Software Engineer — full-stack · applied AI · Doha, Qatar
I build web products end to end and make the AI parts inside them reliable enough to ship.
- Whole products, not prototypes — React and Next.js frontends, Node and FastAPI backends, Postgres underneath, and the deployment nobody posts about
- AI that gets checked — generated SQL reviewed before it runs against a read-only database, retrieval with real access control, safety logic that doesn't depend on the model behaving
- Finding where the time goes — profiling slow endpoints, indexing, caching, and fixing the thing that's actually slow instead of the thing that looks slow
- Building with the people who'll use it — sitting with users first, shipping, then changing it based on what they say
BS in Artificial Intelligence, FAST NUCES.
Most of what I build day to day is closed source, so it isn't here. The current one is a sports analytics platform I'm the only engineer on: a 40+ page React app with custom visualisations, Node and FastAPI services, a Postgres database built from a pipeline covering millions of rows, subscription-based access control, and a chat that turns plain questions into SQL and checks the query before letting it near the database. Before that, email and calendar integration for an AI assistant, where nothing ran without the user approving it first.
grayhatdevelopers/vidxp — a video indexing engine that lets agents search inside video in natural language. I contributed bulk indexing for large media sets and path-resolution test fixes.
| What it does | Built with | |
|---|---|---|
| MindEase | Bilingual Urdu/English mental health companion with text and voice chat. Screening results feed into the conversation, and crisis handling runs on rules rather than trusting the model. Led development on a 3-person team. | Next.js · Django · pgvector |
| Secure Multi-Tenant RAG | Separate groups query private knowledge bases. Isolation is enforced at retrieval, not in the prompt, with ACLs, PII masking and injection detection. A red team suite of 10 attack prompts treats any leak as a failed test. | Flask · ChromaDB · Groq |
| IoT Vulnerability XAI | Predicts CVSS severity and CWE type from vulnerability text, with SHAP token attributions showing which words drove each prediction. 0.729 F1 across 28,902 records. | SecureBERT · PyTorch · SHAP |
| Story2Audio | A prompt becomes a narrated story with cloned voices. gRPC service defined in protobuf, with a REST proxy in front for clients that don't speak gRPC. | gRPC · Protobuf · XTTS |