Rafiqspace.ai is an Indonesian AI software company building responsible, enterprise-grade speech intelligence for Bahasa Indonesia—serving government agencies, national health institutions, and regulated enterprises across the country.

Indonesia’s public sector generates hours of high-stakes meeting audio daily. Parliamentary sessions at DPRD DKI Jakarta and DPR RI require verbatim records with full speaker attribution. Manual transcription delivered 70%–80% accuracy and required 3–4 hours of human effort per hour of audio, making same-day distribution of official records impossible.

Commercial speech-to-text tools offered no viable alternative. None were fine-tuned for Bahasa Indonesia’s regulatory vocabulary or multi-speaker environments, and streaming transcription ran below 1x real-time—too slow for live note-taking. Rafiqspace.ai needed a platform that could meet government accuracy thresholds, deliver sub-300 ms streaming latency, and identify individual speakers automatically.

Rafiqspace.ai built its platform on NVIDIA Nemotron models and libraries, fine-tuning the Parakeet-CTC ASR model on a proprietary dataset spanning parliamentary proceedings, Google FLEURS benchmarks, internal business meetings, lectures, and broadcast talk shows. The core infrastructure stack:

• NVIDIA H100 GPUs for model training and fine-tuning
• NVIDIA L40S GPUs for production inference, deployed on premise or in hybrid configurations with GCP and AWS
• NVIDIA NeMo™ ClusterDiarizer for speaker diarization, combining Silero VAD, TitaNet-Large speaker embeddings, and two-stage agglomerative clustering
• Bidirectional gRPC streaming for sub-300 ms interim transcription results
• A punctuation and capitalization model to post-process raw output into structured, readable meeting records

Rafiqspace.ai partnered with Lintasarta—NVIDIA’s only Cloud Partner in Indonesia—to provide the on-premise infrastructure backbone for government deployments, ensuring that sensitive parliamentary and health records remain in controlled environments.

Rafiqspace.ai’s NVIDIA-powered platform delivered step-change improvements in speed, accuracy, and cost—enabling government clients to receive same-day meeting minutes where turnaround previously took 2–3 days, and reducing manual transcription effort by 95%.

On accuracy, Rafiqspace.ai’s fine-tuned Parakeet model set a new benchmark for Bahasa Indonesia speech recognition—reaching 97.7% on its test set and outperforming OpenAI Whisper, Google Gemini 2.5 Pro, ElevenLabs Scribe, and Alibaba Qwen3. For government meeting documentation specifically, accuracy came in at 95.57% versus 91.78% for Google Gemini Pro—a margin that directly affects the reliability of official parliamentary records.

The efficiency and cost gains were equally significant:

• 569x real-time batch transcription on H100—up from a 6x–10x baseline
• 1.65x real-time streaming, up from 0.75x, enabling live note-taking that keeps pace with speakers
• Per-hour transcription cost reduced from up to $50 (outsourced) to approximately $5—a 90% reduction

Rafiqspace.ai is evaluating NVIDIA Nemotron 3 Nano for LLM-powered regulatory insight and compliance report generation, and plans to integrate NVIDIA AI Enterprise and NVIDIA Run:ai for infrastructure optimization as deployments scale. The roadmap targets an end-to-end agentic compliance platform combining transcription, diarization, and autonomous document generation.

The company plans to launch B2C services in 2026 and develop industry-specific models for finance, legal, and healthcare sectors. Longer term, Rafiqspace.ai aims to expand to additional Indonesian regional languages and ASEAN markets—bringing compliance-grade Bahasa Indonesia speech AI to Malaysia, Thailand, and the Philippines.