Case Study: BOB
HAI in Production Development
Every feature, every bug fix, every spec — tagged with who did what.
BOB is an AI customer communication platform built at serban.eu.com. From the first endpoint to production deployment, every piece of work was tracked with HAI. This is not a theoretical exercise — these are real numbers from a real platform serving real customers.
0+
API endpoints
0+
Tests passing
0%
Work tracked with HAI
AI-led
~65% of work itemsProduction code, CRUD endpoints, test suites, migrations
Example: Adding a new API endpoint following established patterns. Spec is clear, AI implements, human reviews the output against the spec.
Collaborative
~25% of work itemsArchitecture decisions, feature design, complex integrations
Example: Designing the multi-channel conversation engine. Human specified the requirements, AI proposed architectures, human evaluated tradeoffs, AI implemented the chosen approach.
Human-led
~10% of work itemsBusiness decisions, pricing strategy, partner negotiations
Example: Defining the pricing model and client communication strategy. Human drove every decision, AI provided market research and drafted templates.
8 Communication Channels
Voice, SMS, WhatsApp, Web Chat, Email, Facebook Messenger, Instagram DMs, Telegram — unified under one agent configuration.
Multi-Tenant Architecture
Full client isolation with per-agent database schemas. Separate test, dev, and demo environments with safety guards.
Admin Dashboard
Next.js admin with live conversation monitoring, agent configuration, analytics, and the HAI Development Tracker itself.
Security Audit
50-finding security audit (47 resolved). JWT auth with Redis blacklist, TLS, encrypted backups, fail2ban, structured audit logging.
AI Engine
OpenAI Responses API integration with function calling, conversation memory, sentiment tracking, and multi-channel context.
HAI Development Tracker
The tracking system itself was built using HAI. Work items, phase transitions, assessments, and retrospectives — all tracked in the admin dashboard.
Spec Quality Impact on Delivery
Features with detailed acceptance criteria, testable conditions, and defined scope were implemented with minimal iteration. The AI could execute confidently because the specification was clear.
Security audit implementation (SEC-001): 50 findings documented with specific acceptance criteria each. 47 resolved in first pass.
Vague specifications led to implementation that missed the intent. The AI produced working code, but it wasn’t what was needed. The problem was always the spec, not the production.
Initial dashboard design: “make it look good” produced technically functional UI that didn’t match the team’s vision. Respecified with reference examples and design tokens — second iteration was right.
Most collaborative work starts at medium spec quality. Direction is clear, details emerge through iteration. This is where the human-AI feedback loop is most visible.
HAI Framework itself: human provided direction (“replace time estimates”), AI proposed structures, human redirected (“too technical for clients”), AI refined. Six iterations to reach the current form.
“This framework was created using the process it describes. That's not a coincidence — it's validation.”
HAI emerged from a real development session on the BOB Platform. The creation process itself demonstrates the human-AI collaboration it describes. The human provided direction, judgment, and values. The AI researched, synthesized, iterated, and documented. Every decision was recorded.
- ✓Specification quality is the strongest predictor of delivery success. Invest in specs before production.
- ✓AI-led work with high spec quality is the most efficient mode. But getting to high spec quality requires human judgment.
- ✓Retrospectives compound. The team got measurably better at estimating size and spec quality over the first month of tracking.
- ✓Work mode drift is real and healthy. Features that started as collaborative shifted to AI-led as patterns were established.
- ✓Integration work always surprises. External API dependencies capped confidence at medium until tested, regardless of spec quality.
- ✓The framework works for solo practitioners directing AI agent fleets. In fact, it’s arguably more important when working alone.
Explore the BOB Platform
Architecture, API reference, channels, and deployment guides.
Ready to track your own work with HAI?
Start with the Quick Start guide — five steps, fifteen minutes.