AI-Engineer (Full-time) at UrbanMatch (2024-08 – Present)
Technical Blog
- Architected Advanced agentic hybrid search engine over a joint embedding space with Qwen3-Embedding-4B (dense) and Learned SPLADE (sparse) vectors fused search using Distance-Based Score Fusion (DBSF) to preserve similarity accuracy for higher top-K precision vs RRF. Implemented iterative depth/breadth semantic exploration with learning-driven query refinement and progressive filtering, extended to multimodal retrieval (CLIP) and optimized memory/latency via int 8 scalar quantization for lower latency.
- Designed a fault-tolerant real-time streaming retrieval pipeline using Redis Stream Groups + SSE with consumer ownership validation and background processing, enabling result streaming beyond 60-second request limits; optimized retrieval performance via scalar/binary quantization and extended search to multimodal retrieval (text→image, image→image) using CLIP embeddings.
- Architected a hierarchical hybrid memory layer (session + Redis + Mem0 graph + text search) enabling long-term personalization with short-term conversational continuity, while actively managing inference stability through GC pressure tuning, Redis TTL controls.
- Eliminated GPU OOM from large embedding models by orchestrating Qwen (dense) and SPLADE (sparse) inference in separate containers across heterogeneous GPUs (L40S + T4), preventing VRAM contention and removing dependency on A100-class hardware through workload-aware GPU partitioning.
- Optimized serverless cold-start latency by pre-baking model weights into container images, caching to blob storage (60s → 25–30s), implementing CUDA checkpointing to persist compiled CUDA graphs and kernels (12–14s startup), and replacing vLLM boot paths with TEI (Text Embeddings Inference) to bypass CUDA graph creation (eliminating 4–5 minute warmups)—enabling high-throughput, low-latency, cost-efficient inference via coordinated batching and GPU-saturation tuning.
- Vertically scaled Clustered Monolith architecture to 20K+ users and drove UrbanMatch acceptance into Modal Labs startup program (secured $5K GPU credits) and mem0-ai 6-month accelerator
AI-Engineer (Consultant) at Burni Consulting (2024-07 – 2024-07)
- Worked on ideation and designing system architecture, while researching various speaker diarization techniques to separate audio based on individual speakers in the recording.
- Complied with the EU's Data Protection Act (GDPR) by implementing Whisper.rn on-device, eliminating the need for cloud data storage.
- Researched Medical-LLM for summarization and drug suggestion, and learned how to quantize models and deploy them to edge devices using Qualcomm AI Hub.
AI-Engineer (Consultant) at InTouchAPP (2024-05 – 2024-05)
- Designed and developed an end-to-end architecture to integrate an advanced retrieval-augmented generation (RAG) feature as part of an experiment to connect private data with a large language model (LLM)
- I worked on a subset of sales data to implement a natural language Q/A bot for managers and the CEO and made key decisions on the choice of database (Qdrant DB), LLM (GPT-4), and embedding model (OpenAI's text-embedding-ada model) with respect to Self-hosting, Cost, Scalability, reliability, and model evaluations
- Key features I implemented include query translation, reranking, advanced message-based chunking, and improvements to accuracy, achieving approximately 90 percent
