Hammad Sarfraz

DocuRAG | LangGraph, RAG, BM25, Vector Search, RRF A 7-agent RAG pipeline with LangGraph, conditional routing, bounded retrieval-retry loops, and ambiguity-triggered clarification. Implemented hybrid BM25 + vector retrieval, weighted score fusion, cross-encoder reranking, semantic chunking, and dynamic context-token budgeting.

Software Engineer (AI and Backend) at Taleemabad - Powered by Orenda (2025-05 – Present)

• DocuRAG | LangGraph, RAG, BM25, Vector Search, RRF
• A 7-agent RAG pipeline with LangGraph, conditional routing, bounded retrieval-retry loops, and ambiguity-triggered clarification.
• Implemented hybrid BM25 + vector retrieval, weighted score fusion, cross-encoder reranking, semantic chunking, and dynamic context-token budgeting.
• Built grounded, citation-backed answers with LLM-based groundedness evaluation and confidence-gated web-search fallback.
• Implemented semantic answer caching with embedding similarity and corpus-mutation invalidation, reducing redundant LLM calls and query cost.
• Hardened the pipeline with timeouts, exponential backoff, circuit breakers, rate-limit-aware retries, path-traversal protection, and upload-size guards.
• Added per-run observability for stage latency, token usage, and agent paths via a pluggable LLM provider layer.
• Added adversarial regression tests covering retrieval, caching, concurrency, and file-handling failures.
• Meetings MCP | Python, MCP, PostgreSQL, RLS
• Built an MCP server exposing 500+ meetings to AI agents through progressive-disclosure search, resolution, summaries, transcript retrieval, and Notion card generation.
• Designed cost-efficient retrieval workflows using staged search → resolution → summary → transcript access to minimize unnecessary token usage.
• Implemented PostgreSQL RLS-based access control with public-corpus and participant-scoped modes, preventing unauthorized meeting access at the database layer.
• Added startup security assertions to detect misconfigured database roles before serving restricted data.
• Added access-aware responses distinguishing unavailable meetings from restricted meetings to prevent misleading agent responses.
• Migrated years of teacher-training data to a new participant identity model using reversible set-based SQL across 7 profile types, 6 live/historical tables, and 2 legacy ownership schemas.
• Implemented backward-compatible Django migrations for business-driven data transitions with zero downtime.
• Owned the Grand Quiz module end-to-end across Django and Retool, including subject-driven modes, MCQ locking.
• Restored a broken staging-to-production data transfer pipeline, ensuring reliable data integrity across environments.
• Built signal-driven Celery pipelines for asynchronous processing across internal workflows and external services.
• Optimized complex REST APIs, reducing response payloads from 70MB to 3-4 MB.
• Diagnosed and resolved Redis memory spikes, improving cache stability and resource efficiency.
• Reduced server memory usage through AWS S3 presigned URLs for file delivery.
• Translated requirements from non-technical stakeholders into production-ready deliverables.

Backend Engineer at Wave (Contractual) (2026-02 – 2026-04)

• SpeedSpeak Backend | Django, DRF, PostgreSQL, JWT, AWS S3, Firebase
• Built a Django REST API for an interpreting/translation marketplace connecting clients, interpreters, and admins across booking, bidding, assignment, and payment workflows.
• Designed the booking domain with DB-level CheckConstraints, interpreter bidding, primary/backup assignments, timesheets, and role-based communication.
• Implemented collision-resistant monthly booking serials with transactional retries and concurrency-safe uniqueness handling.
• Implemented OTP-based email verification/password reset with expiry, resend limits, cooldowns, and automated cleanup.
• Built resilient Firebase push notifications with graceful failure handling and automatic removal of dead FCM tokens.
• Implemented admin-managed client/interpreter payment tracking independently from original booking estimates.

Python Developer at Botnostic Solutions (2024-05 – 2024-09)

• Built ETL scripts for data preprocessing and pipeline optimization.
• Conducted exploratory data analysis and API-compliant data parsing.
• Created data visualizations for analytics dashboards and stakeholder reporting.
• Automated regression and functional testing using Selenium with JSON output validation.
• Analyzed Prime Minister Youth Program datasets to extract insights supporting program evaluation.
• Performed automated testing on chatbots and web applications using Selenium.

Bachelor of Science in Data Science – FAST National University of Computer and Emerging Sciences (NUCES) (2021 – 2025)