oll.am · oll.in POC showcase · live on synthetic data · 2026-07-07
oll.in — the POC, running
A simple one-process frontend that runs all three oll.in POCs live on synthetic data: cited retrieval with honest abstain, the three grounded dossier verbs, and the Career Radar digest. One command, no keys —
sh scripts/poc_demo.sh → http://localhost:8099.
Branch: poc/ollin-demo (merges #86·#87·#88) → PR #89
Run:
sh scripts/poc_demo.sh
Verified: live server + 3 endpoints + screenshot
What it looks like Playwright screenshot of the running demo
Three panels — Your evidence (retrieved cited passages + the honest-abstain control), Your dossier (the tailored CV, cover letter and interview questions from the live verbs, with [n] citations and gap-markers), and Career Radar (the live scored job digest). Captured from http://localhost:8099 after clicking "Generate the dossier".
Click to open full size.
How it runs one keyless process
The demo is a tiny Flask "demo BFF" (demo/server.py) that loads the three POCs in-process — no external services, no keys. It seeds the synthetic RAG corpus on startup, then serves one vanilla page plus three endpoints.
| Endpoint | Runs | Live result (verified) |
|---|---|---|
| POST /api/demo/ground | RAG retrieval over the synthetic corpus (relevance floor + abstain) | 15 cited passages over 6 job facets; ?abstain=1 on "quantum biology…" → cosine 0.072 < floor 0.08 → abstained, no fabricated citation. |
| POST /api/demo/dossier | The 3 oll-write verbs (offline mock model) on the retrieved context | Tailored CV + cover letter + interview prep, with [n] citations and an honest [gap] marker; 15 real citations attached. |
| GET /api/demo/radar | The oll-scout synthetic tick (fixture jobs → score) | Pulled 7 / scored 7 / 4 strong (≥60); top match 91 — Senior Backend Engineer (Python), Helvetia Fintech AG · Zürich. |
Real vs mock no overclaiming
Real computations
Genuinely live (on synthetic data)
- Retrieval, the relevance floor, and the honest abstain path.
- Citation numbering & passage → evidence mapping.
- Job scoring, ranking, and the strong-match digest.
- The full pipeline wiring end-to-end, in one process.
Mocked for the POC
Not proven here
- Text generation is an offline stub (
provider=mock-offline) — pipeline demo, not prose quality. - The candidate + jobs are fictional synthetic data.
- Real-model send-worthiness = a Sam gate (needs a live model).
- Live Swiss job coverage = a Sam gate (needs a free Jooble key).
The point of this showcase: it makes the architecture tangible — you can click through the exact loop oll.in will run (evidence → grounded dossier → radar) and see the honesty guarantees (abstain, gap-markers) actually firing, before we spend anything on real models or data. Nothing here is merged or deployed; it's a draft PR you can run locally. See the POC results and the solution architecture.