Roger Dali — Industry-specific AI copilots for Swiss SMEs

§ research & autonomous systems / live systems that prove the methodology

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.

research / romanwerk

builds/30d112

Romanwerk.

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

14 phasesquality gates4 cloud models

14-phase architecture with ~20 specialised agent roles across 4 cloud models (Kimi · MiniMax · GLM · Qwen3-Next 80B). Quality gates and self-repair between every phase. Four novels produced autonomously — 100 chapters, more than 900 generated pipeline artefacts. The novels themselves are written in German.

→ view_postmortem.sh

agent topology (romanwerk v5.0 / Book_TemplateTest)

 ┌─ director (kimi-k2.6) ─────────────────────────────────┐
 │                                                         │
 │   P0.5 cast-warden          (kimi-k2.6)                │
 │                                                         │
 │   P1a  creative-planner     (minimax-m2.7)             │
 │   P1b  rule-validator       (kimi-k2.6)                │
 │   P2   scene-planner        (kimi-k2.6)                │
 │                                                         │
 │   P3   writer               (glm-5.1)                  │
 │   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)                  │
 │   P4C  world-bible-audit × 6  (parallel, non-blocking) │
 │                                                         │
 │   P5   repair-writer        (qwen3-next:80b)           │
 │   P6A  archivist-merge      (kimi-k2.6)                │
 │   P6B  archivist-updates    (kimi-k2.6)                │
 │   P7   compactor            (glm-5.1)                  │
 │   P7A  kurator              (qwen3-next:80b)           │
 │   P8   exporter             (kimi-k2.6)                │
 │   P9   lektor               (glm-5.1)                  │
 │   P10  final-reader         (qwen3-next:80b)           │
 │                                                         │
 │   quality-gate ──▶ self-repair ──▶ next-phase          │
 │                                                         │
 └─────────────────────────────────────────────────────────┘
          │
          ▼
 pipeline: 14 phases · ~20 roles · 4 cloud models
 ✓ Lena_Schatten         25 chapters · 150 artefacts   (no human)
 ✓ Ankerbrecher          25 chapters · 267 artefacts   (no human)
 ✓ Leo_Gibt_Nicht_auf    25 chapters · 282 artefacts   (no human)
 ✓ Tickender_Herzschlag  25 chapters · 236 artefacts   (no human)
 ▸ Torvin_Kasch          in production

phaseP0.5 · cast-warden16.6 tok/s

14

Phases

~20

Agent roles

4

Cloud models

100

Chapters autonomous

research / 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

research / dungeon-forge

builds/30d61

Dungeon Forge.

Autonomous agent swarm · Steam-ready in 10 days

● running7 phasesvertical backtrack

Eleven roles, one pipeline: nothing enters the codebase that hasn't passed architect (contracts) → coder (implementation) → reviewer (spec check) → qa (boot test) → integration (smoke). Specialists — game logic, game feel, level design, shaders — are consulted in parallel; their rulings are appended to the design doc, never silently merged. Producer holds scope override. This is the pipeline that replaced the V1 retry-loop.

→ 14-agent topology · live phase view

dungeon-forge / covergame artwork · concept

Live

agent topology (dungeon-forge v2.3.0 / Dungeon Forge)

                     ┌──────────────────┐
                     │     PRODUCER     │   scope · overrides
                     └────────┬─────────┘
                              │
                     ┌────────▼─────────┐
                     │     DIRECTOR     │   decomposes · dispatches · backtracks
                     └────────┬─────────┘
                              │
    ┌──────────────────────────┴──────────────────────────┐
    │                 build pipeline                       │
    │                                                      │
    │   architect ─► coder ─► reviewer ─► qa ─► integr.   │
    │   contracts   impl.    spec-chk    boot   smoke     │
    │                                                      │
    └──────────────────────────┬──────────────────────────┘
                               │
                    ┌──────────▼──────────┐
                    │  PHASE-END GATES    │   hard · veto-capable
                    │                     │
                    │  architectural_     │   CANONICAL.md diff
                    │  auditor            │   semantic drift detect
                    │                     │
                    │  reality_check      │   browser runtime probe
                    │                     │   console · render · fps
                    └──────────┬──────────┘
                               │
                        ┌──────┴──────┐
                    PASS│             │FAIL
                        ▼             ▼
                   next phase    vertical backtrack
                                 │
                                 ├─► rollback to phase N-1
                                 ├─► contract revision
                                 ├─► SLBB arm penalty
                                 └─► re-dispatch
                               │
                        consultation bus (director-brokered, MCP)
                               │
    ┌─────────────┬────────────┼────────────┬──────────────┐
    │ game logic  │ game feel  │ level des. │ shader artist│
    │ rulings     │ tuning     │ encounters │ materials    │
    └─────────────┴────────────┴────────────┴──────────────┘
                               ▲
                               │ reward signals
                      ┌────────┴─────────┐
                      │   SLBB broker    │   phase-scoped
                      │                  │   reality-weighted
                      └──────────────────┘

phaseproducer · scope67 tok/s

tok/s = output · single-shot · n/a = deterministic, no llm · gates: hard + veto-capable

—·Pending·—

code: private · architecture: public → Dungeon Forge postmortem

14 agents · 7 phases · 2 phase-end gates

4 cloud models + 2 deterministic tools

vertical backtrack · phase-scoped SLBB reward

~2 400 LOC · 10 days repo → steam

§ verticals / industry-specific AI copilots in development

What I'm working on
for Swiss SMEs.

Markets that structurally lack marketing and digital capacity. AI copilots don't replace specialists — they cover the capacity a specialist can't.

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

Gipfel AI — AI copilot for Swiss tourismGuest communication, content, reviews. Swiss-hosting, nFADP-compliant. Research partnership with HSLU. Currently in pilot acquisition.Product page →

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

Until the products launch, I work with clients directly.

Freelance & Workshops

freelance · workshops

available now

AI Readiness Audit	from CHF 2,500 (1 day)
AI Pilot Project	CHF 8,000–15,000 (2–4 weeks)
Architecture Consulting	CHF 250/h
Workshop “AI for decision makers”	CHF 490 (half day)
Workshop “Building AI agents”	CHF 1,490 (2 days)

Send inquiry →

§ 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 I brought to solar project management: nothing ships that doesn't beat its metric.

Outside the workshop: long walks across Alpine ridges, a surfboard on the roof rack, photography — and roots in Central Switzerland that don't travel well.

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 landed.

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.

AI systems that work.

Three systems,in production use.