104 lines
6.7 KiB
Markdown
104 lines
6.7 KiB
Markdown
# AI Photography Assistant - Hermes Discovery Experiment (Jesslyn's Workflow)
|
||
|
||
**Status**: Experimental / Discovery Project
|
||
**Primary Goal**: Test the effectiveness and limitations of persistent AI agents (via Hermes) in a real creative workflow.
|
||
**Secondary Goal**: Explore whether this can provide practical value to a professional photographer (Jesslyn's business).
|
||
|
||
**Date Started**: 2026-06-17
|
||
**Owner**: Leonard / Tony
|
||
|
||
**Key Reference**: [Odysseus Deep Research Report - The Dawn of Persistent AI Agents in Professional Photography (rp-222259c2ba04)](research/odysseus-report-rp-222259c2ba04.md)
|
||
|
||
This report (accessed via provided credentials) validates our direction, expands on the 8 use cases, and explicitly defines "Hermes-style persistence." New insights have been integrated below.
|
||
|
||
## Core Concept
|
||
A Hermes-based persistent AI partner that learns a photographer’s style, preferences, and business operations. It acts as a creative second brain across pre-shoot, on-set, post-production, and admin.
|
||
|
||
This is **not** primarily about building a product. It is about discovery: What does "persistent" actually mean in practice? Where does it add value? Where does it break? How much human oversight is truly required?
|
||
|
||
## Key Capabilities to Explore
|
||
|
||
| Area | Potential Hermes Role | Persistence Questions to Test |
|
||
|------------------|---------------------------------------------------|-------------------------------|
|
||
| Pre-Shoot | Personalized shot lists, lighting plans from briefs + past work. | Can it recall style from previous shoots without re-prompting? |
|
||
| Post-Production | Culling, tagging, edit suggestions, gallery sequencing. | Does style memory improve suggestion quality over multiple sessions? |
|
||
| Business | Emails, proposals, invoicing, scheduling via Leonard-style automation. | How reliably can admin tasks run with state carried forward? |
|
||
| On-Set | Voice co-pilot suggestions. | Can short-term state + long-term memory combine for useful real-time help? |
|
||
|
||
## Experimental Focus: Persistence (Updated with Odysseus Report Insights)
|
||
|
||
This project will specifically stress-test Hermes persistence mechanisms, aligned with the five pillars of **Hermes-style persistence** identified in the Odysseus report:
|
||
|
||
1. **Cross-Shoot Memory** — Remember actual style decisions from previous jobs (not just static presets). Dynamic memory that evolves (e.g., client-specific preferences like warmer tones for certain portraits).
|
||
|
||
2. **State + Context Across Workflow Stages** — Understand interconnectedness: pre-shoot decisions inform post-production; client feedback informs future culling. Example: Prioritize candid shots based on pre-shoot notes and adjust retouching style accordingly.
|
||
|
||
3. **Sub-Agent Coordination** — "Culling agent talks to retouching agent talks to delivery agent." Seamless information flow and coordinated decision-making across modules.
|
||
|
||
4. **Self-Improvement from Corrections Over Time** — Continuous learning from human overrides and manual adjustments. Aligns with emerging tools like Aftershoot's self-improving profiles.
|
||
|
||
5. **Local-First + Explainable Decisions** — Photographers must understand *why* an AI made a choice. Maintains creative control, supports client explanations, and keeps sensitive data private.
|
||
|
||
**Hermes Capabilities to Leverage**:
|
||
- Artifact-based memory (long-term style bible, client history, decisions)
|
||
- Structured state files for ongoing projects
|
||
- Sub-agent handoff with context passing
|
||
- Self-improvement from human feedback/corrections
|
||
- Cross-session recall (days/weeks apart)
|
||
- Hybrid local (MacBook) + VPS execution
|
||
|
||
**Success will be measured by**:
|
||
- Reduced need to re-explain style/preferences over time
|
||
- Measurable time savings on repetitive tasks
|
||
- Quality of outputs (human judgment)
|
||
- Number and type of interventions required
|
||
- Agent's ability to surface relevant past work without explicit search
|
||
- Demonstrable application of the 5 pillars above
|
||
|
||
**Report Recommendation**: Prioritize **deep workflow mapping** of Jesslyn's actual process *before* building specific agents. Ground development in real needs.
|
||
|
||
## Realistic Assessment (from original proposal + report)
|
||
|
||
**Upsides**:
|
||
- Time savings on repetitive tasks (culling, admin)
|
||
- Deep style learning via persistent memory
|
||
- Strong privacy (local-first)
|
||
- Extensible skills hub
|
||
- Hybrid MacBook + VPS setup
|
||
- Potential for genuine self-improvement and cross-workflow intelligence
|
||
|
||
**Downsides & Limitations** (to be tested):
|
||
- Requires human oversight (especially creative decisions)
|
||
- Slower performance on older hardware for vision tasks
|
||
- Initial setup and tuning effort
|
||
- Not fully autonomous
|
||
- Integration may need debugging
|
||
- Ethical risks (authenticity, bias, copyright, privacy) — must be addressed transparently
|
||
|
||
## Project Principles
|
||
- Discovery first, utility second.
|
||
- Log everything: what worked, what didn't, why.
|
||
- Start narrow, expand only after evidence.
|
||
- Use real workflows (Jesslyn's business) for validity.
|
||
- Keep it grounded — no hype about full replacement of the photographer.
|
||
- Incorporate advanced memory concepts (e.g., Retrieval-Augmented approaches, ACAN-style attention) where they enhance Hermes capabilities.
|
||
|
||
## Next Steps (Discovery-Oriented) — Updated with Report
|
||
|
||
1. **Map Jesslyn’s actual workflow in detail** (pre-shoot → on-set → post → admin). This is the report's #1 strategic recommendation before technical work.
|
||
2. Ingest sample briefs, past shoots, and style references into a persistent profile (start with top candidates: retouching style learning + culling).
|
||
3. Build minimal Phase 1 agent focused on one high-persistence area (e.g., Personal Style Learning for Retouching or Automated Culling, per report).
|
||
4. Explicitly implement/test the 5 Hermes-style persistence pillars.
|
||
5. Run controlled tests across multiple "sessions" (simulated and real).
|
||
6. Document findings in this repo (what persisted, what required re-instruction, surprises, alignment with report insights).
|
||
7. Consider ethical guardrails and explainability from the start.
|
||
|
||
## Related
|
||
- Parent research: Ty_Tech/research
|
||
- Odysseus Report: `research/odysseus-report-rp-222259c2ba04.md` (full integration of new research)
|
||
- Previous use cases baseline: `research/use-cases-qwen.md`
|
||
- Hermes capabilities to leverage: artifacts, memory, state tracking, sub-agents, skills, cron, tool integration, local-first architecture.
|
||
|
||
This experiment will help answer: In what ways can persistent agents actually compound value in creative professional work?
|
||
|
||
**New from Odysseus Report**: Strong validation of our focus. Recommends starting with workflow mapping + the two highest-potential use cases. Introduces concrete 5-pillar framework and memory architecture ideas for implementation. |