SENIOR SOFTWARE ENGINEER · SENIOR ML ENGINEER
Backend, cloud, and ML across computer vision, LLMs, and data science, with a research streak in trustworthy ML. Based in Karachi, working remotely with US teams.
1ST PLACE Task 2 — MediaEval Medico 2026 · calibrated, safety-aware medical VQA
01 — ABOUT
I am a senior software and ML engineer. I build production systems end to end: the model, the backend services that serve it, and the cloud and DevOps that keep them running. I move as comfortably through systems, backend, and infrastructure work as through the machine learning on top, whether that's computer vision, LLMs, or data science. And I ship the whole thing, not just a notebook.
I also do research on trustworthy, calibrated ML, which took 1st place at MediaEval Medico 2026. That rigor carries into everything I build, but it is one part of the picture: most of my work is getting real systems designed, deployed, and dependable in production. I like problems that are well specified and hard.
02 — EXPERIENCE
Couro · US company · remote FLAGSHIP
My flagship role. I build production systems across ML, backend, and cloud DevOps for a sports AI platform: FastAPI services for video upload, processing orchestration, model inference, and result delivery, on AWS with blue/green deployments and scalable workflows for pose estimation and athlete performance analysis.
Folio3
Owned computer-vision and MLOps work end to end across a range of projects: trained and shipped detection, segmentation, tracking and classification models, and built the streaming pipelines and AWS infrastructure to run them reliably at scale.
Folio3
Worked across several real-time computer vision systems, including license plate recognition, sports ball tracking, and foul detection. Built detection and tracking pipelines and tuned them for real-time inference.
03 — SELECTED WORK
OPEN SOURCE — PUBLIC REPOSITORIES
Tesseract Hackathon 2026: composed Julia (JutulDarcy adjoint), Python/JAX autodiff, and C++ finite-difference solvers into one end-to-end differentiable pipeline for CO₂-storage caprock inversion from InSAR uplift.
3 SOLVERS · 1 DIFFERENTIABLE FN
Julia · JAX · C++
Production agentic RAG: OpenAI tool-calling with an extensible tool registry, retrieval, observability, and human escalation. Deployed on AWS with Terraform IaC.
TOOL-CALLING + RAG + IaC
FastAPI · OpenAI · Docker
A white-box estimator of neuron activation with a spec-parity test harness: two independent implementations cross-checked against a written spec.
TWO IMPLS vs SPEC
Python
Retrieval-augmented medical QA built around a formal eval harness; hybrid search over Qdrant with chunking strategies benchmarked. Dockerized, Streamlit.
RETRIEVAL QUALITY MEASURED
LangChain · Qdrant · Docker
A reusable, production-grade full-stack foundation for building multi-tenant AI SaaS apps fast.
MULTI-TENANT FOUNDATION
TypeScript · full-stack
Award-winning gastrointestinal VQA: calibrated aspect-reliable explanations with safety-aware abstention. 1st place, MediaEval Medico 2026 (Task 2).
1ST PLACE · MEDIAEVAL 2026
PyTorch · Computer Vision
paper · MediaEval Medico 2026 (Task 2) · paper9.pdfINDEPENDENT BUILDS — PRIVATE
Products I'm building. Ledger and Sentinel are sibling tools, designed to work together.
Config-driven document extraction with certified selective acceptance. It auto-accepts only the fields that clear a proven error bound and routes the rest to review, so teams get automation they can actually trust. I'm building it into a product.
CERTIFIED ACCEPTANCE
Python · LLM · vision
Spec-driven RPA: an LLM solves a task once, then compiles it to a deterministic plan that runs with zero model calls. A sibling to Ledger, built to integrate with it.
ZERO MODEL CALLS ON REPLAY
Python · LLM
Coming soon
Demand and sales forecasting with calibrated conformal prediction intervals whose coverage is validated on held-out windows.
VALIDATED INTERVAL COVERAGE
Python · time-series
Tabular AutoML with calibrated uncertainty via conformal prediction. Trustworthy probabilities, not just point predictions.
CALIBRATED UNCERTAINTY
Python · scikit-learn
FDR-controlled anomaly, fraud, and tamper detection tuned to alert economics. Precision and recall get traded on cost, not a threshold guess.
FDR-CONTROLLED ALERTS
Python
A calibrated churn and customer-analytics suite: trustworthy probabilities with SHAP reason codes.
CALIBRATED CHURN + REASONS
Python
Agentic tender and contract analysis that returns every obligation with a page-level citation and a reasoned bid/no-bid.
CITATION-OR-NOTHING
Python · LLM
A code-first, AI-native workflow automation and integration engine.
CODE-FIRST INTEGRATIONS
Python · FastAPI
A config-driven AI voice-agent platform for inbound and outbound phone agents.
INBOUND + OUTBOUND AGENTS
Python · telephony
Shift-aware bounded-risk explanations for medical VQA. It's the research lineage behind the award-winning Medico work.
BOUNDED-RISK UNDER SHIFT
PyTorch · Computer Vision
A config-driven NLP engine: classification, aspect-based sentiment, topics, NER, and summarization.
ONE ENGINE, MANY TASKS
Python · NLP
A config-driven recommendation engine whose deliverable is a leakage-safe evaluation.
LEAKAGE-SAFE EVAL
Python
A retrieval platform: a Python scraper aggregating web text (TXT/DOCX/PDF) into a LangChain vector store, with RAG over OpenAI and Hugging Face models, cross-lingual querying, and a Streamlit chat UI.
CROSS-LINGUAL RAG
LangChain · OpenAI · Hugging Face · Streamlit
COMPANY WORK — EMPLOYER PROJECTS
Founding Engineer: FastAPI services for video upload, orchestration, inference, and delivery, plus CV pose estimation and athlete performance, on AWS (ECS/Fargate, SageMaker, Cognito, DynamoDB) with Terraform and blue/green.
POSE + PERFORMANCE, IN PROD
FastAPI · AWS · Computer Vision · Terraform
Detect and track logos across ~1,900 brands. Warm Lambda with provisioned concurrency for low-latency synchronous image ops; async video via an ECS cluster on g4dn EC2, orchestrated with API Gateway and SQS; an incremental YOLOv8 pipeline that learns new logos without full retraining; all infrastructure in Terraform.
1,900 BRANDS · INCREMENTAL YOLOV8
YOLOv8 · AWS Lambda · ECS · Terraform
Four 1080p 60 FPS cameras detect the ball's exact contact point against court zones (service box, no-go, out); a math algorithm estimates position when high-speed motion outruns the cameras. Real-time.
4 CAMERAS · 60 FPS · REAL-TIME
Python · OpenCV · Computer Vision
One system running 12-13 ML models (PPE, smoking, fall, license-plate, phone-use, fighting) over many simultaneous live camera streams. Modular via the Factory pattern and SOLID; Kafka for real-time inference flow, DynamoDB for results, GStreamer ingest, Docker Compose deploy.
12–13 MODELS · MANY LIVE STREAMS
Kafka · DynamoDB · Docker · GStreamer
Real-time ALPR with OCR, location mapping, blacklist/whitelist, multi-channel alerts, transport-authority verification, and video management system integration.
REAL-TIME ALPR + OCR
Python · OpenCV · OCR · object tracking
Architected and deployed the AWS processing pipeline for a lacrosse performance-analysis app (pose via MediaPipe): API Gateway, Lambda, ECS Fargate and EC2, ECR, CloudWatch alarms, and SNS alerts. I owned the cloud infrastructure and deployment.
AWS PIPELINE · MONITORED
AWS Fargate · Lambda · CloudWatch · Terraform
An AWS Lambda image-processing pipeline: pull from S3, detect page contours and crop margins, watermark, and emit a JPEG plus light, medium, and small thumbnails, with Google Document AI integration and a structured callback. Built multi-tenant with automated per-tenant deployment and cost-configurable steps.
MULTI-TENANT · CONFIGURABLE STEPS
AWS Lambda · S3 · OpenCV · Pillow
A full-stack internal operations platform for a large engineering firm: a document and drawing register, project records, transmittals, and role-based access control, built as a reliable system of record.
RBAC · SYSTEM OF RECORD
Next.js · TypeScript · PostgreSQL · Docker
04 — PUBLICATIONS · 2026
MEDIAEVAL MEDICO 2026
A frozen-backbone system with a training-free reliability layer that knows when to trust its own answers, delivering calibrated, safety-aware responses to medical visual questions. It won its task.
→ read paper (PDF)MEDIAEVAL SYNTHETIC IMAGE 2026
Detecting synthetic images by stacking frozen features from multiple backbones into an out-of-fold ensemble, an approach that posted the highest F1 of any open run on the task.
→ read paper (PDF)05 — SKILLS
DOMAINS
Computer vision spans detection · tracking · pose · segmentation.
ML & AI
CLOUD (AWS)
INFRASTRUCTURE & DATA
LANGUAGES
PRACTICES
CREDENTIALS — VERIFIABLE
ACHIEVEMENTS
06 — CONTACT
I'm open to senior ML and applied research work, and to hard, well-specified problems in general. The fastest way to reach me is email or LinkedIn.
Available for senior software and ML engineering work