Mission
Tahrirchi Bot gives Uzbek editorial teams an always-on copy editor inside Telegram. Journalists can forward drafts, receive grammar suggestions, and push approved text straight into their CMS—all without leaving the chat.
System Snapshot
- Telegram updates land in the gateway bot (Aiogram) which authenticates editors.
- Requests are streamed to a FastAPI edge service for rate limiting and telemetry.
- gRPC workers run specialised NLP tasks: morphological analysis, grammar correction, and style scoring.
- Results and metadata are persisted in PostgreSQL, while Redis queues decouple long-running jobs.
Feature Highlights
- Offline resilience: when Telegram throttles webhooks, the bot falls back to polling with exponential backoff.
- Editor context: Redis stores recent conversation context so suggestions stay relevant to the article at hand.
- Batch proofing: editors can drop a
.docx, and the pipeline fans it out into paragraphs processed in parallel. - Quality bar: confidence scores below 0.7 are flagged with “second opinion” messaging to keep humans in control.
Usage Patterns
| Metric | Value | Comment |
|---|---|---|
| Daily active editors | 1.4K+ | peaks before newspaper deadlines |
| Texts processed / day | 6,500 – 8,200 | 40% run through batch mode |
| Average turnaround | 1.2 s | end-to-end, including Telegram I/O |
Tooling for Editors
Text
Draft drop → Pipeline fan-out
├─ Core grammar worker
├─ Terminology check (custom newsroom dictionary)
└─ Style heuristic scorer
Workers stream back to the review hub → formatted message → Telegram reply
The bot posts inline diffs with highlighted replacements. Editors can tap suggested fixes or reply with /reject to feed the learning dataset.
Operational Playbooks
- Elastic workers scale horizontally via Docker Swarm based on queue depth.
- Alerting pings Telegram admins when success rates drop below 98% over 5 minutes.
- Full audit trails allow compliance teams to trace who approved which changes.
Future Iterations
Upcoming releases will introduce per-editor glossaries, multi-language pipelines, and a dashboard that surfaces content quality trends to newsroom leads.
