design choices, not compliance checkboxes. Multi-platform thinking from day one. Including Gitee and CODING alongside GitHub from the start — rather than as an afterthought — made the app genuinely relevant for Chinese engineering teams. For a Baidu-sponsored hackathon, that cultural awareness mattered.
What's next for DevPulse
DevPulse is built on data companies already have. The natural next steps are:
Team-level dashboard for Engineering Managers showing aggregate wellbeing trends across the entire team Workload rebalancing suggestions — redistribute tasks before burnout hits, not after Calendar integration to factor in meeting load alongside commit activity Predictive attrition scoring per developer over a rolling 6-month window Baidu ERNIE integration as the primary AI model for Chinese enterprise customers Extension beyond developers to QA engineers, DevOps, and other Git-adjacent roles Anonymous peer support nudges within the team
The $300 billion annual cost of developer attrition is a problem waiting for a product. DevPulse is the start of that product.
Built With
- airtable
- airtable-plugin
- gemini
- medo
- react
- resend
- supabase
- tailwind
- typescript
Log in or sign up for Devpost to join the conversation.