Sahil Chachra

Joined a pre-seed startup building vision analytics for physical retail spaces. An AI stack existed — but it was expensive to run and didn't scale economically. My job: rebuild the pipeline to run on cheaper infrastructure while supporting far more cameras, inference throughput, and VLM API calls simultaneously.

• ▸Built a real-time multi-camera VLM pipeline from scratch — asyncio-native, with a sliding-window persistence gate that filters hallucinations before they reach the backend, cutting false alert rates without adding latency.
• ▸Contributing to a patented hardware-accelerated video compression algorithm (Intel VAAPI) and building cross-platform utility applications (Windows, macOS, Linux) that onboard customers into the BLUE ecosystem.
• ▸Profiled the full inference stack across quantization levels and GPU/edge targets — found and eliminated the bottlenecks preventing real-time SLA compliance.

First engineer hired at a pre-seed startup building AI for factory floors. The domain was manufacturing — CAD files, machine catalogs, compliance specs, and factory-floor constraints. The job was to make an LLM reason reliably across all of it, at production scale, with no prior art to follow.

• ▸Designed a 3-tier, 27-agent orchestration framework — role-based CrewAI agents, async direct-LLM agents, and an 11-stage DAG task executor — decomposing complex manufacturing tasks into verifiable, structured sub-problems.
• ▸Built a 7-stage document enrichment pipeline (extract → plan → classify → enrich → normalize → cross-doc → assemble) with conditional enrichers and concurrency semaphores handling 5 document modalities in parallel.
• ▸Integrated bidirectional MES sync with Odoo (JSON-RPC 2.0), connecting AI-generated factory plans to live shop floor data for planned-vs-actual dashboards — the first time those two worlds talked to each other.

Not everything I do is full-time. At Mesa I mentor student teams during intensive AI hackathons — helping founders and postgrads go from idea to working prototype using modern AI tooling, without needing deep engineering backgrounds.

• ▸Guided teams across PG and UG cohorts through hands-on builds using Google AI Studio, Relevance AI, and n8n — covering prompt design, agent workflows, and no-code/low-code AI automation.
• ▸Helped non-technical teams translate business problems into AI-powered prototypes within hackathon time constraints.

Three years scaling a computer vision platform from early contracts to enterprise deployments across hundreds of cameras. I went from training models to owning system architecture — and eventually added LLMs to a stack that was purely CV when I joined.

• ▸Scaled a computer vision platform to 600+ cameras across enterprise clients, maintaining production reliability across 30+ active use cases simultaneously.
• ▸Built a person re-ID pipeline that tracked individual customer journeys across camera zones — measuring dwell time and staff engagement patterns that didn't exist in any other data source.
• ▸Engineered an LLM + RAG layer that translated natural language requirements directly into structured CV pipeline configurations, cutting new use-case deployment cycles by 20%.
• ▸Upgraded the platform with LLaVA and Qwen VLMs, achieving up to 95% accuracy improvement across 350 cameras.

Started my career applying deep learning to real automotive and mobility problems. The fundamentals I built here — training discipline, production mindset, and understanding hardware constraints — shaped everything that followed.

• ▸Accelerated data annotation pipelines using deep learning models for the Smart Mobility Group, cutting manual labeling time on automotive datasets.
• ▸Integrated trained CV models into customer-facing products across mobility and automotive domains — one of the first times I shipped ML to real users.
• ▸Awarded the Technical Excellence Award for outstanding contributions to the organization.