Skip to main content

Domain-Specific AI Agents — Why I Build My Own myGPTs

· 3 min read
Sampras Zheng
Sampras Zheng
Engineer

General-purpose LLMs are powerful, but when it comes to deep analysis in specialized domains, they have two hard limitations: knowledge that lags behind the cutting edge, and factual dilution for niche fields. Take the Polkadot ecosystem as an example — key 2024–2026 developments like Agile Coretime, RegionX, and JAM are either unanswerable by general models, or return outdated information.

The solution is not a bigger model — it's giving the model a structured, trustworthy external knowledge base, then wrapping it with a persona and task prompt. That's the motivation behind my myGPTs.

The Two I Currently Use​

Jamia — Polkadot × JAM Expert​

  • 🔗 Jamia on ChatGPT
  • Training data: Gavin Wood talk transcripts, JAM Gray Paper summaries, Polkadot Wiki, cross-source synthesis from 10+ Polkasharks podcast episodes
  • Strengths: JAM architecture, Coretime economics, XCM cross-chain messaging, OpenGov governance analysis
  • Weaknesses: Real-time prices, simple term definitions (use Grok/ChatGPT for those — it's faster)

Spacesharks — Space Industry Analyst​

  • 🔗 Spacesharks on ChatGPT
  • Training data: Upstream/midstream/downstream breakdown of the LEO industry, key Orbital Data Center (ODC) players, Taiwan supply chain map (Qorvo/WIN Semi, Auden, Tripod, etc.)
  • Strengths: LEO lifecycle analysis, Starlink/Amazon Leo/Starcloud comparisons, ISL optical communications, Taiwan's position in the global LEO ecosystem
  • Further reading: The Space section of 2026 Tech Roadmap

The System Behind Them: Obsidian Wiki + LLM Ingest Loop​

These two GPTs are not arbitrarily tuned. They share an Obsidian vault as their knowledge backbone:

wiki/
├── sources/      one file per source, individually ingested raw material
├── entities/     pages for people, companies, and products
├── concepts/     protocols, architectures, industry chains, and other concepts
└── synthesis/    long-form pieces produced from cross-source synthesis

Every time I read new material (podcasts, papers, earnings reports), I run this workflow:

  1. Discuss key takeaways with Claude Code
  2. Write it into sources/<slug>.md
  3. Update or create relevant entities/ and concepts/ pages (a single ingest typically touches 5–15 pages)
  4. Update index.md and log.md

The text content of this vault is exactly what feeds Jamia and Spacesharks. When a reader asks a cross-source question, the model is retrieving from the context graph I built by hand — not guessing from across the entire internet.

Why Not Just Use RAG or an API Integration?​

I tried that. For a solo-maintained knowledge base, Custom GPTs have a clear advantage:

NeedCustom GPTSelf-built RAG
No backend, freely shareable✅❌ Need to host a vector store
Knowledge base under 10 MB✅ Direct upload⚠️ Over-engineering
Frontend UX (works on mobile)✅ ChatGPT app❌ Need to build your own
Custom tools (news fetching / price lookup)⚠️ Limited actions✅ Fully controllable

Once the knowledge scale or interaction patterns exceed what Custom GPT can handle, migrating to a RAG / Agent architecture is always an option. Get the knowledge right first, then worry about the system.

Next Steps​

  • The long-form pieces currently in wiki/synthesis/ will gradually be published to this blog — start with LEO × Taiwan ODC Gap.
  • Jamia and Spacesharks' actions (market price / on-chain data fetching) are in progress; I'll write a separate post once they're done.
  • The next myGPT will probably be an RF/hardware domain debug assistant, grounded in the taxonomy from RF/SoC Debug Playbook.