Roger Dali — AI Architect · Workshop Lucerne

§ workshop / whats_running

Three systems,
in production use.

Notes from a live workshop journal. No pitches, no roadmap slides — just what is currently drawing power, with versions, phases and metrics.

PID 01 / romanwerk

builds/30d112

Romanwerk.

Multi-agent pipeline · autonomous long-form production
note: the novels are written in German.

9 phasesquality gates4 cloud models

9-phase architecture with 15 specialised agent roles across 4 cloud models (Kimi · MiniMax · GLM · Qwen3-Next 80B). Quality gates and self-repair between every phase. Two novels produced autonomously — 50 chapters, more than 400 generated pipeline artefacts. The novels themselves are written in German.

→ view_postmortem.sh

agent topology (romanwerk v4.5.0 / Book_TemplateTest)

 ┌─ director ─────────────────────────────────────────────┐
 │                                                         │
 │   P1  strategic-planner    (kimi-k2.5)                 │
 │   P2  scene-planner        (kimi-k2.5)                 │
 │   P3  writer               (minimax-m2.7)              │
 │   P3B parallel-critique    × 3  (slop · structure · voice)
 │                                                         │
 │   P4  reviewer             (glm-5.1)                   │
 │   P4A special-auditors     × 5  (qwen3-next:80b)       │
 │        ├─ system           ├─ numeric-timeline         │
 │        ├─ cast-agency      ├─ narrative-drift          │
 │        └─ cosmos                                        │
 │   P4B voice-reviewer       (glm-5.1)                   │
 │                                                         │
 │   P5  repair-writer        (minimax-m2.7)              │
 │   P6A archivist-merge      (minimax-m2.7)              │
 │   P6B archivist-updates    (minimax-m2.7)              │
 │   P7  compactor            (glm-5.1)                   │
 │                                                         │
 │   quality-gate ──▶ self-repair ──▶ next-phase          │
 │                                                         │
 └─────────────────────────────────────────────────────────┘
          │
          ▼
 pipeline: 9 phases · 15 roles · 4 cloud models
 ✓ Lena_Schatten     25 chapters · 150 artefacts  (no human)
 ✓ Ankerbrecher      25 chapters · 267 artefacts  (no human)

phaseP1 · strategic-planner847 tok/s

9

Phases P1–P7

15

Agent roles

4

Cloud models

50

Chapters autonomous

PID 02 / gemma4bolmor

builds/30d38

Gemma4Bolmor.

LoRA stack · 7 specialised adapters · Ollama Cloud

● servingbase · gemma4-31bvisiontoolsthinking

7 specialised LoRA adapters on a shared base model, trained on curated style pairs. Target metric: output is indistinguishable from human writer references in blind reviews (val-loss < 0.25, Turing-style A/B at n=40). No generic LLM German — a trained voice, verifiable.

↓ Download model on Ollama

→ view_postmortem.sh

$ bolmor-train --stack all --report lora base pairs epochs val-loss status ───────────────── ─────── ────── ────── ──────── ────── writer-s1 gemma4:31b 180 4 0.213 ✓ writer-s2 gemma4:31b 142 4 0.198 ✓ planner-p1p2 gemma4:31b 164 3 0.241 ✓ reviewer-p4 gemma4:31b 98 6 0.267 ✓ voice-p4b gemma4:31b 231 5 0.171 ✓ archivist-p6 gemma4:31b 112 4 0.254 ✓ constraint-grammar gemma4:31b 128 3 0.234 ✓ ───────────────── ─────── ────── ────── ──────── ────── total · 7 LoRAs 1 055 mean 0.225 all green scaling: 4× more pairs queued · base swap → gemma4-omni-31b deploy: ollama cloud · ch-zurich-1 · cold-swap < 800ms

7

LoRAs

1 055

Curated pairs

4×

Scaling

0.21

Mean val-loss

PID 03 / dungeon-forge

builds/30d61

Dungeon Forge.

Autonomous agent swarm · Steam-ready in 10 days

● running3-lane FSM11 agentsself-iterating

11 agents in a 3-lane FSM: Director v2, SLBB budget broker, rate guard, regression gate. 10 days from empty repo to Steam-ready chassis — without human iteration.

→ view_postmortem.sh

dungeon-forge / covergame artwork · concept

agent topology (dungeon-forge v2.1.1)

 ┌─ director ─────────────────────────────────────┐
 │                                                 │
 │   ┌─ world-gen      ┌─ combat-sim              │
 │   ├─ quest-writer   ├─ balance-tuner           │
 │   ├─ lore-keeper    ├─ ai-opponent             │
 │   └─ asset-namer    └─ telemetry               │
 │                                                 │
 │   qa-critic ──▶ repair ──▶ release             │
 │                                                 │
 └─────────────────────────────────────────────────┘
          │
          ▼
 iter-loop(v2.1.1): 8 modules · ~2 400 LOC
 ✓ build_0371  green · 00:47:12  (no human)
 ✓ build_0372  green · 00:51:04  (no human)

phaseworld-gen · L847 ops/s

11

Agents

8

Modules

~2 400

LOC

10 days

Repo → Steam

§ verticals / applications

Two initiatives
for Swiss SMEs.

Markets that structurally lack marketing and digital capacity. AI copilots do not replace a specialist — they make the gap visible.

Gipfel AI

swiss_tourism_copilot

design partner acquisition

Languages	DE · FR · IT · EN · CN · JP
Core flow	Photo → content, review management, guest communication
Market	~38,000 CH tourism businesses
ICP	90% SMEs without a marketing department
Partner	HSLU ITW Lucerne · in progress
Stack	bolmor-v04 · tourism-dach + apac LoRAs

addressable businesses38,000

Fiskal AI

trustee_copilot

pain validation

Target group	5,000 – 8,000 offices
Associations	TREUHAND\|SUISSE · FIDUCIARY\|SUISSE · EXPERTsuisse
Core flow	Receipt → booking · VAT statement
Cantons	26 dedicated models — a tax ruling in Zug does not look like one in Geneva
Compliance	Hallucinations are hard-blocked — output is legally verifiable, professional secrecy preserved
Partner	Early talks with trustee offices in Zug & Lucerne
Stack	bolmor-v04 · trustee-core + vat + 26 ct

addressable offices~6,500

§ bio / roger.dali

From electrician
to AI architect.

roger.dali — 2026 · b&w

Practitioner. I build systems before I talk about them.

Trained electrician. Joined a solar company in its first month and helped build it from the ground up — years on projects sharpened my eye for numbers that have to work: efficiencies, payback periods, capacity. The move into AI I taught myself: no degree, no bootcamp — documentation, experiments, running systems. Same discipline as the job as solar project lead: nothing ships that does not beat its metric.

What stays outside the workshop: long walks across Alpine ridges, a surfboard on the roof rack, photographs without intent — and a deep connection to Central Switzerland that does not export.

Multi-LLM orchestrationLoRA fine-tuningAgent pipelinesConstraint decodingCH SME salesPhotoTravelOutdoor

§ notes / changelog

Workshop notes.

Short build reports from ongoing projects. Written on the day the lesson got expensive.

workshop/notes/*.md3 entries

2026-04-11

LoRA

Why I train one LoRA per canton.

12 min

26 adapters, one base model. What can be learned from Switzerland's tax heterogeneity — for other verticals too.

2026-03-29

Decoding

Constraint decoding against hallucinations.

9 min

A field report: how hard grammar constraints in the decoder keep bookkeeping outputs under control.

2026-03-14

Quality

Pipeline compliance is not reader satisfaction.

14 min

What the Bolmor curator system taught me about the gap between 'pipeline validated' and 'book works' — and why I had to invent a new category of Hard Rule for it.

I build AI systemsthat work.

Three systems,in production use.

Romanwerk.

Gemma4Bolmor.

Dungeon Forge.

Two initiativesfor Swiss SMEs.

Gipfel AI

Fiskal AI

From electricianto AI architect.

Workshop notes.

Why I train one LoRA per canton.

Constraint decoding against hallucinations.

Pipeline compliance is not reader satisfaction.

I build AI systems
that work.

Three systems,
in production use.

Two initiatives
for Swiss SMEs.

From electrician
to AI architect.