How an Experimental Mindset Built Enterprise AI in 2.5 Years
Industry: Traditional Manufacturing/Industrial Distribution Challenge: Build AI capability from scratch with zero foundation Approach: Improvisational experimentation → Pattern recognition → Foundation building Timeline: 2.5 years | Army of 1 → 3-person team Results: Production AI systems, modernized infrastructure, validated roadmap (doubling headcount)
The Starting Point
The mandate: “Figure out AI.”
That was the directive I received 2.5 years ago when joining a traditional industry company’s “Innovations Department”—a 12-person digital/e-commerce team that had built the first-in-industry online store for transmission parts.
The reality:
- Zero AI/ML capability
- On-premise infrastructure (no cloud)
- Legacy stored procedures written by one person, no documentation
- No team, no roadmap, no playbook
- Army of 1
My background:
- AI/ML infrastructure at Wikipedia (content moderation, tooling)
- BlackCrow AI startup (bespoke ML models for D2C brands, predictive analytics)
- AI/ML work since 2016
- Experimental music background (improvisational thinking, pattern recognition)
The challenge: How do you build enterprise AI capability in a traditional industry with limited resources, no technical foundation, and organizational skepticism?
Year 1: Skunkworks & Learning
The approach: Try things. See what sticks.
I didn’t have a perfect plan. I had experience, pattern recognition from previous work, and a willingness to experiment.
POC #1: Intelligent Contact Center
- Built post-call analytics system for sales calls
- Sentiment analysis, entity recognition, topic modeling
- Result: Cool tech, clear value proposition
- Problem: Organization wasn’t ready to maintain it
POC #2: Voice Agent → Chatbot
- Voice agent handling price & availability questions
- Eventually paired down to chatbot
- Demo’d at AAPEX 2024 (industry conference)
- Result: External validation, executive excitement
- Problem: No internal capacity to continue development
The pattern I recognized:
We could build AI. We couldn’t sustain it.
Technical capability ≠ Organizational readiness.
The insight that changed everything:
Stop building flashy POCs. Start building the foundation.
Year 2: Building the Foundation (Still Army of 1)
The pivot: Infrastructure before innovation.
Challenge #1: On-Prem to Cloud Migration
- Legacy on-prem infrastructure blocked modern cloud-based AI tools
- Network issues, security constraints, limited flexibility
- Built AWS cloud infrastructure from scratch
- Enabled access to modern AI/ML tooling
Challenge #2: Legacy Code Debt
- SQL stored procedures written by one person (no docs, no one understood them)
- Business logic tangled in unmaintainable code
- Blocking data access for AI/ML projects
The solution: Use AI to build AI foundations.
- Deployed Claude Code to understand legacy code
- Automated documentation generation
- Risk assessment and test design
- Added observability to critical systems
- Meta-lesson: We used AI coding tools to modernize the platform that would enable AI products
Result:
- Data now accessible for AI/ML projects
- Modern cloud infrastructure in place
- Technical debt systematically addressed
- Foundation ready for scaling
Year 2.5: First Production Wins & Team Building
With infrastructure in place, production AI became viable.
Win #1: Diagram-Mapper (The Zombie Feature)
- Recovered lost capability from dead vendor (6 years broken)
- 1 engineer using Cline + Claude Sonnet 3.5
- 2 quarters to production
- +30% coverage increase, feature went from hidden to sales tool
- See separate case study for details
Win #2: Hybrid Search
- Reduced off-topic search results by 40-60% with semantic filtering
- Ported 2K LOC SQL procedure to Python service in 1 week
- Improved catalog discoverability
Win #3: Platform Modernization
- Untangling business logic from legacy stored procedures
- Upgrading catalog backend to modern stack
- Enabling faster development cycles
First hires (after 2 years solo):
- Python engineer (AI-native, expert with agentic coding tools)
- Data/ML engineer (from traditional industry)
- Hiring strategy: Traditional industry background + modern AI skills
Team growth: 1 → 3 people in 2.5 years
Current State: Validated & Scaling
What exists now that didn’t 2.5 years ago:
- ✅ Cloud infrastructure (AWS) vs. on-prem constraints
- ✅ Modern catalog backend vs. legacy SQL spaghetti
- ✅ 3-person AI team vs. army of 1
- ✅ Production AI systems in customer-facing products
- ✅ Product team able to maintain AI features independently
- ✅ Documented, observable, testable codebase
- ✅ Validated roadmap (company doubling Innovations Dept headcount next year)
Business impact:
- Features that were impossible are now production-ready
- Development velocity increased (modern stack vs. legacy constraints)
- Organization ready to maintain and expand AI capabilities
- Executive buy-in proven (doubling team size investment)
The validation:
When a company doubles headcount on your team, you’ve proven the model works.
Why This Approach Worked
1. Experience + Experimentation I wasn’t learning AI for the first time—I’d built AI/ML systems since 2016. But I approached traditional industry with an experimental mindset: try things, recognize patterns, adapt.
2. Pattern Recognition POCs revealed organizational readiness issues. Instead of pushing harder on flashy AI, I recognized the pattern: infrastructure and processes must come first.
3. Foundation Before Scale 2 years of solo work building infrastructure wasn’t wasted time—it was the necessary foundation. Trying to hire a team Year 1 would have failed.
4. AI to Build AI Using Claude Code and agentic coding tools to modernize legacy systems wasn’t just pragmatic—it proved the value of AI to skeptical stakeholders. “We use the tools we recommend.”
5. Traditional Industry Hiring Hiring engineers from traditional industries (not just FAANG refugees) meant they understood the constraints, pace, and organizational realities.
6. Improvisational Adaptability My experimental music background taught me: you don’t know where you’ll end up until you start. You recognize patterns. You adapt. You don’t force a predetermined playbook.
The Lessons for Traditional Companies
If you’re starting from zero:
Don’t:
- ❌ Hire a big team on Day 1 (you don’t know what you need yet)
- ❌ Build POCs without organizational readiness (they’ll die on the vine)
- ❌ Wait for perfect conditions (you’ll never start)
- ❌ Copy FAANG playbooks (different constraints, different solutions)
Do:
- ✅ Start with one experienced person who can figure it out
- ✅ Try things, learn, adapt (experimental mindset)
- ✅ Build infrastructure before scaling team
- ✅ Use AI to solve your own problems first (prove the value)
- ✅ Hire for traditional industry experience + modern AI skills
- ✅ Prove value before asking for big investments
Timeline expectations:
- Year 1: POCs, learning organizational constraints
- Year 2: Infrastructure, foundation, first production wins
- Year 2.5+: Scale team, expand capabilities
ROI:
- One experienced person + 2.5 years = production AI capability
- Alternative: Hire 5 people on Day 1, watch them fail for lack of foundation
- Or: Hire Big 4 consulting firm, get theoretical roadmap but no execution
Want to Discuss Your AI Journey?
I help mid-market traditional companies build AI capabilities through readiness assessments, vendor selection, and team building coaching.
Contact me via LinkedIn to discuss your specific situation.