Agentic RAG & LLM Systems
Hybrid retrieval pipelines, multi-pass reasoning, prompt orchestration and provider-agnostic LLM gateways that deliver citation-backed answers and cut hallucinations.
Qdrant · Cohere · OpenRouter
Machine Learning Engineer · Islamabad, Pakistan
Muhammad
Saad Hasan
I build intelligent systems and chase the frontier where science meets technology — currently engineering agentic RAG platforms, multimodal AI and computer-vision systems at Levitate Data.
- 4+
- Years building AI
- 2
- Enterprise AI platforms
- 5
- Novel architectures
- 34+
- Open-source repositories
I. About
Engineering AI with the rigour of research and the discipline of production.
I'm a Machine Learning Engineer based in Islamabad, Pakistan, and a Computer Science graduate of FAST (NUCES). My work lives at the intersection of applied AI engineering and hands-on research — I love taking an idea from a paper or a blank repository all the way to systems that real businesses depend on.
At Levitate Data I work in the core AI engineering team behind Thalamus, an enterprise-grade agentic RAG platform, and I led the end-to-end architecture of Aletheia, a multimodal analysis engine that serves as the AI brain behind Cognition Legal. Alongside this, I dig into deep-learning research in medical imaging — rebuilding novel architectures from scratch, training them, and improving on them until they rival far heavier models.
What really drives me is curiosity. I find it astonishing that, by decoding the fundamental forces of nature, we eventually understood enough to build artificial intelligence itself — drawing inspiration from the architecture of the human brain. Chasing that thread, from first principles to working systems, is what keeps me building every day.
- Focus
- Agentic RAG · Multimodal AI · Computer Vision · DL Research
- Stack
- Python · PyTorch · TensorFlow · FastAPI · AWS
- Education
- BSc Computer Science, FAST NUCES (2021–2025)
- Based in
- Islamabad, Pakistan — open to remote
II. What I do
Four disciplines, one engineer.
Multimodal & Computer Vision
Video, audio and image understanding — object detection, tracking, OCR and face recognition — wired into scalable, event-driven processing pipelines.
YOLO · OpenCV · Gemini
Deep-Learning Research
Novel, parameter-efficient architectures — KANs, spectral operators, knowledge distillation — benchmarked rigorously on real medical-imaging datasets.
PyTorch · TensorFlow · KAN · FNO
Full-Stack & Cloud
From React frontends to FastAPI backends and Infrastructure-as-Code on AWS — shipping multi-tenant, fault-tolerant platforms with CI/CD and observability.
React · FastAPI · AWS · Docker
III. Experience
A record across product, research & engineering.
-
Sept 2024 — Present
Machine Learning Engineer
Levitate Data · Islamabad, PK
Part of the core AI engineering team behind two flagship enterprise products — Thalamus (agentic RAG) and Aletheia (multimodal analysis) — for heavily regulated industries; I led the end-to-end architecture of Aletheia.
- Architected a zero-data-loss, multimodal ingestion engine with state-of-the-art parsers (DeepSeek OCR via Vertex AI, Chandra Datalab, Unstructured.io) and dynamic fallback mechanisms.
- Designed a high-precision hybrid retrieval pipeline on Qdrant with Cohere re-ranking, autonomous prompt rewriting and dynamic system-prompt injection.
- Built a provider-agnostic LLM gateway (OpenRouter) unifying OpenAI, Anthropic, Gemini & Mistral, plus an event-driven Celery architecture with autoscaling.
- Hardened compliance with BYOK encryption, API-key rotation, schema versioning, rate limiting and a monetization control plane.
Python · FastAPI · Qdrant · Celery · AWS · Docker
-
Aug 2025 — Sept 2025
Associate AI Engineer
Amanat · Islamabad, PK
- Built an AI-powered lost-and-found app using smart image & text matching for secure item recovery.
- Improved model accuracy and system scalability for faster, more reliable matching.
-
Jan 2025 — Apr 2025
Full-Stack Engineer
MRD Society · Islamabad, PK
- Led end-to-end development of a platform for managing & publishing academic content.
- Built secure REST APIs (Python) and a scalable MySQL schema; React frontend; CI/CD on Git.
-
June 2024 — Nov 2024
Software Engineering Intern
Intrace (APIGen Inc) · Boulder, Colorado
- Contributed to APIGen — generating custom APIs from natural-language prompts (React, Python, FastAPI).
- Built PostgreSQL-backed REST APIs handling 1k+ RPM; shipped CI/CD with 95% test coverage.
-
June 2024 — July 2024
Backend Engineer Intern
Merik Solutions · Islamabad, PK
- Developed backend services with the .NET framework and rigorous test validation.
-
June 2023 — July 2023
ML & Computer Vision Intern
DataInsight Research Lab · Islamabad, PK
- On a fully funded research project, digitised doctors' handwritten prescriptions (Python, TensorFlow, OpenCV).
- Built a YOLO + image-processing OCR pipeline reaching 92% accuracy; integrated TrOCR & PaddleOCR.
IV. Featured Work
Products & systems I've shipped.
I
Live · Enterprise SaaS
Levitate Data · Core AI Engineering Team
Thalamus
Agentic RAG-as-a-Service — “the end of endless searching.”
An enterprise-grade, multi-tenant platform that turns complex, unstructured data into reliable, auditable answers for regulated industries. I architected the multimodal ingestion engine, the hybrid retrieval pipeline and the provider-agnostic LLM gateway powering it.
- 66%
- lower TCO vs in-house
- 6 mo
- faster to production
- Day 1
- production reliability
Qdrant · Cohere · OpenRouter · Celery · BYOK
Visit Thalamus
II
AI brain behind Cognition Legal
Levitate Data · Architect & Lead
Aletheia
Multimodal AI & behavioural-intelligence engine.
A multi-tenant platform for deep, automated processing of video, audio and text — decoding facial expression, tone and language. It powers context-aware chat and deposition analysis for Cognition Legal, coupling Google Gemini's multimodal models with the Thalamus RAG engine.
- 3
- modalities: video · audio · text
- IaC
- AWS CloudFormation · ECS
- Multi-tenant
- strict isolation
FastAPI · Celery · Gemini · ECS · Docker
See it in Cognition Legal
III
Computer Vision · Full-Stack
Flagship Project
TrashCam
Real-time littering detection & offender identification.
An end-to-end system that detects littering in real time, tracks offenders across frames and identifies them — built on a 220K-image dataset, with automatic LLM-generated incident reports and a full React + FastAPI + PostgreSQL stack.
- 0.89
- mAP@0.5 (YOLOv12x + SAHI)
- 98%
- offender ID accuracy
- 220K
- images collected
YOLOv12x · SAHI · DeepSORT · TrOCR · FaceNet · Qwen · Ollama
View the repositoryV. Research & Development
Original deep-learning research, benchmarked hard.
A body of work on parameter-efficient architectures for multi-modal medical imaging — implementing, improving and inventing methods that match heavy baselines at a fraction of the cost.
Spectral-KAN
Kolmogorov–Arnold Networks × Fourier Neural Operators
A novel architecture using custom ChebyshevConv2D layers (learnable degree-3 polynomials instead of linear weights) and a Global Spectral Gating mechanism built on 2D FFTs with learnable complex frequency filters — for global spatial routing across X-ray, MRI and endoscopy.
- ~413K
- parameters
- 93.43%
- test accuracy
- 0.94
- weighted F1
Dynamic Ensemble Knowledge Distillation
Multi-modal medical classification
A 1.9M-parameter MobileNetV2 student with Coordinate Attention, distilled from an EfficientNetV2B0 + ResNet50V2 ensemble using entropy-based dynamic weighting and MixUp.
- 1.9M
- params
- 97.56%
- accuracy
- >0.99
- ROC-AUC
KDLight
Lightweight knowledge distillation · 14 conditions
A compressed CNN distilled from a MobileNetV1 teacher via composite hard + temperature-scaled KL-divergence loss, profiled rigorously for real-world deployment.
- ~90%
- less memory
- 4×
- faster inference
- 86.7%
- accuracy
U-Net + Few-Shot TB Classification
Two-stage segmentation & meta-learning
A 7.7M-parameter U-Net for lung-field segmentation feeding a Prototypical Network (MobileNetV2) for 2-way, 5-shot tuberculosis classification from scarce data.
- 0.96
- Dice score
- 98%
- val accuracy
- 5-shot
- meta-learning
Evolutionary MobileNet-ConvLSTM2D
Hybrid net · evolutionary kernel search
A hybrid MobileNet + Inception + ConvLSTM2D classifier across 9 medical conditions, with a novel “Chain Foraging Cyclone Aging” evolutionary algorithm for dynamic kernel-size optimization.
- 94%
- test accuracy
- 9
- conditions
- 7K+
- images
VI. Projects
Selected builds across the stack.
A catalogue of my open-source AI work — generative models, computer vision, NLP, classical ML and classic-AI search — from my GitHub.
- 01 DCGAN — Image GenerationDeep Convolutional GAN generating realistic images by learning the data distribution. Generative
- 02 GAN — Image GenerationA foundational Generative Adversarial Network trained with adversarial learning. Generative
- 03 RaLSGAN — Image GenerationRelativistic Average LSGAN for more stable, realistic image synthesis. Generative
- 04 VAE — Image GenerationVariational Autoencoder learning a latent space to generate and reconstruct images. Generative
- 05 Conditional GAN — Sketch2FaceConditional GAN (pix2pix) turning hand-drawn sketches into realistic faces. Generative
- 06 CycleGAN — Face2SketchCycleGAN learning bidirectional face to sketch translation. Generative
- 07 Custom GAN — Fake ImagesPyTorch GAN generating synthetic cat images from CIFAR-10. Generative
- 08 GAN — Signature SynthesisGAN synthesizing realistic signatures for forgery detection and augmentation. Generative
- 09 VAE — Signature SynthesisVariational Autoencoder generating synthetic signatures via a learned latent space. Generative
- 10 Custom GANA custom Generative Adversarial Network implementation for image synthesis. Generative
- 11 Vision Transformer — CIFAR-10Vision Transformer treating image patches as tokens for CIFAR-10 classification. Vision
- 12 ResNet — ClassificationResidual network with skip connections for deep image classification. Vision
- 13 CNN — ClassificationConvolutional Neural Network for image classification across datasets. Vision
- 14 Signature ClassificationSignature classification and forgery detection with OpenCV, TensorFlow and sklearn. Vision
- 15 Dermoscopy Image ProcessingSkin-lesion detection via K-Means clustering and connected-components labelling. Vision
- 16 U-Net + Few-Shot TBU-Net lung segmentation feeding a MobileNetV2 few-shot tuberculosis classifier. Vision
- 17 Multimodal FusionFusing LSTM, BERT, CNN and ResNet to combine language and vision signals. Vision
- 18 TrashCam — AppReal-time littering-detection app: YOLO detection, tracking, OCR and face ID. Vision
- 19 Urdu to English TranslationSeq2Seq attention RNN for Urdu to English machine translation in PyTorch. NLP
- 20 English to Urdu TranslationEnglish to Urdu translation with Transformer and LSTM models on UMC005. NLP
- 21 Word2Vec Spam DetectionSpam detection with Word2Vec embeddings and logistic-regression classification. NLP
- 22 NLP Text ClassificationText classification and prediction applying NLP with logistic regression. NLP
- 23 Sentiment & Text PredictionNext-word prediction and movie-review sentiment with deep learning models. NLP
- 24 Spam Detection NNEmail spam classifier built on a custom two-layer neural network. NLP
- 25 ANN — ClassificationArtificial Neural Network built for classification tasks. Classical ML
- 26 Logistic Regression & SVMLogistic Regression and SVM with full EDA and feature scaling. Classical ML
- 27 Linear Regression from ScratchLinear regression implemented from scratch with preprocessing and evaluation. Classical ML
- 28 Voice Customer-Service AgentVoice-enabled conversational AI agent for automated customer service. AI & Search
- 29 Chat WidgetEmbeddable chat-widget front-end for conversational AI assistants. AI & Search
- 30 Genetic Timetable SchedulerGenetic algorithm generating conflict-free university timetables. AI & Search
- 31 Minimax Tic-Tac-Toe AIOptimal Tic-Tac-Toe AI using Minimax with Alpha-Beta pruning. AI & Search
- 32 Sudoku & Magic-Square SolverSudoku solver (backtracking + AC-3) and Magic-Square solver (genetic). AI & Search
- 33 Vertex-Ordering SearchOptimal vertex ordering via BFS, DFS, UCS and Greedy search. AI & Search
- 34 Neural Network on an OSA neural network executed on a multi-core OS with processes, threads and IPC. AI & Search
VII. Toolkit
Technologies I work with.
- Languages
- Python · C / C++ · JavaScript · TypeScript · Java · SQL · C#
- AI / ML
- PyTorch · TensorFlow · OpenCV · YOLO · BERT · TrOCR · Scikit-learn · Pandas · Qwen · Ollama
- Backend
- FastAPI · .NET / ASP.NET · REST APIs · PostgreSQL · MySQL · Celery · Redis
- Frontend
- React.js · HTML / CSS · JavaFX
- Cloud & Tools
- AWS (S3 · ECS) · CloudFormation · Docker · Qdrant · Cohere · OpenRouter · Git · CI/CD
- Concepts
- Agentic RAG · Generative AI · Computer Vision · NLP · Knowledge Distillation · Neural Networks
VIII. Beyond the Code
Off the clock — debater, sportsman & musician.
- 14
- Gold medals
- 7
- Silver medals
- Pakistan
- Selected to represent
Competitive Debate
A Champion Debater at the World Scholar's Cup — 1st place in both Debates & Team — alongside Champion Scholar and multiple wins for Excellence in Writing & Debate. Across the rounds I earned 14 gold and 7 silver medals, and was elected to represent Pakistan at the Global Rounds.
Sport & Fitness
I train at the gym every day and play football whenever I can. Off the pitch I'm an avid spectator — the NFL, golf, polo, basketball and plenty more.
Music
Away from the screen I'm a musician — I play the bass and guitar, a different (and very welcome) kind of pattern recognition.
IX. Correspondence
Let's build something intelligent.
Have a problem worth solving with AI — or just want to talk shop about RAG, vision or research? My inbox is open.
- /in/msaadx
- GitHub
- /msaadx