Latency
<15 ms
End-to-end real-time edge CV inference for global live broadcast feeds (WTT).
Available for Fractional AI Advisory & Architecture Sprints (Q4 2026)
AI Systems Architect & Fractional Founding Engineer
Taking mission-critical Machine Learning and Generative AI systems from prototype to hardened production. Specializing in real-time edge/cloud inference optimization, deterministic multi-agent systems, and multimodal enterprise RAG.
Upstream Contributor to PyTorch & YOLO-WorldEx-Founding EngineerSystems Architect
Production benchmarks
Hallucination cut
30%
Achieved across enterprise RAG and agentic workflows via hybrid BM25 + dense vector routing.
OCR accuracy
+25%
Layout-aware document extraction pipelines using LayoutLMv3 and deep learning inpainting.
Contributor
Upstream
Verified code and optimizations merged into PyTorch and YOLO-World core repositories.
Systems architecture
Latency, failure domain, and cost sit inside the architecture.
01
Real-time edge and cloud inference
The forward pass is not the whole budget. Capture, the compiled graph, and the stream are. PyTorch weights leave through ONNX into TensorRT and run on Jetson-class hardware beside the feed.
PyTorch · ONNX · TensorRT · Jetson · FastAPI
02
Deterministic multi-agent systems
Retrieval answers before a model speaks. Hybrid BM25 and dense search run first. Smolagents and CrewAI sit behind a step limit, and domain models are served on Triton with a cost trace on every query.
LiteLLM · Qdrant · ChromaDB · Triton · Grafana
03
Multimodal document intelligence
LayoutLMv3 reads the page rather than a flattened string. Inpainting removes PII from the image. ECS and SageMaker carry the model out of the notebook, and people only see the low-confidence remainder.
LayoutLMv3 · Inpainting · ECS · SageMaker
Case studies
Production systems with a measured constraint.
Broadcast vision, financial agents, and logistics documents. Each engagement had a ceiling — latency, hallucination, or review load — and a number on the other side of it.
01World Table Tennis
<15 ms
Broadcast-grade edge vision under 15 ms
Context
World Table Tennis international tournaments, including WTT Star Contender Goa and Singapore Smash. A single camera feed, millions of live global viewers, and a broadcast encode that cannot wait on a distant GPU.
Problem
Single non-stereo camera pose tracking with a hard real-time ceiling under 20 ms and zero perceptual broadcast delay. A stereo rig was not available, and a cloud round trip would have shown up on air.
Camera
One non-stereo broadcast camera. No rig calibration and no sync hardware on the critical path.
- End-to-end latency
- <15 ms
- Live viewers
- Global
- Stereo-rig overhead
- Zero
02Fintech and B2B SaaS
30% cut
Multi-agent orchestration and hybrid RAG
Context
Fintech and B2B SaaS platforms that automate document handling and decisions over unstructured financial tables. The cost of a wrong cell is a wrong decision, and the cost of an open agent loop is a surprise invoice.
Problem
LLM hallucinations, token-cost explosions, and unbounded agent loops over financial tables that do not arrive as clean JSON.
Intake
Unstructured financial tables and documents. A naive chunker invents structure the source never had.
- Lower hallucination
- 30%
- Financial logic
- Multi-step
- Cost tracking
- Per query
03Supply chain and logistics
+25%
Document intelligence and automated redaction
Context
Supply chain and enterprise logistics automation. Invoices and bills of lading arrive in whatever layout the sender used, and PII cannot ride along into downstream storage.
Problem
Parsing noisy, multi-layout invoices and bills of lading while scrubbing PII at scale, without parking every document in a human review queue.
Capture
Multi-layout invoices and bills of lading, including scans that defeat template OCR.
- Field accuracy
- +25%
- Less human review
- 70%
- Scrubbed at scale
- PII
Advisory models
Two engagements. Both scoped before the work starts.
A fixed sprint when the question is whether the system will survive contact with production. A monthly retainer when the team needs a senior engineer in the repository.
2-week fixed scope
Architecture & Evaluation Sprint
$6,000 – $8,000
Fixed fee · 2 weeks
Best for
Seed and Series A teams preparing to launch, or already scaling inference and finding the budget in production.
Scope
- Deep teardown of the model and retrieval pipelines
- Triton and vLLM inference profiling
- Token and GPU cost optimization
- Deterministic guardrail roadmap
- Concrete benchmark pull requests
Monthly retainer
Fractional Founding AI Engineer
$10,000 – $14,000 / month
Retainer · 15–20 hours/week
Best for
Funded teams that need a senior hands-on technical lead without a six-month hiring cycle or equity dilution.
Scope
- 15–20 hours/week of dedicated engineering
- Direct model fine-tuning on your data
- Ownership of the RAG and agent platform
- MLOps orchestration across the serving path
- Mentoring the engineers who will run it