MUNEEB

MUNEEB

Applied AI Engineer

MUHAMMAD MUNEEB

Building production AI systems across agentic AI, voice AI, RAG, real-time infrastructure and autonomous workflows.

I build AI systems that do things, not just chat.

View WorkAsk the agent↘

● Open to AI engineering roles

PARTICLES ······

RENDER GPU / GLSL

AGENTIC AIVOICE AIRAGON-DEVICE AILLM INFRASTRUCTUREAI AUTOMATIONREAL-TIME BACKENDSMULTIMODAL AIMODEL RESEARCHPRODUCTION DEPLOYMENTAGENTIC AIVOICE AIRAGON-DEVICE AILLM INFRASTRUCTUREAI AUTOMATIONREAL-TIME BACKENDSMULTIMODAL AIMODEL RESEARCHPRODUCTION DEPLOYMENT
02 / RUNTIME

Reliability by architecture

AXON's execution graph. A ~1B model plans on-device, but nothing it produces is trusted: actions are grammar-constrained, preconditions are gated, and results are deterministically verified. Successful traces compile into skills that replay with zero model calls.

INTENTnatural languagePLANNERGBNF constrainedEXECUTORprecondition gateVERIFIERdeterministic checkSKILL COMPILERtrace → skillLOCAL LLMGemma 3 1B Q4_K_MANDROIDAccessibilityServiceMEMORYSQLite · persistsPASSFAIL · REPLANLEARNED SKILL REPLAY

← swipe the diagram →

TRACEtask receivedSTEP 01 / 08

66 s

Cold LLM plan

2.3 s

Learned replay

0

Model calls on replay

235

Automated tests

03 / AGENT

Ask the portfolio

A retrieval agent running entirely in your browser over an authored corpus of Muneeb's work. It ranks documents, grounds on the best match, and refuses anything below threshold rather than improvising — the same domain-restriction pattern he ships in production RAG.

muneeb://agent — grounded retrieval — 5 stage pipelineIDLE

● agent ready · 22 documents indexed

Ask about AXON, the voice pipelines, RAG, the research, the stack, or how to reach him.

❯
04 / INDEX

Selected work

15 systems across agentic runtimes, voice pipelines, retrieval, research and product. Hover an entry to inspect it.

B2B Communication↗

Offline-first peer-to-peer messaging over the local network.

Weather GUI↗

Desktop weather application built in C++ / Qt.

All repositories↗

github.com/Muneeb50702

RESEARCH

AXON

Offline on-device agentic runtime

A fully offline agentic runtime that lets a ~1B-parameter model operate an Android phone with zero cloud inference.

AXON makes small on-device models reliable through system architecture rather than scale. Grammar-constrained decoding prevents malformed actions, a precondition gate verifies UI elements exist before anything executes, and a deterministic verifier confirms the intended result actually happened. Successful traces compile into reusable deterministic skills that replay later with zero model calls.

66 s

Cold LLM plan

2.3 s

Learned skill replay

0

Model calls on replay

235

Automated tests

Gemma 3 1B Q4_K_Mllama.cppKotlin MultiplatformGBNFAccessibilityServiceShizukuSQLiteJetpack ComposeJNI
SOURCE ↗
05 / TRACK

Where he's built it

Two years of shipping AI into production environments where it has to keep working after the demo.

SocioOn Ltd

Jan 2026 — Present

AI Engineer

Configurable AI receptionists, WhatsApp business automation and connected AI assistants in production.

  • 01Building a universal prompt-configurable AI receptionist platform that generates domain-specific receptionist behaviour from minimal business input and ships immediately testable voice agents.
  • 02Engineering low-latency real-time voice-to-voice pipelines on LiveKit with Deepgram streaming STT, GPT-4o mini reasoning and Cartesia / ElevenLabs speech synthesis.
  • 03Integrating HubSpot, OpenDental, Calendly, Cal.com, Google Calendar, Twilio and Telnyx for scheduling, phone provisioning, CRM logging and call-recording workflows.
  • 04Developing WhatsApp business automation on the Meta WhatsApp Cloud API, with Baileys-based WhatsApp Web onboarding for rapid prototyping and client activation.
  • 05Developing Nexivo AI — a Flutter-based unified AI assistant architecture orchestrating workflows across connected apps, social platforms and productivity services.
PythonFastAPILiveKitDeepgramGPT-4o miniCartesiaElevenLabsTwilioTelnyxLangGraphPostgreSQLDockerFlutter

Devsinn Technologies

Jan 2025 — Jan 2026

AI Engineer

Production RAG systems, real-time streaming AI backends and constrained tool-calling agents for fintech.

  • 01Designed and deployed production-grade RAG systems using LangChain, ChromaDB, Hugging Face embeddings and OpenAI APIs.
  • 02Built real-time AI backends with FastAPI, WebSockets and Socket.IO for streaming LLM responses and interactive agent workflows.
  • 03Developed tool-calling agentic systems for fintech automation, including constrained actions for payments, invoices, memos and scheduled transactions.
  • 04Optimized local LLM inference with GGUF and 4-bit quantization; deployed Dockerized AI microservices to cloud and VPS environments.
  • 05Implemented OCR and document-automation pipelines with PaddleOCR and Tesseract, plus PostgreSQL / SQLAlchemy storage for conversations, embeddings and AI configuration.
PythonFastAPILangChainChromaDBHugging FaceOpenAIMistralLLaMAGGUFDockerPostgreSQLSocket.IOPaddleOCR
06 / RESEARCH

Understanding the machine

Production work answers whether a system behaves. Research asks why. Two tracks: transformer internals built from first principles, and multimodal bridges that push speech into a frozen LLM's token space.

◆ AXON CORE

PROCEDURAL · BLENDER 4.5

GLB / DRACO · 4,506 VERTS

Mechanistic interpretability

He's implementing a decoder-only GPT-2 from first principles in PyTorch — every component written by hand so nothing is hidden behind an abstraction. The point isn't the model; it's the instrumentation: exposed activations, logit lens, activation patching, attribution, head ablation, residual-stream analysis and induction-head experiments.

Speech into token space

A two-stage Q-Former cross-attention bridge aligning Arabic speech with a frozen LLaMA-3-8B, pretrained on 128+ hours with contrastive, matching and language-modeling objectives — then evaluated on WER/CER, hallucination behaviour and cross-dataset generalization.

0.6396

Validation loss

Arabic speech adapter

0.72

Semantic similarity

vs. Whisper baseline

17k+

Evaluation samples

cross-dataset

128 h

Speech pretraining

NVIDIA A40 · FP16

Active interests

Mechanistic InterpretabilityTransformer InternalsSmall Language ModelsOn-device AIModel QuantizationAgent ReliabilityConstrained DecodingAI Safety & GuardrailsAgent VerificationMultimodal AISpeech-to-LLMRAG EvaluationReal-time AI InfrastructureAutonomous Systems
READ THE CODE↗

01I'm an Applied AI Engineer based in Lahore, Pakistan, currently at SocioOn Ltd while completing my BS in Computer Science at Superior University.

02I build production-oriented AI systems across agentic AI, Voice AI, Retrieval-Augmented Generation, real-time backend infrastructure, multimodal systems and business automation — low-latency LiveKit voice agents, FastAPI and WebSocket backends, LangChain/LangGraph agents, PostgreSQL-backed AI applications, WhatsApp automation, OCR pipelines, vector search and optimized local LLM inference.

03My research interests go deeper into how these systems actually work. I'm building transformer architectures from scratch for mechanistic interpretability, and developing AXON — an offline on-device agent runtime exploring how small local language models can reliably drive Android through constrained decoding, deterministic verification and learned skill replay.

07 / STACK

Capability matrix

65 tools, frameworks and techniques currently in rotation across six domains.

Agentic AITool CallingLangChainLangGraphConstrained DecodingPrompt EngineeringLLM OrchestrationModel Context ProtocolLocal InferenceQuantization
08 / CONTACT

LET'S BUILD

Open to AI engineering roles, freelance builds & research collaboration. If you're working on agentic systems, voice infrastructure, retrieval at scale — or something that hasn't been built yet — he'd like to hear about it.

Start a conversationDownload CV ↗
GitHub@Muneeb50702↗LinkedInmuhammad-muneeb-yaseen↗WhatsAppMessage directly↗Curriculum VitaeDownload PDF↗
Based inLahore, Pakistan--:--:-- PKT