Admitly runs specialized Planner-Executor-Critic pipelines — Opus where judgment compounds, Haiku where work is mechanical. The Critic scores every draft on a 4-axis rubric calibrated against real SOPs from a study-abroad consulting partner, looping up to 3x until it passes.
Built with
Click any stage to inspect its input, output, and prompt. Toggle scenarios to see how the routing path changes.
Task Decomposer
opus-4.7 · $15/MStrong-model planning. Decomposes the request into ordered sub-tasks for the Executor, decides what RAG context is needed, and sets the iteration budget (2–3 loops) so latency stays predictable.
Prompt Pattern
You are the Planner. Given pipeline="{tag}", profile, target → emit ordered sub-tasks[] with required RAG sources. Cap iterations at 3.Mock Input
{
"pipeline": "sop",
"profile": { "publications": 1, "research_area": "instruction tuning" },
"target": "CMU LTI PhD"
}Mock Output
█Not a wrapper. Every component is a deliberate engineering decision with measurable tradeoffs.
pgvector cosine similarity + BM25 keyword re-ranking over 2,000+ program profiles. Retrieval latency under 50ms.
Strong models where judgment compounds (Planner, Critic). Light models where work is mechanical (Executor). Loop capped at 3 iterations.
Rubric calibrated against real SOP cases from a study-abroad consulting partner. In-product 👍/👎 ratings flow back into prompt iteration.
Offline
LLM-as-judge calibrated on partner-scored SOPs
Online
User 👍/👎 → prompt + rubric refinement
4-axis LLM-as-judge rubric. Tested on 10 real prompts: 78% passed without human intervention; failures pinpointed weak personalization.
Ephemeral cache_control on system prompts >1024 tokens. Up to 90% token reduction on repeated RAG contexts.
Cache hit rate — avg 84% this week
Tool-use with Zod-derived JSON schemas. No "please return JSON" prompting — Claude fills a type-safe tool call every time.
const SchoolMatchSchema = z.object({
schools: z.array(z.object({
name: z.string(),
fit_score: z.number().min(0).max(100),
reasons: z.array(z.string()),
deadline: z.string(),
})),
confidence: z.number(),
});
// → Claude tool_use, never hallucinated JSONWatch the pipeline execute end-to-end on real applicant scenarios.
Match Schools · CS PhD applicant · NLP focus
Name: Alex Chen GPA: 3.8/4.0 (top 5% of class) GRE: 165Q / 158V / 5.0W Research: NLP — 1 pub at ACL 2024 (co-author) Advisor interest: LLM alignment, instruction tuning Location preference: East Coast or no preference Budget: prefer funded positions