Sayed RaheelHussain

Founding engineer leading the production AI surface. Took Kayaan AI from 0 → $1M ARR in 8 months. Now leading the build of Kayaan Freight, an autonomous freight brokerage on top of the same agent platform.

• ›Designed and shipped a 0→1 production voice-driven logistics system (ASR → LLM → TTS), deployed end-to-end in real-world operations.
• ›Built a latency-critical real-time voice pipeline, optimising TTFB / TTFT, throughput (TPS) and cost for conversational workloads.
• ›Defined offline and online evaluation pipelines for LLM systems — tracking WER, latency distributions, task success, and post-deployment regressions.
• ›Implemented golden-prompt regression testing to catch quality regressions before model releases.
• ›Shipped production safety: PII redaction, schema-validated tool calls, rule-based fallbacks, and human-in-the-loop escalation.
• ›Achieved 30–40% workflow automation; scaled operator capacity from ~5 to ~50 concurrent workflows (target 80%).
• ›Shipped per-user and per-agent long-term memory — procedural + semantic, vector-backed (LLM-OS style) — enabling persistent user personality, context-aware reasoning, and iterative agent improvement via feedback-driven re-prompting. Built in ASP.NET (C#) with a React frontend.
• ›Lead a team of 7 (2 FDEs, 1 AI Architect, 4 platform engineers) across the Kayaan platform and the Kayaan Freight build.

Embedded with clinical teams to ship a multi-agent assistant for clinical information retrieval under strict safety constraints.

• ›Built and deployed a multi-agent AI assistant for clinical information retrieval used by care teams.
• ›Evaluated and adapted domain-specific language models under strict safety, accuracy, and usability constraints.

Fine-tuned diffusion models for interior design generation and productionised the segmentation + generation stack.

• ›Fine-tuned and deployed ControlNet-based diffusion models for personalised interior design image generation.
• ›Built and productionised image segmentation + generation pipelines — improving end-to-end throughput by 30% and enabling scalable inference.

Pre-trained a domain-specific 1B-parameter transformer for veterinary medicine and shipped a few-shot CV system for livestock health.

• ›Built VetMedGPT: curated a 500GB+ corpus and trained a 1B-parameter Transformer from scratch using distributed training on 6× NVIDIA A100 GPUs.
• ›Developed a few-shot computer vision system for livestock health assessment — custom CNNs and fine-tuned YOLOv8 segmentation/detection — achieving 54%+ AP after labeling 47k frames.