GPU inference · CPU inference · vLLM & Kubernetes · Agentic AI · Model routing
01 / Experience
Machine learning and AI infrastructure.
I started in computer vision and document AI. My recent work is in LLM serving, agent platforms, and the systems around them.
Jun 2026 to presentPanasonic Avionics
Machine Learning Engineer II · promoted from MLE I
Designed and optimized custom MiniMax-M3 inference on NVIDIA RTX PRO 6000 GPUs, achieving 160 tokens/s decode throughput at batch size 1.
Implemented and optimized an Ideogram forward pass for vLLM-Omni, reducing image generation time from roughly 2 minutes in the reference PR implementation to 24 seconds.
Built an end-to-end LLM router for Claude Code, Kilo Code, and Cursor, improving cost utilization by 4Ă—.
Started and expanded a unified inference platform for open and closed models, raising productivity by 60% across compatible LLM SDKs.
Built an agent platform that cut the time to create a production-grade agent from days to under 10 minutes.
Designed a Kubernetes controller for lifecycle management, budgets, observability, and operations. It reduced debugging and incident-resolution time by 90%.
KubernetesLLM routingAgentic AIObservability
Jan to Jul 2025Calsoft, Inc.
Data Scientist · consultant for Panasonic Avionics
Built input and output guardrails with open and closed-source tools, contributing to an 88% safer AI system.
Built a WAF-protected vLLM pipeline that served dense models at 80 tok/s for proprietary coding and application workflows.
vLLMAI guardrailsRayWAF
Sep 2023 to Jan 2025Softsensor.AI
Associate Data Scientist
Built cybersecurity workflows for Text2SQL and unstructured MITRE document search.
Built document translation software that preserved layouts, tables, and diagrams across different file types.
Led claims-document extraction with linked document relationships, cutting extraction cost by up to 85%.
RAGText2SQLOCRDocument AI
Mar 2022 to Feb 2023Blackcoffer Consulting
Associate Data Scientist · remote
Worked on satellite-image vehicle detection, document summarization, data visualization, market intelligence, and e-commerce reporting.
Computer visionNLPFlaskAnalytics
02 / Work
Selected work.
Open-source contributions and projects I have built outside my day job.
Vulkan · Merged upstream contribution
RX 560X 316.60 AGG. TOKENS/S
Pushing legacy hardware
Built and validated a Qwen3.5-2B inference path on an RX 560X, reaching 316.60 aggregate decode tokens/s across 128 short-context sequences. Portable runtime fixes and a 2.55Ă— CPU embedding-decoder speedup merged upstream.
A reusable coding-agent skill for self-hosted LLM inference, from one GPU to Kubernetes: memory sizing, runtime selection, API verification, workload benchmarks, and measured optimization.
Turns repository history into coding tasks, verifies each task against the historical fix in Docker, and reports how candidate models perform on your code.
Built and tested a custom vLLM branch with SM 12.0a backend support and parser fixes after a community report of Chinese and nonsensical output. The upstream PR remains open and unmerged; review raised concerns about broad parser-side normalization.
An agent skill for Claude Code, OpenCode, and Codex that plans LoRA, SFT, or continued pretraining, catches local training failures, and evaluates the artifact that will actually be served.
What's unique about Abhiram is that he is both an incredibly gifted and competent coder, and is still able to work and communicate with product managers who do not necessarily understand the technical ins and outs of the product he helps create. He has creative solutions in getting around problems; he is very well-read on the literature surrounding machine learning, leading to a great awareness of different models.
On top of his abilities, what stands out is his work ethic. Abhiram responds to requests promptly and is always willing to help you with running and improving the programme. He is helping us build a product we are proud of, and we are sure he will continue to build great things and come up with great ideas with whomever he works with in the future.
I worked with Abhiram for 8 months. During this time he gave us frequent updates about our project, fantastic explanations and dedicated much of his time to help us develop our ideas and turn our ideas into practical ideas. He acted with the up most professionalism and conducted himself through his knowledge and understanding of computer science, specifically python. He was accessible when we needed to ask questions and very happy to hop onto a call.
Overall, I enjoyed working with Abhiram and the knowledge and creative solutions he brought to the table were outstanding.
I had the privilege of working alongside Abhiram in the AI team at SoftSensor.AI, and I can confidently say that his technical prowess and innovative thinking set him apart. What impressed me most was his ability to tackle complex challenges in our document chatbot project, where he seamlessly integrated multiple data sources while maintaining high accuracy in information retrieval.
Beyond his technical skills, Abhiram is an exceptional team player who consistently brings fresh perspectives to discussions. His ability to collaborate effectively with both data scientists and engineers while maintaining a clear focus on deliverables made him an asset to our team.
