Each case study leads with the result. Client names appear only where the work is public, and confidential engagements stay anonymous.
AI Workflow Pilot / Implementation Sprint
4 days → 4 hours
A core proposal workflow, 75–80% automated
A California EdTech company: an enterprise proposal system
Before4 days
After4 hours
A California EdTech company wrote enterprise training proposals and course outlines by hand, from scratch, every time. Over a six-week engagement in March and April 2026, Daria Gavrilova, brought in as AI Content Strategist, designed and architected a GitHub-backed application on the Claude API that generates personalised learning packages. A person reviews every output before it goes out. She then led the technical handoff to the developer and the stakeholder who would own the system.
AI Workflow Pilot / Implementation Sprint
29.4% → 48%
Patient adoption of a WhatsApp booking bot, month one to month two
A London dental clinic: a WhatsApp booking and intake bot
Adoption
Month 129.4%
Month 248%
Conversation to booking
Month 145.3%
Month 254%
Handled without staff
Month 156.9%
Month 268%
Daria Gavrilova co-led the rollout with a technical partner, owning the product decisions and the UK GDPR design: pseudonymisation, webhook security and the retention and deletion policy. Month-one adoption was lower than hoped, so the intake flow was tightened to cut question fatigue, reception promoted the channel more actively, and the link and QR code were made easier to find. The results were measured, not assumed.
AI Opportunity Diagnostic
300+ students
Reached by AI-integrated courses built after a readiness assessment
UAB Study Abroad: assess first, then design
- Assess readiness
- Prioritised roadmap
- Design courses
- Measure and iterate
Before designing a single course, Daria Gavrilova assessed the programme's actual AI readiness across curriculum, faculty capability and governance. The findings became an AI-integration roadmap with prioritised use cases, sequencing and expected learning outcomes. The courses that followed, International Marketing Strategies and Career Management with AI, reached more than 300 international students, with assessment frameworks tying AI theory to measurable professional outcomes.
AI Adoption Workshop
6 of 6 built a prototype
Cohort rating: 4.33 out of 5 for usefulness and for confidence to build with AI
EAE Business School: a vibe-coding masterclass
Useful4.33 / 5
Confidence4.33 / 5
EAE invited Daria Gavrilova to modernise a Global Master's residential week with AI. She designed and delivered a half-day, hands-on masterclass, "Vibe-Coding: building a landing page with AI", taking non-technical students from a one-line brief to a working prototype. Every student in the cohort of six built their own. The second rating measures what an adoption programme is for: whether people now believe they can build with AI.
Own system · AI adoption in practice
7 agents · 7 skills
One shared context, four version-controlled repositories, 100+ commits in about three and a half months
A multi-agent operating system for running HCD Barcelona
90%of video clips accepted first time
0unapproved sends: publishing scripts refuse anything not marked approved
30/30test tasks routed to the right agent in a blind test
- Writing
- Podcast production
- Business development
- Consulting
- Coordination
- and more
- Write the spec
- Direct the build
- Verify against the spec
Daria Gavrilova built this for her own work, not for a client, which makes it proof of hands-on range rather than advisory talk. Role-specific agents are wired into real workflows and share a context architecture that keeps consulting, podcast, writing and newsletter work separate, with a second brain that maintains its own daily notes and project state. The method is the same throughout: write the specification, direct Claude Code to build against it, and check the result against the specification before calling anything done. That is how a first build of the writing system was rejected: it looked right but still read as AI-written, so the generation rules were tightened. No before-and-after outcome is claimed here. The figures above describe how reliably the system behaves, not business results.
Own project · AI Workflow Pilot in miniature
343 MB → 87 MB
A model moved into the browser when the planned hosting needed a paid tier
A Barcelona recycling-sorting app
Original343 MB
Quantised87 MB
A fine-tuned Vision Transformer sorts a photo of an item into the right Barcelona recycling bin, with local edge cases built in and an honest "not sure" answer whenever confidence drops below 55%. When the planned hosting route turned out to need a paid tier, the model was exported to ONNX, quantised and moved entirely into the browser: no server, no hosting cost, and photos never leave the device. A side-by-side check on a test image showed the confidence within about 0.2 percentage points of the original. Tests on deliberately messy photos exposed a limitation, which was recorded rather than hidden. Try the app
