3.02x
LLM training throughput
Measured gpt_small speedup in WSL2 validation
Chicago based, open to California and remote
Mir Athfan Ali
Full-Stack AI Engineer
I build real AI products end-to-end across mobile, backend, AI agents, vector search, recommendation systems, and ML performance tooling.
Chicago, IL
Open to Chicago, California, and Remote
Product engineering map
Mobile app to AI systems, with measurable performance work.
Mobile
React Native / Expo
Backend
Firebase / FastAPI / Node
AI
OpenAI / LangGraph / retrieval
ML systems
PyTorch / CUDA / ONNX
Wardrobe retrievalAgent reasoningMobile product
const product = await build({
app: "React Native",
backend: ["Firebase", "FastAPI", "PostgreSQL"],
intelligence: ["LangGraph", "vector search", "feedback memory"],
proof: ["TestFlight", "benchmarks", "case studies"],
});Featured results
52.4%
Peak GPU memory reduction
1,120 MB baseline to 533 MB optimized
TestFlight
Mobile product shipped
AURA end-to-end AI styling platform
Full stack
Operating range
Mobile, backend, AI, infrastructure, and ML systems
Featured project
AURA is the flagship AI product
The portfolio leads with the strongest evidence: an end-to-end mobile AI product with real app screens, backend architecture, agent orchestration, retrieval, personalization, and TestFlight workflow.
AI mobile product
AURA - AI Personal Stylist
TestFlight release / active development
A mobile AI styling platform that turns a real wardrobe into personalized outfit recommendations using wardrobe embeddings, vector retrieval, structured agents, and long-term style feedback.
TestFlight
Release path
10 nodes
Agent workflow
23 golden cases
Eval harness
0 hallucinated items
Closet validity
Built closet-grounded outfit generation with validate-and-repair logic for item membership and role canonicalization.
Turned 18 feedback types into style memory signals for future recommendations.
Added privacy-safe metrics, fallbacks, rate limits, quota controls, and account deletion flows.
React NativeExpoTypeScriptFirebaseFirestoreCloud FunctionsOpenAILangGraphVector Search
Product
AI stylist chat
Digital wardrobe
Outfit recommendation cards
Saved looks
Calendar outfit planning
Wear tracking
Wardrobe insights
Personalized shopping recommendations
Intelligence
Wardrobe embeddings
Vector retrieval
Role-balanced outfit generation
Recommendation validation
Style memory
Feedback personalization
Wardrobe-gap detection
Shopping recommendation scoring
Agent architecture
Intent classification
Style-memory retrieval
Wardrobe-context retrieval
Outfit generation
Refinement
Explanation
Feedback handling
LangGraph orchestration
Production engineering
Firebase Authentication
Firestore
Cloud Functions
Firebase Storage
TestFlight distribution
Rate limiting
Account deletion
Privacy controls
AI quota gating
Fallbacks and error handling
Major engineering projects
Full-stack, AI systems, and performance depth
These projects make the site less generic for recruiters: each one connects a role-relevant domain to concrete architecture, tooling, and results.
Full-stack mobile platform
Forge - Fitness Intelligence Platform
Active development
A production-style workout tracking platform with reliable set logging, progressive-overload recommendations, analytics, and a tested FastAPI backend.
Active development
Status
FastAPI + PostgreSQL
Backend
JWT + Argon2
Auth
Idempotent logging
Reliability
Implemented authenticated sessions, reusable templates, exercise search, workout history, and analytics foundations.
Persisted active workout drafts with set type, units, RPE, notes, and rest-timer timestamps.
Used client mutation IDs and backend reconciliation to make set creation and completion safer.
React NativeExpoTypeScriptFastAPIPostgreSQLSQLAlchemyAlembic
ML systems performance
LLM Training Performance Lab
Reproducible benchmark lab
Optimized GPT-style Transformer training on an RTX 3060, increasing throughput by 3.02x while reducing peak GPU memory by approximately 52.4%.
51.0K tok/s
Baseline throughput
154.2K tok/s
Optimized throughput
3.02x
Throughput lift
52.4%
Peak memory reduction
Benchmarked FP32, TF32, SDPA, FP16, BF16, fused AdamW, and torch.compile configurations.
Generated CSV and JSON artifacts with tokens/sec, step time, max memory, status, and environment disclosure.
Documented Windows/Triton limitations and WSL2 validation instead of fabricating compiled Windows results.
PythonPyTorchCUDAGPT-style TransformersBF16FP16SDPA
GPU inference performance
GPU AI Performance Lab
Reproducible inference lab
A hands-on inference optimization lab benchmarking PyTorch, ONNX Runtime, TensorRT, CUDA, and batch-scaling behavior on an NVIDIA RTX 3060 Laptop GPU.
0.860 ms
ResNet18 p50
3335 samples/s
Batch-16 throughput
+131.6%
TensorRT lift
7.25x
CUDA finance speedup
Validated model output across PyTorch and ONNX Runtime with small expected CUDA numerical differences.
Benchmarked FP32 and FP16 inference paths across dynamic batch sizes and GPU execution backends.
Added a C++/CUDA Monte Carlo option-pricing benchmark to study GPU scaling outside vision inference.
PythonPyTorchONNX RuntimeTensorRTCUDAC++WSL2
Machine perception systems
Comma Video Compression Challenge
Metric-aware compression experiment
Developed and evaluated metric-aware video-compression approaches for machine-perception workloads, reducing the official evaluation score from 2.99 to 1.3546 while keeping the final archive below 1 MB.
2.99
Baseline score
1.3546
Final score
994,341 bytes
Archive size
4,808 bytes
Side channel
Established a reproducible baseline and iterated through validated compression variants.
Reduced model-perception error while staying under the final archive-size limit.
Documented an MPS sharp variant separately because its CPU evaluation path crashed.
PythonVideo CompressionMetric EvaluationPoseNetSegNet
Additional projects
More applied ML work
Focused experiments that support computer vision, ML evaluation, and security-oriented machine learning roles.
Computer vision
EfficientFormer vs ResNet18
Implemented and compared an EfficientFormer-L1-inspired hybrid CNN/Transformer architecture against ResNet18 on CIFAR-10.
87%
Test accuracy
~25% lower
Inference time
Built modular training, validation, and evaluation pipelines.
Used augmentation, tuning, and deployment-focused optimization.
Benchmarked against a ResNet18 baseline for edge-oriented inference.
PyTorchVision TransformersCNNsQuantizationCIFAR-10
Security ML
KDD Network Intrusion Detection
Built an end-to-end ML pipeline for classifying malicious network traffic from KDD Cup 1999 records using ensemble models and explainability.
494K+
Records
99.94%
Accuracy
0.832
Macro F1
Mapped 23 attack subtypes into 5 attack categories.
Handled class imbalance with feature selection and SMOTE.
Used SHAP to interpret high-impact traffic features.
RPythonRandom ForestXGBoostLightGBMSHAPSMOTE
Experience
Product ownership plus engineering execution
The wording stays grounded in actual work: no inflated company claims, no invented titles, and no unsourced metrics.
October 2025 - Present
Full-Stack AI Engineer / Founder
AURA, KASAT Labs
Designed and shipped an AI-powered mobile styling platform from concept through TestFlight distribution.
Built the React Native application, Firebase backend, Cloud Functions, storage workflows, vector retrieval, recommendation engine, and LangGraph-based styling agent.
Developed multimodal wardrobe ingestion, outfit recommendation cards, wardrobe intelligence, personalized shopping, style feedback, saved looks, and calendar wear tracking.
Implemented privacy controls, account deletion, rate limiting, AI usage gating, quota controls, fallbacks, and error handling.
Owned product design, backend architecture, AI orchestration, deployment, testing, and beta preparation as the sole engineer.
January 2023 - March 2023
Software Engineering Intern - Machine Learning
Tech Mahindra
Developed TensorFlow-based computer vision modules in Docker containers and used Jupyter for model experimentation.
Built preprocessing and evaluation workflows that helped standardize model benchmarking on sample images.
Skills
Recruiter-friendly technical map
Organized by the kinds of roles you are targeting: full-stack, AI application development, GenAI, React Native, Node, Python, and ML systems.
Languages
TypeScriptJavaScriptPythonSQL
Frontend and mobile
ReactReact NativeExpoNext.jsTailwind CSS
Backend and data
FastAPINode.jsFirebasePostgreSQLSQLAlchemyAlembic
Applied AI
OpenAI APIsLangGraphEmbeddingsVector searchRetrieval pipelinesRecommendation systemsMultimodal AI workflows
ML systems
PyTorchONNX RuntimeCUDATensorRTMixed precisionModel profilingPerformance benchmarking
Infrastructure and quality
DockerGitHub ActionsPytestCI/CDEASVercelGit
About
Recent MS graduate focused on applied AI products
I recently completed a Master of Applied Science in Computer Science with an AI specialization at Illinois Tech. I am interested in AI products, full-stack engineering, GenAI applications, mobile apps, and ML systems.
AI productsFull-stack engineeringGenAI applicationsMobile appsML systemsReact / React NativeNode / Python
My best work sits where product, backend reliability, and applied AI meet: building the app experience, wiring the data model, grounding AI output in retrieval, and validating the result with tests or benchmark evidence.
Education
Computer science and AI foundation
Illinois Institute of Technology
Master of Applied Science, Computer Science - Specialization in Artificial Intelligence
Chicago, Illinois
August 2024 - May 2026
Dayanand Sagar University
Bachelor of Technology, Computer Science
Bengaluru, India
August 2019 - May 2023
Contact
Open to software, full-stack, and AI application roles
Best fit: teams building practical AI products, mobile or web applications, reliable backends, retrieval systems, recommendation logic, and ML performance workflows.