News
What the Head Nerd has been building and what the Department Sub-Nerds have been learning. Two streams: platform releases on the left, live evidence drops on the right as each specialist's nightly distil cron writes new high-tier findings into the brain.
Platform releases
8 entries- sourcesd7106a6
52 sources tracked · 4 new live APIs · /sources transparency page
Every search now fans out to 7 live sources in parallel instead of 3. The 52-source registry documents every place the Head Nerd can reach — Australian colleges and government first, global colleges next, with a clear link-out posture for licensed products like eTG, AMH and UpToDate.
- ›Semantic Scholar (200M papers + citation graph) wired into every specialist fan-out
- ›CORE (275M OA full-texts) with direct full-text URLs where available
- ›Crossref (170M DOI-registered works) — catches non-MEDLINE journals
- ›Epistemonikos — largest multilingual systematic review aggregator
- ›52-source registry with Australian-first priorities: RACGP, NHFA/CSANZ, RANZCOG, ADS, ATAGI Immunisation Handbook, CARI, Jean Hailes, Australasian Menopause Society, Cancer Council AU and more
- ›Licensed products (eTG, AMH, UpToDate, DynaMed, BMJ Best Practice, Embase, CINAHL) registered as link-out only — never scraped, never fabricated
- ›/sources public transparency page lists every source with coverage, country, licensing status
- ui6afbc86
PRISMA-style clinician review + Head Nerd's conclusion + copy everywhere
Results pages now render a full PRISMA 2020-aligned clinician review with a dedicated Head Nerd's conclusion section written in the router voice. Every section has a one-click copy button for pasting into notes, Word, EMRs, or emails — from novice 'copy summary' to clinician 'copy full review'.
- ›Animated magnifying-glass loading screen with Head Nerd / Sub-Nerd status messages
- ›PRISMA report structure: Title, Abstract, PICO, Introduction, Methods, Results (with study characteristics table), Discussion, and Head Nerd's conclusion
- ›GRADE strength badge on every conclusion (high, moderate, low, very low, insufficient)
- ›Copy to clipboard on every section — conclusion, abstract, citations (Vancouver), full review
- ›Open-access 'Full text' badges on citations that have reachable full text via CORE, PMC, Semantic Scholar PDF, or Crossref CC-licensed content
- ›LLM synthesis now uses structured outputs (ai.generateObject with Zod schema) — no more JSON parse failures
- infrabf3e2cc
Vercel AI Gateway with layered Claude → GPT → Gemini fallback
All LLM calls route through the Vercel AI Gateway with an explicit per-task model chain. If Anthropic rate-limits or has an incident, searches fall over to OpenAI, then Google, with zero code changes — outages won't take down synthesis.
- ›Synthesis chain: anthropic/claude-sonnet-4-5 → openai/gpt-5 → google/gemini-2.5-pro
- ›Classification chain: anthropic/claude-haiku-4-5 → openai/gpt-5-mini → google/gemini-2.5-flash
- ›Single AI_GATEWAY_API_KEY — one provider bill instead of separate Anthropic + OpenAI + Google accounts
- ›Built-in caching: identical PICO questions return instantly from edge cache
- ›Proven in production against your real key
- feature40615be
MCP server live at /api/mcp — Claude Desktop, Gemini, Cursor ready
The PICO-SEARCH brain exposes 7 public MCP tools over Streamable HTTP. Point Claude Desktop (or any MCP client) at https://www.picosearch.ai/api/mcp and you get pico_search, clinician_query, plain_language_query, classify_question_type, list_departments, get_department, and get_pico_search_status — all running against the live evidence pipeline.
- ›Async job pattern: returns a job_id in <1s, poll get_pico_search_status until completed
- ›Compound routing: PCOS routes to endocrine + gynaecology + mental_health automatically
- ›Same pipeline hit by Claude Desktop, the web app, and the bearer-token HTTPS API
- ›Full client setup guide at docs/mcp-client-setup.md
- feature547f2c4
Live PubMed + Europe PMC + ClinicalTrials.gov with evidence-pyramid ranking
Every search now fans out live to PubMed, Europe PMC, and ClinicalTrials.gov. Records are deduped by DOI → PMID → NCT → fuzzy title, then ranked by the question-type-specific CEBM pyramid with recency decay. Proven end-to-end against real clinical questions like 'Does metformin reduce HbA1c in type 2 diabetes?'.
- ›Head Nerd 3-stage router: compound rules → keyword → MeSH fallback
- ›10 compound fan-out rules including PCOS, cardiorenal, diabetic retinopathy, adult ADHD
- ›CEBM 1a → 5 tier assignment per question type (treatment / prognosis / diagnosis / aetiology)
- ›Recency decay: half-life 5y for treatment/diagnosis, 10y for prognosis
- ›40 deduped citations per search returning in 4–15s
- launche6583e8
25 agents live: Head Nerd + 24 ICD-11 specialists
The full taxonomy is live. One Head Nerd router plus 24 Department Sub-Nerds across Neurology, Ophthalmology, ENT, Cardiovascular, Respiratory, Gastrointestinal, Endocrine, Renal/Urology, Musculoskeletal, Dermatology, Mental Health, Obstetrics, Gynaecology, Men's Health, Paediatrics, Geriatrics, Haematology, Oncology, Infectious Disease, Immunology/Allergy, Emergency/Critical Care, Rehab/Pain/Palliative, Public Health, and Dental/Oral Health. Each has its own ICD-11 scope, MeSH roots, and specialty-specific system prompt.
- ›Two-layer architecture: Markdown files source of truth, Supabase rows are runtime state
- ›Sync script with sha256 drift detection
- ›Each agent has its own preferred evidence hierarchy, red flags, and cross-department overlap rules
VPS corpus stand-up: Postgres 16 + pgvector + MeSH 2026 live
Corpus-worker deployed to medasr-training (BinaryLane Brisbane). Dockerised Postgres 16 with pgvector, isolated from existing services, holding 31,110 MeSH descriptors, 235,902 synonyms, 65,360 tree positions, and 42,519 parent-child relations from NLM's 2026 MeSH release.
- ›pgvector/pgvector:pg16 in Docker on 127.0.0.1:55432 with a 768MB memory cap
- ›Fastify HTTP API on 127.0.0.1:8788 via systemd, node-cron scheduler running
- ›MeSH /mesh/expand returns matched + synonyms + ancestors + descendants via recursive CTE
- ›pico_admin, pico_worker, pico_reader Postgres roles with granular privileges
- ›Independent of the web app — can be swapped without downtime
PICO-SEARCH genesis — all 10 build phases shipped
Greenfield to production in one day. All 10 phases of the original plan shipped: monorepo scaffold, Supabase schema with 13 migrations, MCP server, corpus worker, live evidence pipeline, LLM synthesis, web UI, auth, Stripe scaffold, cross-project bearer API, and the VPS corpus bootstrap.
- ›pnpm monorepo: apps/web (Next.js 16) + apps/corpus-worker (Fastify)
- ›Supabase project qftuklfjegrhnexliexm (Tokyo region) with 14 migrations
- ›13 database tables: users, orgs, agents, search_jobs, search_results, question_classifications, saved_*, ingestion_runs, agent_memory + pgvector, subscriptions, usage_events, api_keys
- ›HTTPS Bearer API at /api/v1/search for sister projects (Tele360, TeleScribe, NRS, shescooked)
- ›picosearch.ai live on Vercel with Next 16 + AI Gateway