Claudio Bustamante

Problem. Infrastructure engineers at a major Spanish bank write change requests for the bank’s production ITSM platform. Requests that break the bank’s rules are rejected, and an assistant in a bank cannot invent data or take an unpredictable path.

What I did. I am the core developer of a multi-agent LLM assistant that helps these engineers write their change requests. I built it from an empty skeleton, and it now creates changes in the bank’s production ITSM platform.

Key decisions.

• The model reasons; the code controls the flow. Agents run in a fixed sequence defined in code, with no LLM planner.
• Model output passes strict Pydantic schemas and a single write point; anything invalid is discarded.
• The bank’s change-rejection rules became tested checks, so users see what would be rejected while they are still filling in the form.
• An LLM judge reviews text quality, with a filter that drops any suggestion containing invented data.
• Long first messages go to a larger model that extracts every field at once; a regex-based detector decides when.

Result. Changes created in production, and 900+ tests that run without network, LLM or databases, in a component that started with none. I also ran demos for operations and management and wrote the handover docs a colleague used to run the component while I was away.

Problem. An APEC aquaculture project needed a mobile tool to help diagnose eight diseases of tilapia and trout. The knowledge lived in an 82-page guidebook. Hand-coding the rules would be slow, and an LLM answering in the field could not be audited.

What I did. I built a Unity editor pipeline in which the LLM only proposes rule candidates from the PDF. Ten deterministic stages then validate, auto-fix, canonicalize, deduplicate, gate, selectively retry (best-of-N), dry-run and apply them. The Android app runs the resulting rules in a diagnostic engine on the device, with no LLM at runtime. AR is used separately, for standardized photo guides and training.

Key decisions.

• The LLM proposes; the pipeline decides.
• Content that can’t be extracted exactly is discarded and logged, never approximated. Auto-fix applies only when the exact value is in the source text.
• A semantic rule key catches duplicates; a quality gate blocks the run on parser errors, missing packs or too much auto-filling.
• Every rule keeps its source page.

Result. 101 typed rules in 10 packs, each traceable to a page of the guidebook. The gates blocked 8 of 22 batch runs. The deduplication stage removed 21 duplicates before parsing and cut parser warnings from 261 to 180. Fidelity was checked with an LLM-assisted, rule-by-rule audit against the guidebook and 18 golden cases; there are no precision/recall metrics yet.

Problem. Game NPCs usually choose between hand-written dialogue, which runs out, and a live LLM, which forgets, contradicts the story and is hard to debug. I wanted NPCs that react to what this player did without breaking canon.

What I did. I designed and built Eón, a Unity package where each agent has one narrow job. Offline, a seed baker prepares narrative anchors. At runtime, one agent generates branching conversation trees from the player’s history; another writes single replies in free dialogue. Buffer Souls keeps what each NPC believes. All of it runs on a shared, provider-agnostic LLM package.

Key decisions.

• Canon, the objective record and what an NPC believes are separate layers. The prompt stacks them in order of authority, and a belief never overwrites canon.
• NPC state changes through one write path only: validated events, pure bounded deltas, and a read-only projection for prompts.
• Generated trees are parsed and checked against structural rules. A failure is a typed result that names its stage; nothing throws.
• Versioned prompts, a workbench with per-attempt traces and cost, and 100+ tests.

Result. Eón started as research funded by ANID, Chile’s national research agency, and was first tested in a small RPG where it generated a story the player completes by visiting regions and talking to NPCs in order. It now drives the narrative of Alshaat, a tactics game in development at BCS Studios: it weighs each player’s actions and changes how every NPC talks to that player.