Kaleem Ullah Q.

Most agent projects break in the same place: nobody measured whether the thing actually works until users found the bug. I build agent, fine-tuning and retrieval systems, then instrument them so failure rates show up in your dashboard, not your support inbox. ⭐ 100% Job Success | 20+ projects | Every review 5 stars ⭐ PhD in AI. Peer-reviewed papers in IEEE T-ITS, Information Sciences, JAIR and IEEE INFOCOM | 48 citations ⭐ Lead researcher on a Huawei Technologies research contract (recursive reasoning agents) ⭐ One client: "went so far above and beyond, I would give higher than 10 if possible" ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ❓ Your agent works in testing and falls apart on real users? ❓ Need to know if fine-tuning beats prompting before you pay for GPUs? ❓ Sitting on a document pile or a database nobody can query? ❓ Need it self-hosted, with no data leaving your infrastructure? 👉 If yes, send me the details. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 & 𝗠𝘂𝗹𝘁𝗶-𝗔𝗴𝗲𝗻𝘁 𝗦𝘆𝘀𝘁𝗲𝗺𝘀 Planning loops, tool use, memory and orchestration that recovers from failure instead of hanging. Built multi-agent systems hitting 87% human-quality scores with 65% faster production, cut agent API latency 40%, raised routing accuracy 35%. ✅ MCP, OpenAI Agents SDK, LangGraph, Pydantic AI, smolagents, Google ADK ✅ Tool calling, routing, retry policies, human-in-the-loop checkpoints ✅ Multi-agent reasoning is my published research (IEEE T-ITS), not a tutorial I followed 𝗟𝗟𝗠 𝗘𝘃𝗮𝗹𝘀 & 𝗔𝗴𝗲𝗻𝘁 𝗥𝗲𝗹𝗶𝗮𝗯𝗶𝗹𝗶𝘁𝘆 The part most builds skip. Task-specific eval sets, regression gates in CI, LLM-as-judge with human calibration, hallucination and memory-faithfulness auditing. ✅ lm-evaluation-harness, OpenCompass, custom eval harnesses, Weights & Biases ✅ Published a causal audit of agent memory: 4,200 instances, 7 models, 3 memory systems ✅ You get a number for "is it working", not a vibe 𝗙𝗶𝗻𝗲-𝗧𝘂𝗻𝗶𝗻𝗴 & 𝗣𝗼𝘀𝘁-𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 (𝗥𝗟𝗛𝗙 / 𝗚𝗥𝗣𝗢) For when prompting stops paying off. SFT, LoRA/QLoRA, preference optimization (DPO/ORPO), GRPO and verifiable-reward RL, reward modeling. I will also tell you when fine-tuning is the wrong answer. ✅ TRL, verl, OpenRLHF, PEFT, DeepSpeed, FSDP ✅ +25% task accuracy from domain fine-tunes of open-weight models ✅ Certified in GRPO reinforcement fine-tuning and LLM post-training (DeepLearning.AI) 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 & 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹 (𝗥𝗔𝗚) Hybrid retrieval (BM25 + vector), reranking, context compression, citation-grounded answers. 10K+ documents a month at 92% answer accuracy on domain queries. Cut one client's API bill 40% through caching and context compression. ✅ Qdrant, Weaviate, Chroma, FAISS, Pinecone | LlamaIndex, LangChain ✅ Source citations and hallucination guardrails as standard 𝗡𝗟𝟮𝗦𝗤𝗟 & 𝗗𝗮𝘁𝗮 𝗜𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲𝘀 Plain-English querying over your database, for teams who do not write SQL. ✅ 5 minutes of manual SQL → 10 seconds automated | 94% intent routing accuracy ✅ Client: "Revolutionized our SQL Chat Agent's performance with 79% faster response times" 𝗦𝗲𝗹𝗳-𝗛𝗼𝘀𝘁𝗲𝗱 & 𝗢𝗻-𝗣𝗿𝗲𝗺 𝗗𝗲𝗽𝗹𝗼𝘆𝗺𝗲𝗻𝘁 Nothing leaves your infrastructure. vLLM serving, quantization, GPU and NPU inference tuning, containerized stacks with test suites so regressions surface before release. ✅ vLLM, Hugging Face, Docker, Kubernetes, AWS, Huawei Ascend ✅ Frontier APIs (Claude, GPT, Gemini) and open weights (Llama, Qwen, DeepSeek, Mistral) ✅ GDPR-compliant private deployments 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 𝗥𝗲𝘃𝗶𝗲𝘄 & 𝗙𝗲𝗮𝘀𝗶𝗯𝗶𝗹𝗶𝘁𝘆 Before you spend six months on the wrong approach. ✅ Client: "His insights into the feasibility of our approach and feedback on the tech stack were incredibly valuable" ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 𝗕𝗲𝘀𝘁 𝗙𝗶𝘁 Teams with a defined AI problem who need production-grade execution and evidence that it works. If you are still exploring "maybe we need AI", I am probably not your match. If you know what you need, message me with the details. #AIAgents #AgenticAI #LLM #LLMEngineer #MultiAgent #MCP #LangGraph #AIEvals #LLMEvaluation #FineTuning #RLHF #GRPO #LoRA #vLLM #ContextEngineering #RAG #NL2SQL #AIIntegration #AIAutomation #Qdrant #SelfHostedAI #OpenSourceLLM #AIConsulting #GenerativeAI #Python