Personal Industry Radar Pipeline — Precis + Karpathy LLM Wiki + Claude Code Subscription Synthesis
On this page
ADR-049: Personal Industry Radar Pipeline — Precis + Karpathy LLM Wiki + Claude Code Subscription Synthesis
Status amendment (2026-05-22, ADR-050 filed 2026-05-28): The “deploy Precis + plugins” implementation choice in this ADR is superseded by ADR-050-Personal-Industry-Radar-Native-Implementation (Accepted — native
packages/rootweaver-radar/workspace package, deployed to production 2026-05-27). Other decisions (vault folder layout §2.2, source list §2.1, Karpathy synthesis layer plan §2.1 Tier 3, ADR-047 provenance integration) remain valid and are inherited by ADR-050.
Status: Proposed (2026-05-21) Date: 2026-05-21 Jira: VW-424 (parent), VW-418 (research umbrella context) Related: ADR-047-LLM-Content-Trust-Boundary (this pipeline consumes via the untrusted lane), ADR-029-KEDA-GPU-Swap-Controller (radar uses Ollama-CPU only, swap-aware by accident), ADR-034-Data-Source-Connector-Architecture (future migration target) Research basis: 2026-05-21-LLM-Input-Format-Trust-Pipeline-Migration-Archaeology (internal reference; not included)
1. Context
1.1 Problem
The 2026-05-21 research session (VW-418 (internal reference; not included)) surfaced a structural gap in Rootweaver’s architecture: the platform has 9 systems-of-record connectors (Jira, Confluence, GitLab, Discord, Slack, Gmail, IMAP, Web, File) but no curated industry-signal-ingestion path. Industry signal (ArXiv papers, vendor PSIRTs, lab blogs, newsletters, conference talks, RFCs) is currently only ingested when the operator actively goes looking — meaning the platform is structurally behind the field by however long it takes between someone explaining a topic and the operator reading about it.
The acute trigger: today’s session discovered three architectural patterns (HTML accessibility tree, dual-LLM defence, migration archaeology) that have been published consensus for 6-12 months and weren’t in the vault. A one-conversation knowledge transfer shouldn’t be the mechanism by which architectural patterns reach the vault.
1.2 Constraints
- Single GPU (RTX 4080, 15 GB VRAM) time-shared between vLLM and TEI per ADR-029-KEDA-GPU-Swap-Controller
- Anthropic API costs are unwelcome — operator has Claude Code Pro/Max subscription; per-call API billing should be avoided
- One-person operation with leverage-via-Claude — manual curation does not scale; automation is mandatory
- Provenance must follow ADR-047 — radar content is external-authored web content, must carry
trust_level=untrusted-external-webend-to-end - Storage cost must be modest — vault is currently 148K+ chunks in Qdrant; radar should add bounded volume (~6K/month)
- No new operational surface — must use existing K3s + Ollama + vLLM + Cloudflare + Discord webhook infrastructure
1.3 Existing landscape (researched 2026-05-21)
| Class | Best option | Status |
|---|---|---|
| Paid SaaS aggregator | Inoreader Pro (£72/yr) | Replaceable by self-host |
| Enterprise CI platform | Crayon / Klue / Contify (£20-40K/yr) | Wrong shape — competitor-focused not tech-evolution-focused |
| Self-hosted RSS reader | FreshRSS / Miniflux / NewsBlur | Pure RSS, no AI summary |
| AI-enhanced RSS reader | Precis (MIT, Ollama-native), RSSbrew (AGPL — reject), auto-news (too big) | Precis fits |
| Paper-discovery academia | Scholar Inbox (free, daily-indexes arXiv with personalised recs from user ratings) | Use directly |
| Email→RSS bridge | Kill The Newsletter (free hosted by leafac, or self-host) | Use hosted |
| Newsletter signal layer | AlphaSignal (100K subs, ML-ranked), Latent Space (200K subs, AI engineer focus), The Batch (Andrew Ng) | Subscribe directly |
The 5 candidate radar projects were cloned and code-reviewed in the same session — findings documented in 2026-05-21-LLM-Input-Format-Trust-Pipeline-Migration-Archaeology (internal reference; not included) §11. Key license + maintenance findings:
| Project | License | Last commit | Verdict |
|---|---|---|---|
| Precis | MIT | Dec 2024 | ★ Deploy |
| RSSbrew | AGPL v3 ⚠️ | Sep 2025 | Reject — viral copyleft, future productisation blocker |
| auto-news | MIT | Jul 2025 | Too large (Notion+iOS+Android+Airflow), harvest Helm pattern only |
| ArxivDigest | MIT | Jul 2023 | Stale; harvest relevancy prompt only |
| Kill The Newsletter | (check) | Feb 2026 | Use hosted version |
1.4 Pattern selection — Karpathy LLM Wiki vs swyx ai-notes
Two reference patterns examined:
- swyx’s
ai-notesrepo: hand-curated, no automation; 10 topic pages + monthly diary + stub notes. 2.4 MB total. Proven across 2+ years. - Karpathy’s LLM Wiki (2026): LLM agent reads
/raw, writes/wikitopic pages. Bypasses traditional RAG for personal-KB use. Self-updating.
Decision: adopt swyx’s structure (topic + monthly + stubs + resources + blog ideas) with Karpathy’s automation (LLM-driven synthesis via Claude Code subscription). Best of both — proven structure, automated curation, zero manual maintenance burden.
2. Decision
Deploy a three-tier radar pipeline composed entirely of existing infrastructure plus one new K3s service (Precis):
2.1 The three tiers
Tier 1 — Ingestion (free, mostly hosted)
| Source class | Mechanism | Authority |
|---|---|---|
| Curated newsletters (AlphaSignal, Latent Space, Smol AI, The Batch, Pragmatic Engineer) | Email → kill-the-newsletter.com → per-subscription Atom feed → Precis | Hand-curated by editors (highest quality-per-byte) |
| ArXiv papers (cs.AI, cs.CR, cs.SE, cs.LG, cs.CL) | Scholar Inbox personalised recs (account, ratings-trained) + raw arXiv RSS as backup | Personalised academic signal |
| Vendor advisories (Cisco PSIRT, Juniper PSIRT, NVD/CVE feed) | RSS direct → Precis | Security-critical signal |
| AI lab blogs (Anthropic, OpenAI, DeepMind, HuggingFace, Cloudflare) | RSS direct → Precis | Vendor-of-record signal |
| Curated discovery (Simon Willison, Lilian Weng, Sebastian Raschka, Eugene Yan) | RSS direct → Precis | Individual high-signal authors |
| Tech aggregators (HN front-page filtered, GitHub trending) | RSS direct → Precis | Community signal |
Tier 2 — Filter + summarise (free, local LLMs)
- Precis deployed in
rootweaverK3s namespace as a Deployment - LLM backend: existing Ollama with
qwen2.5:3b(CPU-only, ~1.9 GB RAM, ~10 tok/s) - Per-item relevance scoring (1-10 scale, threshold ≥7) using the ArxivDigest relevancy prompt (proven prompt template, ported into Precis config)
- Per-item ~100-word summary for items above threshold
- Provenance auto-tagged per ADR-047:
trust_level=untrusted-external-web,source_system=radar/<feed-name>,field_owner=external-untrusted-web - Two custom Precis plugins (~80 LoC total, MIT, upstreamable):
notification/discord.py— Discord webhook handler (existing pattern, copyslack.py)notification/vault_writer.py— writes summaries to04-Resources/AI-ML-Research/Raw/YYYY-MM-DD/item-<slug>.md
Tier 3 — Synthesise (free, Claude Code subscription)
- Two Claude Code scheduled agents via the
/scheduleskill (uses subscription quota, not API tokens):ai-radar-nightly— runs 02:00 UK daily; readsRaw/<today>/, updatesWiki/<Topic>.md(Karpathy pattern), createsWiki/_stubs/for emerging topics (swyx pattern), appends to01-Journal/AI-Radar/YYYY-MM.mdai-radar-weekly-digest— runs 18:00 UK Friday; top-5 items + gap analysis against current HLD/LLD + Discord webhook post
2.2 Vault folder layout (adopted from swyx with Karpathy automation)
04-Resources/AI-ML-Research/
├── Raw/ ← Precis writes here, untrusted-external-web
│ └── 2026-05-21/
│ └── item-<slug>.md
├── Wiki/ ← Claude Code agent maintains, Karpathy pattern
│ ├── Prompt-Injection-Defence.md
│ ├── Accessibility-Tree-Consumption.md
│ ├── GraphRAG-Patterns.md
│ ├── Migration-Archaeology.md
│ └── _stubs/ ← swyx pattern — watch-list items
│ └── *.md
├── Resources/ ← existing — standing references
└── (existing files...)
01-Journal/AI-Radar/ ← chronological diary
└── 2026-05.md2.3 What does NOT happen here
To prevent scope creep, this ADR explicitly does not mandate:
- Migration to a native RSS connector in the connector framework (this is a future option in Sprint 3+ review, not a requirement)
- Indexing of Raw items into Qdrant beyond what vault-indexer naturally does (no special-case logic)
- Multi-user / multi-tenant operation (Precis is single-user, fine)
- Real-time alerting (daily cadence is sufficient; sub-daily not justified by signal velocity)
- Synthesis using Anthropic API (Claude Code subscription only — see §3.2 cost rationale)
3. Consequences
3.1 Positive
- £0/yr marginal cost vs. £72/yr (Inoreader Pro), or £20-40K/yr (enterprise CI). All inference free at marginal level (Ollama CPU, Claude Code subscription)
- Zero new infrastructure — Ollama already deployed (HLD-02 §1); Claude Code subscription already paid; Cloudflare/Discord webhooks already configured
- Vault becomes synthesis substrate (Karpathy pattern) — radar output is searchable via existing Qdrant indexing, joinable via vault-postgres, and traceable via FalkorDB once provenance is implemented (VW-420/421)
- Trust boundary preserved — all radar content carries ADR-047
untrusted-external-webprovenance; downstream consumers (Oracle agent, recall_decisions) can downweight via existing trust-aware retrieval - GPU-swap-aware by accident — radar uses Ollama CPU only, so remains operational during TEI embedding windows when vLLM is paused (ADR-029)
- Survives operator vacation — automated daily ingest + automated synthesis means going on holiday doesn’t accumulate signal debt
- Composable — Precis’s handler abstraction (
LLMHandler/NotificationHandler/StorageHandler/ContentHandler) makes future extensions cheap (e.g., add Telegram notification by writing one Python class) - Strategic option-value — if Personal Radar succeeds, the same architecture can be productised as a service for client work (one of the strategic options flagged in 2026-05-21-HLD-LLD-Review-LLM-Trust-Boundary (internal reference; not included) §8.1)
3.2 Negative
- Precis is semi-stale (last commit Dec 2024, 5 months ago). Risk: critical bugs or security issues remain unpatched. Mitigation: MIT licence — operator can fork at any point; codebase is ~3K LoC Python (auditable in one sitting); maintain a local patch set if upstream goes dormant.
- Hosted Kill The Newsletter is a third-party dependency (run by leafac). If service shuts down, newsletter→RSS layer breaks. Mitigation: self-host backup option exists (Docker variant at
semiformal-net/kill-the-newsletter); migration is ~1 day of work if needed. - Ollama
qwen2.5:3bquality cap — 3B params is small; relevance-judgement quality may be insufficient for nuanced topics. Mitigation: bump toqwen2.5:7b(CPU still works, slower; ~30s per call instead of ~5s) if accuracy audit reveals gaps. Alternative: use vLLM Qwen3-8B-AWQ for filter step when vLLM is idle. - Claude Code subscription quota — heavy synthesis use could brush against monthly limits. Mitigation: nightly+weekly cadence is bounded; if quota exhausted, fall back to vLLM Qwen3-8B-AWQ for synthesis (lower quality but free).
- No Discord notification in Precis upstream — one of the few notification destinations missing. Mitigation: write
notification/discord.py(~30 LoC, MIT contribution back); already-committed work in VW-424 acceptance criteria. - No vault-writer notification in Precis upstream — needed for Karpathy pattern. Mitigation: write
notification/vault_writer.py(~50 LoC); upstream-able as a generic markdown-writer plugin if desired. - Karpathy LLM Wiki pattern is unproven at scale — the academic claim is recent (2026), few production references. Mitigation: hybrid with swyx’s proven structure; if Karpathy automation fails, fall back to chronological diary (still useful, just less curated).
- Signal-quality risk — automation may surface noise instead of signal; relevance scoring may miscalibrate. Mitigation: 2-week pilot before declaring victory; signal-quality audit in acceptance criteria; threshold and prompt tuning iteratively.
3.3 Risks (severity × likelihood)
| Risk | Likelihood | Impact | Mitigation |
|---|---|---|---|
| Precis upstream goes fully dormant | Medium | Medium | Fork; ~3K LoC, owner-able |
| Hosted Kill The Newsletter dies | Low | Medium | Self-host (1-day migration) |
| Ollama qwen2.5:3b too small for relevance scoring | Medium | Low | Bump to 7b or use vLLM idle windows |
| Claude Code subscription quota exhaustion | Low | Low | Reduce cadence or fall back to vLLM |
| Radar output drifts toward low-signal aggregator content | Medium | Medium | Re-tune relevance prompt + threshold; rotate sources quarterly |
| Discord webhook gets compromised → injection vector | Very low | Low | Webhook is outbound only; no inbound risk |
| Prompt-injection content in newsletters reaches Claude Code synthesis agent | Medium | Medium | ADR-047 spotlighting + provenance tagging contains it; agent has no destructive tools |
3.4 Compliance posture
This pipeline is compatible with future EU AI Act / NIST AI RMF compliance documentation because:
- Every radar chunk carries provenance with traceable source URL + ingest timestamp + transformation chain (ADR-047 §2.2)
- Source list is config-driven and auditable (YAML in
configs/precis.yaml) - Synthesis decisions are logged (Claude Code session transcripts + memory-writer captures all reasoning rounds)
- “Right to know” queries answerable via
SELECT chunks WHERE provenance->'chain' @> '[{"step":"ingest","system":"radar"}]'once VW-420 lands
4. Alternatives Considered
4.1 Paid SaaS (Inoreader Pro + Readless or similar)
Subscribe to Inoreader Pro (£72/yr) + an AI-digest service like Readless. Bypass self-hosting entirely.
Rejected because:
- Recurring £72-200/yr cost vs free
- Vendor lock-in for the curation layer
- Provenance metadata can’t be added at the SaaS layer (no integration with ADR-047 trust boundary)
- No path to vault-native synthesis (Karpathy pattern requires file system access)
4.2 Pure-build RSS connector in existing framework
Skip Precis entirely. Build the RSS connector planned for Sprint 3+ as part of the connector framework (packages/rootweaver-connectors/sources/rss/). Filter + summarise as KEDA-scaled consumers.
Rejected (for now, kept as future option) because:
- ~1-2 sprints of original work vs ~2 hours deploying Precis
- Reinvents Precis’s handler abstraction without clear benefit
- Operator gets zero radar signal during the build period
Future migration path retained: after 6 months of running Precis, re-evaluate. If radar is heavily used and Precis goes dormant, port functionality into the native connector framework. ADR-049 supersession would be the trigger.
4.3 RSSbrew (AGPL v3) self-hosted
Active maintenance (Sep 2025), Docker-deployable, similar feature set to Precis.
Rejected because:
- AGPL v3 viral copyleft — any modification must be released under AGPL when run as network service
- Future productisation (migration archaeology pivot in 2026-05-21-HLD-LLD-Review-LLM-Trust-Boundary (internal reference; not included) §8.1) would be legally compromised
- No Ollama support in
requirements.txt— would need fork to add local LLMs, ironically triggering AGPL release requirements - “Still under development” — author’s own assessment of completeness
4.4 auto-news (MIT, K8s-native via Helm)
Largest scope of the 5 reviewed projects; ships with a Helm chart matching Rootweaver’s deployment model.
Rejected because:
- Scope mismatch — multi-source (Twitter, Reddit, YouTube, RSS), Notion-based reading UI, iOS+Android mobile apps. ~10K LoC.
- Tech stack mismatches: Airflow DAGs (we use Prefect), Notion (we use Obsidian), embedded ChromaDB (we use Qdrant)
- Helm chart pattern is interesting but not worth importing the whole project for
Retained value: the Helm chart pattern is a useful reference for future Sprint 3+ RSS connector if we build native.
4.5 ArxivDigest deployed as a service
Direct deployment of AutoLLM/ArxivDigest as a GitHub Actions cron.
Rejected because:
- Last commit Jul 2023 — 3 years stale
- Hardcoded to
gpt-3.5-turbo-16k(OpenAI API only) - Would need a fork to swap to Ollama, defeating the “use OSS as-is” benefit
Retained value: the relevancy prompt is excellent and portable — adopted verbatim in Precis config.
4.6 Anthropic API for synthesis (instead of Claude Code subscription)
Use Anthropic Messages API directly from a Prefect flow for synthesis, with prompt caching for cost reduction.
Rejected because:
- Per-call API billing vs already-paid Claude Code subscription
- Estimated cost: £100-300/mo at typical synthesis volume even with prompt caching
- Claude Code scheduled agents are the natively-supported path (
/scheduleskill,CronCreatetool) - API path adds an Anthropic billing dependency that doesn’t exist today
4.7 Build-from-scratch in connector framework (no external project)
Skip Precis. Build the radar pipeline natively in packages/rootweaver-connectors/sources/radar/ with KEDA consumers for filtering and summarisation.
Rejected (for now) because:
- ~2 sprints of original work
- Operator gets zero radar signal during the build period
- Precis already does 80% of what we’d build, at MIT license
Future migration path: if Precis goes dormant or doesn’t suit, this is the natural fallback. ADR-049 would be superseded by a new ADR articulating the native-connector decision.
5. Implementation Plan
5.1 Weekend / today (~2 hours, zero infrastructure changes)
| # | Action | Effort | Cost |
|---|---|---|---|
| 1 | Sign up to Scholar Inbox | 5 min | £0 |
| 2 | Subscribe to AlphaSignal + Latent Space + Smol AI News + The Batch + Pragmatic Engineer (Gmail label ai-radar) | 10 min | £0 |
| 3 | Set up kill-the-newsletter.com — one feed per newsletter | 10 min | £0 |
| 4 | Deploy Precis to K3s rootweaver namespace (Deployment + Service + PVC for data dir) | 30 min | £0 |
| 5 | Configure Precis: Ollama backend (qwen2.5:3b), 8-10 initial sources, Discord webhook | 20 min | £0 |
| 6 | Copy ArxivDigest relevancy prompt into Precis system prompt | 5 min | £0 |
| 7 | Pull qwen2.5:3b on Ollama: kubectl exec -n rootweaver ollama-... -- ollama pull qwen2.5:3b | 5 min | £0 |
| Total weekend | ~85 min | £0 |
5.2 Sprint 2 (0.5 sprint, plugins + agents)
| # | Action | Effort |
|---|---|---|
| 8 | Write notification/discord.py Precis plugin (~30 LoC, copy slack.py pattern) | 0.5 day |
| 9 | Write notification/vault_writer.py Precis plugin (~50 LoC, writes markdown to vault-mounted PVC) | 0.5 day |
| 10 | Configure provenance metadata in vault-writer plugin per ADR-047 | 0.5 day |
| 11 | Create Claude Code scheduled agent: ai-radar-nightly (02:00 UK daily) | 0.5 day |
| 12 | Create Claude Code scheduled agent: ai-radar-weekly-digest (18:00 UK Friday) | 0.5 day |
| 13 | Optional: upstream the Discord plugin to leozqin/precis as a PR | 0.5 day |
| Total Sprint 2 | ~3 days |
5.3 Sprint 3+ (review + iterate)
| # | Action | Trigger |
|---|---|---|
| 14 | 2-week signal-quality audit: % relevant items, % missed important items | After 2 weeks running |
| 15 | Prompt + threshold tuning based on audit | After audit |
| 16 | Source-list rotation (add/remove based on signal-to-noise observed) | Quarterly |
| 17 | Optional: bump Ollama model to qwen2.5:7b if 3b quality insufficient | If audit shows accuracy gaps |
| 18 | Optional: native RSS connector migration (re-evaluate need vs Precis) | 6 months in |
| 19 | Optional: productise externally if pattern proves valuable for client work | If §8.1 strategic decision goes that way |
6. Validation Approach
6.1 Pilot success criteria (2 weeks post-deployment)
- Coverage metric: of the 12 “tier 1 + tier 2” gaps identified in 2026-05-21-HLD-LLD-Review-LLM-Trust-Boundary (internal reference; not included), does the radar surface ≥6 of them organically in the 2-week window?
- Signal-to-noise: of items reaching the operator (above threshold), what fraction are flagged as relevant on manual review? Target: ≥70%
- False-negative audit: review a known industry event (e.g., major model release, paper publication) — did the radar catch it?
- Wiki growth: at least 5 new Wiki pages or stub notes created per week by the synthesis agent
6.2 Failure modes to monitor
- Aggregator-of-aggregators effect: if Smol AI / AlphaSignal references content that other sources also reference, the radar would surface it multiple times. Dedupe by article URL in vault-writer plugin.
- Drift toward low-signal sources: any source contributing < 10% of through-threshold items over 4 weeks gets a quarterly review for removal.
- Synthesis agent hallucination: Claude Code’s nightly synthesis pulling concepts from injected content. ADR-047 spotlighting + provenance check should catch — verify in audit.
6.3 Decision points
- Week 2: continue or kill? If signal-quality < 50%, prompt-tuning sprint. If < 30%, reconsider Precis vs build-native.
- Month 3: scope expansion? Add more sources, multi-modal (YouTube transcripts via auto-news pattern), or hold steady.
- Month 6: integration depth? Migrate to native RSS connector OR stick with Precis for another period.
7. Related
- VW-424 (internal reference; not included) — implementation Jira ticket
- VW-418 (internal reference; not included) — research umbrella that produced this ADR
- ADR-047-LLM-Content-Trust-Boundary — radar consumes via the untrusted lane; Mode A defence applies to synthesis-agent reading of radar content
- ADR-029-KEDA-GPU-Swap-Controller — radar’s CPU-only Ollama use is naturally swap-aware
- ADR-034-Data-Source-Connector-Architecture — future migration target if native RSS connector built
- ADR-046-Memory-System-Embedder-Architecture — radar wiki content gets embedded by the existing pipeline (Qwen3-Embedding-4B, 2560d)
- 2026-05-21-LLM-Input-Format-Trust-Pipeline-Migration-Archaeology (internal reference; not included) — research synthesis (companion canonical doc)
- 2026-05-21-HLD-LLD-Review-LLM-Trust-Boundary (internal reference; not included) — gap analysis that surfaced the radar need
- HLD-02-RAG-Platform (internal reference; not included) — current connector framework (potential future home)
8. References
Deployed projects (cloned and code-reviewed 2026-05-21)
- Precis (leozqin/precis) — MIT, Ollama-native, FastAPI; the deployed radar service
- Kill The Newsletter (leafac/kill-the-newsletter) — email→RSS, hosted version preferred (kill-the-newsletter.com)
Pattern donors (concepts harvested, not deployed)
- ArxivDigest (AutoLLM/ArxivDigest) — relevancy prompt template ported verbatim
- swyxio/ai-notes — folder structure pattern (topic + monthly + stub + resources + blog ideas)
- Karpathy’s LLM Wiki pattern (2026) — automation pattern (LLM writes wiki pages from raw)
Source list (initial, evolvable)
- Scholar Inbox — personalised arXiv recs
- AlphaSignal — weekly ML-ranked research breakdowns
- Latent Space — AI engineer weekly
- Smol AI News — daily AI-Discord/Twitter recap
- The Batch (DeepLearning.AI) — Andrew Ng’s weekly digest
- Simon Willison’s blog — high-signal individual author
- Anthropic news — vendor-of-record
- Cisco PSIRT — security advisories
- NVD CVE feed — security advisories
Foundational references
- Karpathy’s LLM Wiki step-by-step
- Obsidian AI Second Brain Complete Guide 2026 (NxCode)
- AI Newsletter Arena (Arize comparison)
Proposed 2026-05-21. Awaiting architect review (operator self-review) before VW-424 implementation begins. First concrete action: §5.1 weekend deployment (~85 min).