Voice2Policy: citizen feedback at machine scale.

Public consultation produces thousands of free-text submissions. Reading them all is impossible, sampling them is biased, and summarizing them by hand is slow and opaque. Policy ends up shaped by whoever shouts loudest, not by what people actually said.

Voice2Policy was built to close that gap: verify contributors at the door, make the full body of input readable, and keep the trail from published conclusion back to individual voices.

Three moving parts. Verified intake runs polls and collects opinion from confirmed contributors. A fine-tuned deBERTa-v2 model, served on AWS SageMaker, classifies submissions by topic and stance at volume. A synthesis layer assembles whitepaper drafts and policy artifacts from that classified base, with poll results published alongside.

The workhorse is a classifier, not a text generator. Classification on a fine-tuned encoder is cheap, fast, and auditable at scale. Generation only enters at the drafting stage, where humans review before anything publishes. Judgment where it earns its keep, deterministic software everywhere else.

Built as a full-stack platform: NestJS services behind a Next.js front end, with model inference isolated on SageMaker so classification scales independently of the application.

The public surface runs today: a whitepaper hub, live polls with published results, and a publications archive. The featured whitepaper lays out a strategic framework for applying AI to national development in Bangladesh, from healthcare access to disaster preparedness.

• Live in production at voice2policy.aiwhitepaper hub, polls, poll results, publications
• Classification runs on a fine-tuned encoder, not per-item LLM callscost and latency hold at volume
• Published artifacts stay linked to the opinion data behind themthe transparency claim is structural, not marketing