I fix hard bugs in Celery
Third-year CS student. I chase root causes in concurrency, memory and shared state, then get the fixes merged into the open-source projects production systems run on.
$ git log --merged upstream --author=namraa
✔ celery#10998 redis chord race: callback fired twice
✔ django#22177 FETCH_PEERS overwrote loaded values
✔ agent-framework#8819 failed run leaked state forward
✔ celery#10493 traceback_clear() was a silent no-op
✔ celery#10497 re-entrant PubSub lock deadlock
✔ agent-framework#7944 compaction dropped its summaries
$
Consistency and achievement
Open source activity on one side, a hackathon win on the other.
Upstream engineering
Merged into Celery, Django, Microsoft Agent Framework and Open Food Facts. The interesting part is never the diff. It is finding the root cause.
What broke
Concurrent workers could both complete a Redis chord and both publish its callback, so the callback ran twice.
The fix
Completion detection and dispatch became one atomic, Lua-scripted claim with unique claim tokens and ownership-validated cleanup, so a stale claim can never release a newer one. Added regression tests.
celery#10998 · two workers, one callback
What broke
traceback_clear(exc) silently failed on a live frame, so a traceback–frame reference cycle kept memory alive after hard timeouts.
The fix
Broke the cycle with exc.__traceback__ = None and added regression and smoke tests.
What broke
A redundant UNSUBSCRIBE re-entered redis-py's non-reentrant PubSub lock and deadlocked the result consumer.
The fix
Guarded the redundant call and added regression tests.
What broke
A certificate that expired after it was loaded could still pass signature verification.
The fix
Added an explicit expiry check before verification, with regression tests.
Systems engineering
A price-time-priority limit order book in C++20. Changes were profiled, and the experiments that regressed were reverted.
Cancel-heavy load
Swapping a map/deque book for a sorted vector with intrusive lists made cancels O(1).
Bars use a log scale (1K to 10M ops/s).
cancel-heavy throughput
object-pool hit rate
lower variance (up to)
runs under ASan, UBSan, TSan
Ingestion path
Order feed
ingestion thread
SPSC ring buffer
lock-free, acquire/release
Matching thread
single writer
Order book
sorted vector + intrusive lists
Featured Work
Full-stack applications, ML systems, and cloud architectures built with production-grade patterns.
Engineering intelligence platform that turns GitHub history into verifiable evidence of real work — PR-centric contribution analysis, an auditable repo-eligibility engine, deterministic six-dimension scoring, and evidence-backed resume bullets and STAR answers.
Built a snapshot-safe evidence pipeline where every resume claim must link to the PRs that prove it — claims without evidence are dropped.
Key Engineering Challenges
AI-powered HR analytics platform built for a hackathon. An async Celery and Redis pipeline moves compute-heavy feedback processing off the request cycle, backed by a multi-entity PostgreSQL schema and a Dockerized stack.
Cut synchronous processing load by 60% by moving feedback processing into an async Celery–Redis pipeline.
Key Engineering Challenges
Fully serverless AWS application for secure note and file management. Lambda and API Gateway serve the REST APIs, S3 presigned URLs handle file I/O, DynamoDB stores data, Cognito secures access and CloudWatch tracks custom metrics.
Shipped a secure serverless backend with Cognito JWT auth, least-privilege IAM and an admin-only monitoring dashboard.
Key Engineering Challenges
Full-stack career development platform with gamified learning, MCQ and coding exams, resume generation and a coin-based reward store. Built with Django, DRF, PostgreSQL and WeasyPrint for PDF generation.
Built async resume and PDF generation with RBAC-backed learning workflows.
Key Engineering Challenges
Deep learning project classifying images into 15 animal categories using MobileNetV2 transfer learning. Features data augmentation, early stopping, and a Streamlit web app for predictions.
Trained and deployed transfer-learning inference for 15-class image classification.
Key Engineering Challenges
Experience
Open source and production backend work, in the order I did it.
Celery · Django · Microsoft Agent Framework · Open Food Facts · Remote
Techaronic IT Solutions Pvt. Ltd. · Ahmedabad, India
Recognition
Competitive programming, hackathons and certifications. The Code for Good win is up top.
Winner. Delivered production-ready features under time constraints.
Top 2,500 of 30,000+ participants.
Qualified for Round 2.
Finalist, on technical merit.
Generative AI Professional and Data Science Professional.
Technical Expertise
A living visualization of connected technologies floating in space.
Academic Background
Computer science and engineering degrees.
B.Tech Computer Science and Business Systems (CSBS)
Diploma Computer Engineering
Get in Touch
Open to backend engineering, distributed systems, and AI infrastructure roles.