Skip to content

NilaMind v1.9.9

Choose a tag to compare

View all tags
@sampathmannam sampathmannam released this 12 Jul 18:03
v1.9.9
a862fc6

Catch-up release closing the gap since v1.9.2 — a large batch of work had accumulated on main without a published build. Highlights, grouped:

Safety (§9)

  • Tamil/Telugu negation-first crisis-phrase coverage locked in, plus a native-review flag
  • Hindi precision hardening + distortion negation-guards (adversarial review)
  • Crisis takeover now fires only on genuine risk (was over-triggering); solution-forward overlay
  • Removed the persistent CrisisPill from the app shell — crisis stays reachable, not constantly visible
  • Region & crisis-lines selector; Tele-MANAS first-line + pesticide means-safety category

Companion voice

  • Short, non-lecturing companion replies (persona + exemplar-RAG) instead of long explainers
  • Fixed first-person mirroring / role-confusion in replies
  • Per-turn steering for "why/how" questions so the model reflects instead of lecturing
  • Anti-sycophancy: closed a paraphrase-evasion gap

Clinical research wave 2

  • BA/PST implementation intentions + fidelity, GAD-worry and DBT-skills routing fixes
  • ACT values-based-action micro-protocol, inhibitory-learning exposure prompts
  • Alliance-voice validate-first sequencing, engagement-onboarding personalization
  • EMA cadence copy + streak de-emphasis; Social Rhythm Metric (IPSRT) tracker
  • Lethal-means counseling flow (Stanley-Brown step 6); assertion training module

Reports & exports

  • Clinician-facing PDF report (behavioral insights, sleep analysis, assessment trajectory)
  • Structured JSON clinical export; experimental FHIR R4 export
  • Selectable 7/30/90-day report window + on-device usage section

UI

  • NilaFace redesign (biophilic breathing, crisis differentiation, haptics)
  • Manic-first state layer — elevation reaches the UI, not just copy
  • Accessibility pass: tap targets, focus rings, text floor

Housekeeping

  • README no longer hardcodes a stale version number
  • Dead code removal (unused components, dead exports)

Default model is still the stock, un-fine-tuned Gemma-3-1B-Instruct (~806MB) — see docs/TRANSPARENCY.md for the full model card and safety architecture. All inference, crisis detection, and voice transcription run on-device; no account, no backend, no analytics.

Try it now, no install: https://nilamind.netlify.app
Android: download the APK below, or track releases automatically via Obtainium (see DISTRIBUTION.md).

Assets 3
Loading