Skip to content
mahadizulfiker
All work
Mobile app2025–2026Talktivity

Talktivity

Real-time voice-AI English coach: talk to an AI tutor, get a CEFR report, replay every mistake.

Core contributorCore frontend + recent React Native mobile work (team of ~6–8).

Cut a 32% incomplete-score rate on AI reports to near zero.

An AI English-speaking coach. Learners hold real-time voice conversations with an AI tutor over WebRTC, then receive an LLM-generated CEFR-scored report, wrapped in an 8-week course with spaced-repetition review, gamification, and community. Answers are now recorded and analysed on a durable pipeline, so a learner can replay the exact moment of each mistake rather than read about it. Ships as an installable Android build with over-the-air updates. Freemium, billed on two rails: AamarPay (BDT) and Lemon Squeezy (USD).

Talktivity session and scoring pathA learner session runs over LiveKit and WebRTC into Groq speech-to-text and the language model. A deterministic scorer, not prose parsing, produces the CEFR report. A BullMQ-armed token expiry cuts the session at the LiveKit layer when the plan budget runs out.SESSIONweb / React NativeLIVEKITWebRTCGROQSTT / LLMKOKORO TTSspeech outDETERMINISTIC SCORERschema, not proseCEFRreportBULLMQtoken expiryends session
session to score · accent marks the deterministic path

Problem

AI-generated score reports failed silently, a measured 32% of sessions produced incomplete or missing CEFR scores, and per-user voice minutes had no hard ceiling, so cost could run away.

Constraint

Voice sessions are live and stateful; a retry can’t replay a conversation, and the fix had to work inside an existing LiveKit + LLM pipeline the team already shipped.

Decision

Replaced probabilistic parsing of the model’s report with deterministic scoring, and enforced per-plan voice-time budgets server-side with BullMQ-armed token expiry that ends a session at the LiveKit layer.

Tradeoff

Deterministic scoring is stricter and occasionally rejects a valid-but-oddly-shaped model response, so it needed a fallback path and more up-front schema work than trusting the model’s prose.

Result

Incomplete-score rate dropped from 32% to near zero, and voice-time cost became bounded and predictable per plan.

next step

Open to full-time engineering roles.

I reply within a day.

Next project

SeloraX Voice AI
Chat on WhatsApp