Ciro Canton / AI portfolio
AI systems, built around real work.
I design AI products, agent workflows, and automations that turn complex work into clear next steps.
Selected work
From a working problem to a working system.
Three systems for real work: student briefs,film planning, and format-specific production automation.
OpenClass
Query → Route → Agent
Typed decisions. Clear handoffs.
A multi-agent, open-source briefing workflow for students that connects class deadlines, progress evidence, rubric feedback, and curated recommendations through typed judgments.
View project
Slatework
Script → Draft → Review
Read the breakdown beside the source.
A film-producer workspace that turns screenplay passages into source-linked draft requirements, production efficiencies, and a cross-scene continuity ledger.
View project
Production automation
Batch → Transform → Export
From incoming files to format-specific exports.
A workflow system for an Argentinean social media production company, connecting batch file intake, image transformations, format routing, and controlled exports.
View project
Slatework
Read the draft beside the source.
Process screenplay passages into source-linked draft requirements, production efficiencies, and a cross-scene continuity ledger for easier department handoff. For micro-studios, solo producers and guerilla-filmmakers.
AI proposes producing tips, passage validation keeps the source visible, and the filmmaker decides what belongs in the plan.
SLATEWORK / REVIEW FLOW
01
Source passage
Local screenplay parsing
02
AI draft
Consent-based interpretation
03
Passage validation
Check against saved evidence
04
Human review
Correct, confirm, or question
05
Export
JSON · CSV · print
Workflow schematic · AI output remains a draft
Local parsing. Open Fountain, guarded FDX, and supported text PDFs in the browser.
BYOK consent first: bring your own inference API credentials for extra analysis; keys stay transient.
Source validation. Draft findings must point back to saved screenplay passages.
Requirements, scene review, and group continuity occurrences can be human-reviewed and edited.
Portable work. Keeps revisions in IndexedDB and exports in JSON, CSV, or print.
Currently developing paid agentic tiers to autonomously cast talent, find suitable locations based on the filmmaker's production hub, and automatic cloud-sync.
Production automation
One intake & different delivery rules.
For an Argentinean social media production company, I built a workflow web that route-batches by format, transforms images according to company parameters, and hand files off for download and email logging.
PRODUCTION / WORKFLOW MAP
01
Forms + Sheets
File intake and processing state
02
Tag + aggregate
Cloudinary uploads and batch handoff
03
Route by format
Five (5) dispatched format workflows
04
Deterministic transformations (based on file/format dimensions)
Crop / Gen-fill / Masks / Overlays
05
Export + notify
Drive files and email handoffs
Implementation snapshot · inactive when inspected
Batch intake. Iterate submitted files, track response-and-file keys, and aggregate uploads.
Multiple formats routing: Banners, 1:1 Aspect Ratio, Marquee, Arches, and Billboard workflows .
Image processing: cropping, upscale, generative-fill, masking & overlays. Transformations are deterministic-based (workflow reads file dimensions of ingested files and compares against company guidelines)
OpenClass
A semester becomes a coordinated workflow.
A decision model supplies typed judgement through an MCP server. A Delegator and specialized agents turn those judgments into plans, artifacts, and a brief a student can review everyday.
Implemented and evaluated: Jev MCP tools and routing.
Target workflow: unattended daily triggers and real course ingestion.
Jev advises; the calling agent performs each handoff after task completion
About / approach
Build around the decision.
I’m Ciro Canton. My work spans agent orchestration, AI product design, and production automation. I focus on what a system can verify, where people need control, and how the output will be used.
Keep the evidence close
A finding should lead back to a source, a test, or a clearly stated assumption.
Make human control usable. Always.
Review, correction, and escalation belong in the workflow people actually use. Human review & guardrails are non-negotiable.
Deliver a useful artifact.
A brief, a producer workspace, or an automation should help someone take the next step; not add more headaches.
Have a workflow worth rethinking?
Let’s talk about the people, decisions, and output your next AI system needs to support.