Hanzala Ahsan

AI Lead Developer

Özyeğin University • Full-time

Jun 2026 - Aug 2026 • 2 mos

End-to-End Pipeline Architecture: Engineered a robust, asynchronous Python and FastAPI backend from scratch to handle document ingestion, chunking, embedding generation, and LLM inference flows. Integrated high-performance BGE embeddings to capture dense semantic representations across massive corpora of complex financial data. Advanced Vector Search & Indexing Optimization: Configured and optimized PostgreSQL with pgvector, implementing HNSW (Hierarchical Navigable Small World) cosine indexing to achieve ultra-low-latency similarity search over 158,570+ dense text chunks. Eliminated retrieval noise by implementing multi-query decomposition (breaking complex user audit questions into sub-queries) and integrating a cross-encoder reranker to filter and score top matches before context injection. LLM Integration & Cost/Latency Governance: Connected the retrieval pipeline to Amazon Bedrock DeepSeek generation APIs, establishing strict context window governance and prompt guardrails to ensure deterministic, factual answers. Minimized token overhead and operational costs by ensuring only high-relevance, reranked chunks reached the model generation stage. Rigorous Evaluation Harness & Quality Assurance: Built a frozen-pipeline evaluation harness using golden test datasets and unseen enterprise audit questions to mathematically measure retrieval efficacy rather than relying on subjective output. Measurable Impact: Achieved a top-tier evaluated Hit@5 of 1.00 and Recall@5 of 0.958, completely eliminating hallucinations and successfully resolving multi-hop financial reasoning queries.