Bilal Ahmed

Bilal Ahmed

Senior AI Engineer specializing in Agentic AI, LLM Orchestration, and Real-Time Voice Systems.

Currently: Senior AI Engineer, Product Lead, RTC League (Pvt) LtdExperience: 6 years

Hire me

An experienced AI engineer with over 5 years of expertise in building production-grade voice and conversational AI platforms handling 40,000 calls daily. He specializes in real-time speech pipelines, multi-agent orchestration, and GPU-efficient inference at scale for enterprise clients.

Experience

Senior AI Engineer, Product Lead

RTC League (Pvt) LtdworkJun 2025 – Present

Summary

Architected TelEcho, a production AI voice agent platform handling ∼40,000 inbound and outbound calls/day across enterprise and on-premise deployments. Lead a cross-functional team of 7 engineers and own client delivery for 10+ enterprise clients: gathering requirements, guiding the team through building each client's agents, deploying them on client servers, and acting as the clients' direct point of contact for production issues.

What I did

  • Cut agent response latency from 2,500 ms to under 1,000 ms through model selection, tuned chunking, partial transcripts, speculative LLM responses, streaming, and caching, without losing ASR accuracy.
  • Built real-time agent pipelines on LiveKit Agents: streaming ASR (Deepgram, Whisper) to LLM orchestration (tool calling, RAG) to streaming TTS, with custom VAD, interruption handling, and SmartTurn v3 inspired turn-taking in progress.
  • Built the multilingual TTS pipeline for Arabic, Urdu, and English, with code-switching mid-conversation for mixed-language calls.
  • Built a production RAG pipeline on Qdrant with OpenAI embeddings for sub-second multi-tenant retrieval, with Redis caching to cut redundant LLM calls.
  • Shipped an MCP server with OAuth 2.0/PKCE authentication, published as a public connector that lets Claude and ChatGPT drive the platform in natural language.
  • Improved speech recognition on narrowband telephony audio with a custom denoiser, fine-tuned background voice cancellation (BVC) models, and a bandwidth-extension model, lowering WER across STT providers.
  • Trained intent, sentiment, and churn-risk models on transcribed calls, feeding scores into the CRM to adjust agent prompts and escalation rules.
  • Built call recording on LiveKit Egress to S3, indexed in Postgres against CRM records, feeding a batch pipeline (ASR, diarization, data warehouse) for search and topic detection across millions of calls.
  • Designed multi-tenant agent configuration, giving each tenant its own prompts, tools, and isolated LiveKit rooms, routed by dialed number.
  • Connected the agents to the phone network through a LiveKit to SIP bridge (Twilio, Telnyx, Plivo) with carrier failover and call control (transfer, hold, conference, voicemail drop), reducing dropped calls by ∼30% versus the legacy IVR.
  • Building an outbound call scheduler for automated campaign dialing.
  • Developed a system for transferring calls between specialized agents while maintaining context and real-time conversation speed.

Results

  • Reduced dropped calls by ∼30% versus the legacy IVR.
  • Scaled platform to handle ∼40,000 inbound and outbound calls/day.
PythonLiveKit AgentsDeepgramOpenAIMCPQdrantRedisPostgreSQLS3SIPWhisperOAuth 2.0PKCETwilioTelnyxPlivoLiveKit EgressNative structured outputsStrict mode

Software Engineer, ML & Conversational AI

Afiniti Software Solutions (Pvt) LtdworkMar 2023 – Jun 2025

Summary

Fine-tuned Whisper STT with LoRA on domain audio, cutting WER from 40% to 18% (55% relative improvement) with minimal GPU memory overhead, and applied LoRA/QLoRA to open-source LLMs for support, internal knowledge, and sales assistants, cutting their inference cost 40-60% via AWQ/GGUF quantization and vLLM/Triton serving.

What I did

  • Fine-tuned VITS and XTTS v2 on domain audio for production TTS, improving naturalness while holding a real-time factor suitable for live calls.
  • Built an audio preprocessing pipeline over raw call recordings: VAD-based segmentation, SNR filtering, speaker diarization, noise augmentation with room impulse responses, and automated transcript alignment, producing clean training pairs for TTS and STT.
  • Cut inference latency by ∼35% with no measurable accuracy loss using INT8/FP16 quantization and distillation, profiling with PyTorch Profiler and Nsight Systems and cutting peak GPU memory via paged attention and an FP16 KV-cache to serve on smaller instances.
  • Designed multi-agent orchestration with RAG, webhooks, and tool integrations, implementing MCP, function calling, and agent hand-offs within the OpenAI Realtime framework.
  • Designed a GPU-aware load balancer and multi-threaded inference engine for multi-GPU Kubernetes clusters, adding request queuing with backpressure and autoscaling on GPU utilization to absorb traffic spikes.
  • Tuned vLLM serving under concurrent load (continuous batching, prefill chunking, request scheduling), tracking P50/P95/P99 latency in Prometheus and Grafana.
  • Built the NLU layer for conversational AI (intent classification, NER, and dialogue state management) on fine-tuned BERT-family models.
  • Built gRPC-based microservices for streaming low-latency audio in live customer calls.
  • Automated ML delivery with CI/CD pipelines covering model validation, MLflow tracking, Docker/Triton inference, and staged A/B rollouts.

Results

  • Cutting inference cost 40-60% via AWQ/GGUF quantization and vLLM/Triton serving.
  • Cut peak GPU memory via paged attention and an FP16 KV-cache.
WhisperVITSXTTS v2LoRAQLoRAvLLMAWQGGUFBERTOpenAI RealtimeMCPNsightPrometheusGrafanaKubernetesTritonPyTorch ProfilerINT8FP16MLflowDockergRPCC++Android NDKJNI

Real Time Voice AI Engineer

Younite (Pvt) LtdworkMay 2022 – Feb 2023

Summary

Led development of ML models for audio signal processing (noise reduction, dereverberation, speech enhancement, and source separation) in a decentralized smartphone-based conferencing product, improving noise reduction 15% over the existing pipeline.

What I did

  • Built an AI-driven Voice Activity Detector to remove echo residuals and non-speech segments.
  • Combined DSP algorithms (AEC, AGC, NS, DRC) with ML for adaptive audio processing, and added AI-based stream selection to improve bandwidth utilization.
  • Cut CPU and memory usage 35% by optimizing audio algorithms and libraries on iOS and Android.
  • Researched and developed backend cloud architectures spanning media servers (Janus, Kurento, Pion ion-sfu), SIP servers (FreeSWITCH, Kamailio, FreePBX, Twilio), and SIP-WebRTC gateways.
Deep learningWebRTCJanusKurentoFreeSWITCHKamailioiOSAndroidPion ion-sfuFreePBXTwilioAECAGCNSDRC

Skills

Technical

Whisper
PythonPython
DSP
LiveKit
WebRTC
VITS/XTTS v2
LoRA/QLoRA
vLLM
Multi-agent orchestration
Deepgram
SIP/VoIP
RAG
Qdrant
MCP servers
FastAPIFastAPI
PyTorchPyTorch
C/C++C/C++
TypeScript/JavaScriptTypeScript/JavaScript
Node.jsNode.js
LangChain
Triton
KubernetesKubernetes
TensorFlowTensorFlow
AWS Bedrock
GoGo
Pydantic
Multi-agent call flow
Autonomous planning
Tool Use Orchestration
C++C++
Android NDK
ONNX

Languages

Urdu
English
Arabic

Projects

VoicePing

personal
A latency profiler for voice agents that runs test calls against WebSocket/HTTP endpoints and displays p50, p95, and p99 latency on a live dashboard.
FastAPIPydantic

TelEcho

Architect & Product Leadprofessional
A production AI voice agent platform handling ∼40,000 inbound and outbound calls/day across enterprise and on-premise deployments. Features include real-time streaming ASR to LLM orchestration to streaming TTS, custom VAD, and interruption handling.
PythonLiveKit AgentsDeepgramOpenAIMCPQdrantRedisPostgreSQLS3SIPOAuth 2.0 with PKCEPydantic

Education

BS Electrical Engineering

Balochistan University of Information Technology, Engineering and Management SciencesComputer Engineering Major2016 – 2020
Hire BilalReach out about a role, a contract or a conversation.For recruiters

Bilal's twin is AI, it can make mistakes.

More people on Kanvis

Browse more peopleCreate your own Kanvis