I develop scalable systems using AWS (Amazon Web Services)
LLMs, agent harness, agentic and multimodal systems — I work on the latest technology, deploying on AWS platforms (Bedrock, Lambda, SageMaker, S3, ECS) and RunPod GPUs and workers for training and inference.
Machine learning engineer from Pakpattan, Pakistan, now an M.S. student at Chung-Ang University (2025–present), working on AI for wireless systems.
Before graduate school I shipped production AI: voice agents, calling agents, voice clones, transformer models, RAG, and biomedical systems. I still lead AI work and teach on the latest technology — scalable apps on AWS (Amazon Web Services) platforms and heavy jobs on RunPod GPUs, with an agent harness and models via Bedrock and SageMaker.
B.S. Computer Science, The Islamia University of Bahawalpur (2020–2024).
I work on the latest technology — frontier models, AWS platforms, and RunPod GPUs — as of October 2026.
Meta, 12 Oct 2026. Personal and professional agents with Active Memory and Active Focus. 1M-token context; text, image, video, and audio in. Wonder is on the Meta API and Muse Code; Full is for Muse Code subscribers. Artificial Analysis Intelligence Index: Full 56, Wonder 52, Muse Spark 1.3 at 48.
TypeSafe System One model. It does not write text. You pass state plus a typed question; it returns a choice, a score, or a yes/no probability. Trained with RLCD so the confidence is meant to be calibrated. I use this pattern for routing, filtering, and guardrails inside agent loops.
Kimi K3.1 Pro for long coding runs, Claude Opus and Sonnet 5.5 inside Claude Code, GPT including GPT-6.1 Sol for multi-step work, Gemini 4 Carbon for cybersecurity and long-horizon runs, and Gemini 4 Argon for long trajectories. Muse Spark, Qwen, and DeepSeek when the model needs to run or be fine-tuned locally.
A prime agent plans, writes todos, and delegates to subagents inside an agent harness: filesystem memory, tool use, and a human checkpoint. The harness is executable code. OO-Harness grows a class hierarchy and routes each task; Harness² revises that code on the current task, then keeps the lesson. LangGraph and MCP wire the harness to the apps around it.
LLMs for open-ended work; small language models (SLMs) where latency and cost matter. Adaptation with LoRA and QLoRA, PEFT, and efficient fine-tuning plus GPTQ/AWQ quantization so one base model fits many products.
Agents wired as explicit graphs: conditions, tool calls, and typed outputs your code can branch on — not one long chat loop. Same idea for local SLMs on a GPU box or frontier models on Bedrock.
Production traffic on AWS (Amazon Web Services): Bedrock, Lambda, API Gateway, ECS, RDS, DynamoDB, and S3 buckets. Fine-tuning and batch inference on RunPod GPUs when you need dedicated NVIDIA capacity outside the main AWS pipeline.
Wonder-class models draft, code, and plan. Jev-class models pick among options you already defined. Most of my systems call both: a generator for open work, a decision model where the answer must be a type the code can branch on.
October 2026. I keep the model and the harness separate. Kimi K3.1 Pro, Claude Code, Goose, Kimi Code, OpenClaude, and OpenCode all sit on the same bench.
Moonshot, Oct 2026. Strict terminal tasks. Axis starts at 50 so the gap is visible. The number is the score.
Artificial Analysis. Axis starts at 40. K3.1 Pro is 55, up 11 from the previous Kimi. Wonder numbers are Meta’s 12 Oct 2026 report.
Google, 28 Oct 2026. Harder index than the chart beside this one. Axis starts at 40. Carbon is 59, Argon is 50.
The first K3.1 model. About 2.8 trillion parameters, roughly 104 billion active, and a 1 million token window. Post-trained from Kimi K3 for coding, 3D creation, financial research, and long agent runs. Highspeed is the same model at 2–3× speed. On Kimi’s own long-horizon code bench it scores 61.2, against 55.9 for K3. API ids: kimi-k3.1-pro and kimi-k3.1-pro-highspeed.
Moonshot’s coding agent. I select Kimi K3.1 Pro with /model and stay in the terminal: files, shell, and tools. It also speaks other providers, so the same session can move off Kimi when a task wants Claude, GPT, or Gemini. This is the harness I pair with K3.1 Pro for repo work.
Anthropic’s terminal agent for Opus and Sonnet. Read, write, edit, bash, grep, and glob, plus skills, hooks, subagents, and MCP. I use it when the model should stay in the Claude family and the tool loop is the product. The April 2026 OpenClaude study found Sonnet 4.6 lost about 0.19 points moving from Claude Code to an open harness.
Local-first agent, desktop app and terminal, same Rust binary. Extensions are MCP servers. Recipes pack a prompt, extensions, and settings into a workflow I can hand off. Fifteen-plus providers, including Anthropic, OpenAI, Google, Ollama, OpenRouter, Azure, and Bedrock, plus an ACP path so a Claude or Codex subscription can drive it.
One terminal workflow across OpenAI-compatible APIs, Gemini, Codex, Ollama, and other backends: bash, file tools, agents, MCP, and slash commands. I use it when the harness should stay put and the model should change. The Lawrence study is the reason: model choice carried about 88% of the quality variance, the harness about 2%.
OpenCode is the MIT-licensed agent for terminal, desktop, and IDE, with 75+ providers and local models. Next to it I keep Codex CLI, Aider, and Pi. The pattern is the same as Goose and OpenClaude: pick the model for the task, keep the tools, and do not rebuild the loop every time the weights change.
Scalable systems on AWS (Amazon Web Services) platforms, RunPod GPUs for model work, plus data stores and APIs — the latest technology I ship in production.
Each AWS platform service has a clear job; together they form a scalable AWS (Amazon Web Services) stack for AI products.
From frontier LLMs to on-device small language models.
Logical agentic systems, harnesses, and tool protocols.
RAG, hybrid search, and graph-backed retrieval.
App state, analytics, and raw files — often in S3 buckets before embedding.
Services agents and frontends call into.
RunPod GPUs for LoRA/QLoRA runs, batch jobs, and fast experiments; AWS for everything that must stay in your VPC.
Build, deploy, and watch models on AWS platforms and RunPod.
End-to-end AI engineering across product types.
M.S. research at Chung-Ang University. AI on the physical layer of 5G/6G.
Integrated sensing and communication: one waveform carries data and estimates range, velocity, and biosignals from the echo. Current work covers transmit/receive beamforming and sensing for UAV, UAM, and vehicles.
Passive metasurfaces that steer a link. Latest threads are hybrid RIS phase design from channel or direction-of-arrival estimates, and phase mapping that still works once signaling overhead is counted.
STLC gives full spatial diversity with a simple receiver. I follow IRS-aided STLC: the surface and the line code designed together so a low-complexity radio still gets the rate.
CNNs for RIS phase control, Kalman tracking for urban air mobility and LEO, and WiFi sensing. The point is a receiver or scheduler that adapts without a heavy optimization loop on every slot.
Outbound and inbound calling agents on Twilio. Speech in, a tool-using agent in the middle, appointment booking and follow-up on the way out.
Voice agents that stay in a persona. Voice clones give each line a consistent speaker, then the same agent harness handles turn-taking and tools.
AI CSR platform for support teams: chat and voice, ticket routing, and answers grounded in the company docs instead of a script.
Transformer projects across the stack: T5 summarization, DistilBERT next-sentence prediction, and LLaMA chat adapted with QLoRA.
Upper-arm EMG to tip grip, thumbs up, victory, power grip, and rest. Peak, mean, variance, RMS, skewness. 86.3% accuracy for trans-humeral control. IEEE HITE 2024.
ResNet classifier on fundus images with Grad-CAM heatmaps so a clinician can see which retinal regions drove the grade.
Real-time defect detection on a production line: misprints, damage, misalignment. Training, inference, and a small GUI.
Hybrid detector for objects under 32×32 in DOTA, VisDrone, and FAIR1M. Cross-scale fusion and context enhancement. +5.7 mAP, about 28 FPS.
Streamlit app: upload several PDFs, ask questions, keep history. PyPDF2, LangChain, and Gemini.
Quantized low-rank fine-tune of LLaMA for dialogue, so a chat model fits on modest hardware without dropping context.
LangGraph planner, researcher, and writer with shared state and tool use. A long brief comes back as a sourced draft.
Streamlit dashboard for player and team trends. Live app.
October 2026. The harness is code you can evolve, and semantic memory is starting to keep phase as well as amplitude. These are the papers I design against.
Li, Mishra, Ming, Chen, and colleagues, with UIUC. An evolution agent builds harness subclasses from execution records. Inheritance keeps shared behavior; a router uses routing lessons to pick a specialist or keep the base result. Held-out, with a frozen executor, primary scores move from 61.6 to 77.4 on AutomationBench, 25.5 to 39.5 on Workspace-Bench, and 31.2 to 37.5 strict pass on Toolathlon-GYM. That is up to 15.79 points over the base, ahead of Meta-Harness and Agentic Harness Engineering. PDF.
Xu, Chen, CuiZhu, Mishra, Ming, and colleagues. Model weights stay frozen. Task-level recursion proposes harness edits, probes them on the same task, and composes the changes the behavioral contrast supports, with no ground-truth verifier. Harnessing-level recursion stores harnesses and lessons for the next task. One step improves GLM-5 and Gemini 3.5 Flash by up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench, and 11.8 points on LAB partial pass. A larger recursion budget adds more. PDF · Code · Site.
Listopad. A vector becomes a complex waveform, amplitude times a phase: ψ(x) = A(x) eiφ(x). The resonance score measures constructive interference across dimensions and stays in [0, 1]. A sign-to-phase map keeps ordinary embeddings valid — positive components at phase 0, negative at π. Phase-enriched queries improve top-k when the distinction is negation, inversion, or a context shift. Implemented in ResonanceDB. PDF · Code.
Lee, Nair, Zhang, Lee, Khattab, and Finn. The global-harness search that OO-Harness and Harness² are measured against. A coding agent reads prior harness source, scores, and traces from a filesystem. One discovered harness gains 7.7 points on online text classification at 4× fewer context tokens, and 4.7 points on average across five held-out models on 200 IMO-level problems. The other global line is Agentic Harness Engineering (Lin et al., Apr 2026): ten iterations take Terminal-Bench 2 pass@1 from 69.7% to 77.0%.
Ke and colleagues. The harness changes under the agent: 17 streams, 802 tasks, 520 tools, 42 skills, and 62 specialist agents. Expanding tools, skills, or specialists can drop scores on tasks the agent already solved — harness-induced forgetting. Keeping earlier competence and adapting to the new capabilities pull in different directions. PDF.
The 30 Jan 2026 implementation note covers segmented storage, phase-domain sharding, and mmap I/O. Server v2.0.2 adds a write-ahead log and persisted IVF plus Vamana indexes. The index chooses candidates; the resonance kernel still decides relevance. They report up to 68× lower median latency and 0.994 Recall@10. Version 2.1.0 adds context-modulated queries: a weight in [0, 1] on each phase dimension, applied while scoring.
With Naqash Ahmad and Zareena Kausar. Four daily grips plus rest from upper-arm EMG, 86.3% average accuracy.
Feature selection plus SVM, random forest, and ensembles. Glucose, BMI, age, and blood pressure carry most of the signal.
Vol. 3, Issue 6. Cross-scale fusion, context enhancement, adaptive anchors. Article · DOI 10.5281/zenodo.15605836.
Chung-Ang University · Seoul
Wireless systems and AI.
The Islamia University of Bahawalpur
Machine learning and artificial intelligence.
Essenceware Technologies · Rawalpindi
Also program coordinator (2026) and AI instructor (since 2024).
Aridian Technologies
Production ML and agent systems for client delivery.
Helping Hand · Air University · Sparking Asia
AI consulting, neuroimaging research assistance, and analytics.