README: add end-to-end example with input transcript and annotated output

Author: npow

README.md

@@ -8,7 +8,64 @@
 
 You did the interviews. You took the notes. But now you're staring at a pile of Google Docs trying to remember what that one PM said three months ago about dashboards. Sound familiar?
 
-Elicit reads your interview transcripts and draws out what matters: what jobs your customers are actually trying to do, what's painful, what workarounds they've duct-taped together, and — most importantly — what you should build next, backed by evidence you can trace back to specific quotes.
+Paste in a transcript. Elicit extracts jobs-to-be-done, pain points, and workarounds with supporting quotes — then, after a few interviews, tells you what to build next with a ranked, evidence-backed recommendation you can take straight to your team.
+
+---
+
+**Input — raw interview transcript (excerpt):**
+
+> **Sarah (PM, B2B SaaS):** When planning time comes around, I'm trying to remember what people said three months ago. I'll search through my notes but it's really hard to find patterns across multiple interviews. Probably 2-3 full days per quarter just re-reading notes and trying to pull out themes. I'll highlight things, copy-paste into a spreadsheet, try to tag them. It's a mess.
+>
+> The gap between "I have a bunch of interviews" and "I know what to build" — that's where I feel like I'm just guessing. And when I present my recommendations to leadership, I can't really trace back to specific evidence. It's like "trust me, I talked to customers." That doesn't fly anymore.
+
+**Output — structured extractions, scored and quoted:**
+
+```
+JOB-TO-BE-DONE
+  Synthesize interview data into a defensible, prioritized product decision
+  importance: 9.1/10 · satisfaction: 2.0/10 · confidence: 0.94
+  "connecting the dots across 10 interviews to find the real patterns —
+   that's where I feel like I'm just guessing"
+
+PAIN POINTS
+  • Interview insights rot in unstructured notes, invisible at planning time  [severity: high]
+    "all that information just sits in my Google Docs...trying to remember
+     what people said three months ago"
+
+  • No evidence trail from customer voice to product decision  [severity: high]
+    "I can't really trace back to specific evidence. 'Trust me, I talked
+     to customers.' That doesn't fly anymore."
+
+WORKAROUND
+  • Manual highlight → copy-paste → spreadsheet to find cross-interview themes  [effort: high]
+    "I'll highlight things, copy paste into a spreadsheet, try to tag them.
+     It's a mess."
+
+OPPORTUNITY SCORE: 17.6 / 20
+  Automate the synthesis layer — extract patterns across interviews and surface
+  evidence-backed recommendations without the 2-3 day manual slog.
+```
+
+**After 3+ interviews, Elicit synthesizes across them and generates a ranked recommendation:**
+
+```
+#1 BUILD NOW  (priority: 0.91 · confidence: 0.88 · supported by 4/4 interviews)
+
+  "Ship an automated synthesis view that surfaces recurring JTBD and pain
+   patterns across all interviews, with evidence chains to specific quotes."
+
+  Rationale: Every interviewed PM spends 2-3 days/quarter manually re-reading notes
+  to find patterns they're not confident they're catching. Satisfaction with current
+  workarounds (spreadsheets, Dovetail) is near zero. High importance, no good solution.
+
+  Evidence:
+    [pain]  "all that information just sits in my Google Docs"           — Sarah, interview 1
+    [pain]  "I end up with conflicting information and I don't know why" — Maya, interview 2
+    [job]   "connecting the dots across 10 interviews"                   — Sarah, interview 1
+    [wa]    "copy paste into a spreadsheet, try to tag them. It's a mess"— Sarah, interview 1
+```
+
+---
 
 ## Quick start
 
@@ -95,7 +152,7 @@ Then upload audio files through the API. Requires ~1GB for the base model downlo
 discovery_engine/
   models/      SQLAlchemy data models
   schemas/     Pydantic schemas (API + LLM output parsing)
-  engine/      Core logic (extraction, synthesis, recommendations, coaching, simulation, calibration)
+  engine/      Core logic (extraction, synthesis, recommendations, simulation)
   llm/         LLM client + 13 Jinja2 prompt templates
   api/         FastAPI routes
 streamlit_app/ Streamlit UI (9 pages)