Full-Stack AI Operator
3 Years, 1,000+ Hours, Zero Lines Written by Hand
One non-developer with deep energy-sector domain expertise built a full-stack business operation — three live personal products, two enterprise analytics platforms, 30+ automated intelligence jobs, and a self-managing content engine — entirely by directing AI agents rather than writing code.
The Journey — 4 Phases
From interrogative to directive. Each phase marks a capability shift, not just a platform change.
Treating AI as a smarter Google — brief, one-sentence queries with no context or framing. Zero use of personae or structured prompting.
Style: Interrogative
"tell me about today in history"
The starting point. AI as factual oracle, not co-worker.
Sessions: 45 files
Emergence of role-setting and context-heavy prompts. Professional background provided to narrow output. Perplexity becomes the primary research interface.
Style: Structured
"Acting as a career guide I need you to provide options for a career path with the experience detailed below; UK energy expert [...]"
From information lookup to professional planning and decision support.
Sessions: 545+
Large volume of work migrated to Abacus DeepAgent. Stopped asking questions, started giving instructions. Marks the move from consumer of AI responses to producer of AI-built output.
Style: Command-based
Transition from "what is" to "build this." Up to 66 messages per session — collaborative building, not lookup.
The chasm crossing — from AI consumer to AI producer.
Sessions: 200+
Managing the system from the outside. Directives to autonomous agents. Context pre-loaded via .md files — minimal inline explanation needed. Reviews backlogs and sprint boards with the AI.
Style: Directive
"look at our latest backlog file [...] let me know what are the big ticket items on there."
From builder to orchestrator. The agent is treated as a senior employee.
Sessions: Growing
Hours in the Chair
1,031 hours of active AI operator work across 8 platforms. Not passive consumption — active direction.
1,031 hours
Average session 120 minutes (user-calibrated). Some enterprise sessions ran 4+ hours.
What Has Been Built
Personal products on the left, enterprise delivery on the right. Everything built to be productised and maintained.
Picking Solutions
151 published posts across energy, digital assets, and AI. 5-stage content flywheel. 31 automated daily intelligence jobs. 9 Hermes AI agent profiles.
ProstEnergia
Polish household energy price comparison platform. 63 suppliers, 745 tariffs. Live and maintained by AI agents with human oversight.
HomeEnergyForge
Standalone UK homeowner energy tools site. Lift-and-shift from pickingsolutions.tech calculator tools.
vs 290 days original estimate
0 written by hand
3,328 source tables
40+ dashboard pages
Full agent swarm
Append-only log
Zero warm-up sessions
119 total sessions
Skills Demonstrated Through Building
Not a list of buzzwords. Every skill traces to something documented.
AI Agent Orchestration
9-profile Hermes stack (personal); 7-agent Snowflake Cortex swarm (enterprise)
Enterprise Data Engineering
44.7B rows, 3,328 tables, 3 databases. Signal-first architecture
Prompt Architecture
Anti-fabrication protocols; context pre-loading via .md system files; 42 VQRs in Cortex semantic views
System Design (non-code)
31 automated jobs; 5-stage content pipeline; dual-output architecture; knowledge synthesis system
Session/Context Management
119 consecutive sessions with zero warm-up via handover-as-start-prompt pattern
Data Pipeline Design
Daily intelligence pipeline (27 RSS → vault → wiki → briefs → publish); settlement assurance pipeline
Domain Expertise (Energy)
Wholesale cost mechanics, settlement assurance, margin decomposition, DUoS timeband structure, TNUoS resets
Product Management
5 concurrent products; 346+ architecture decisions; North Star scoring; productised from Session 1
Technical Documentation
USER_GUIDE, TROUBLESHOOTING, RECONCILIATION_DESIGN, 8 agent docs — production-grade
Content Strategy
151 published posts; three-stream content system; autonomous review and approval pipeline
The Operational Method
What separates 'used AI a lot' from 'built a professional system.' Consistent across all 5 build streams.
Mandatory Session Open
GUARDRAILS → BACKLOG → previous handover. Every session. No exceptions.
The reading order matters. GUARDRAILS first (what not to break), then BACKLOG (what to build), then handover (where you left off).
GUARDRAILS as Living Spec
Updated in real time — every rule traces to a specific incident.
Grew from 6 sections to 19 across 68 sessions. Not aspirational — descriptive. Every failure became a structural prevention.
Append-Only Decisions
346+ entries, status-flagged, nothing deleted.
APPROVED / REVERSED / SUPERSEDED / DEPRECATED. The evolution is visible. The log is an audit trail of judgement, not just a list of outputs.
Handover-as-Start-Prompt
Every session close produces a verbatim copy-paste start prompt for the next session.
The single most valuable innovation. Zero warm-up across 119 consecutive sessions. Full context compressed into one document.
SQL Reproducibility
DDL + CTAS saved same session. Introduced after Session 11 loss incident.
Every structural change is preserved at creation time. Prevents repeated rework from context loss across sessions.
The Decisive Variable
In every fork across 119 sessions: domain expertise, not AI capability, was the decisive input.
The AI executed. The operator decided what the AI needed to know. The domain knowledge is what made the intelligence layer intelligent. The gap between "the AI built this" and "the operator built this using AI" is visible in the decision log — 346 entries, each a point where the system would have continued without intervention and produced a result that was correct by the model but wrong for the market.
Audit Findings — What the Data Shows
Quantitative findings from the post-project deep audit of 119 sessions and 59 product documentation files.
Handover length grew 2–3× over 68 sessions
From ~110–160 lines (Sessions 1–3) to ~190–460 lines (Sessions 63–67). Growth reflects accumulated process maturity, not context flooding — no session had to restart due to lost context.
Zero hallucination-forced resets
Across all 119 documented sessions, the GUARDRAILS + handover system produced zero resets caused by agent hallucination. The constraint structure, not model quality, is the explanatory variable.
14% stretch target delivery rate
~7 of ~50 aspirational stories delivered. This is deliberate deferral, not failure — most were scoped out by the North Star filter applied mid-project.
~90% story delivery rate across 4 streams
Approximately 320+ DONE from 356+ attempted. The 10% gap is almost entirely external blockers (admin grants, workspace permissions) rather than scope failure.
What's Next
Active development direction and identified gaps.
HF Agents Course
IN PROGRESSHugging Face — 6 units covering tools, memory, multi-agent patterns, code agents, vision agents.
GitHub Certified: Agentic AI Developer (GH-600)
TARGET — Week 17+Beta certification. All 6 exam domains map directly to evidence from the Snowflake project — architecture, tool use, memory management, evaluation, multi-agent coordination, guardrails.