Applied AI Engineer
Turning AI pilots intoproduction systems.
I design, train and deploy agentic systems, LLMs and forecasting models for enterprise clients, from the first model to the system in production.
AI Engineer at Velocis Systems
01Work
Selected work
Enterprise AI systems I designed and built, from the model to the product around it.
- 01
Velocis Foresight
AI spare-parts forecasting for a leading automotive OEM
17.2% → 14.9%forecast error (WAPE) across 21,957 parts
Problem
Spare-parts planners must decide how much of each part to order, across tens of thousands of parts with irregular demand. Too much stock ties up capital; too little causes stock-outs at dealers.
Approach
- Adapted Cisco's time-series foundation model with a LoRA adapter that trains 3.6% of the parameters in about seven minutes on one H200 GPU.
- Benchmarked it against 15 statistical and deep-learning models on 52 months of history.
- Turned forecasts into order quantities: a business-risk class decides which demand quantile each part is planned to.
- Shipped a planner console and a planning assistant in the console and in Webex. A grounding check blocks any figure that no tool returned.
- Secured it with Cisco Duo and Cisco AI Defense, and connected Splunk alerts for forecast drift.
Results
- Forecast error (pooled WAPE) fell from 17.2% to 14.9%.
- Forecast value added over the naive method rose from 18.5% to 29.5%.
- Parts forecast within 20% error now carry 83.6% of demand volume, up from 78.9%.
- 02
Bullseye Zero
Agentic root-cause analysis for enterprise networks
~24vendor-neutral tools behind one agent
Problem
A network problem usually arrives as a vague complaint, such as “the Wi-Fi is slow”. Engineers then check many tools by hand, and an AI answer without evidence is not something an operations team can act on.
Approach
- A multi-turn agent turns the complaint into a report. The language model chooses the next check; deterministic code computes device state, the ranking of causes and the confidence.
- A dependency graph of the network removes impossible causes before any ranking.
- Read-only connectors for Cisco Catalyst Center and Cisco ISE sit behind vendor-neutral tools, so a new data source adds no new tools for the model.
- Missing or failed evidence is reported as unknown, never as healthy. Every result is stored in an append-only evidence store.
- The agent can stop to ask a person a question and continue when the answer arrives, even after a restart.
Results
- Tested live against Cisco Catalyst Center and Cisco ISE: a case can start from a person's name alone and find their device.
- Covered by 98 automated tests.
More work
Problem
General language models are weak at cyber threat-intelligence tasks, and reinforcement learning wastes compute when many training prompts give no learning signal.
Approach
- Built a pipeline for continual pre-training, supervised fine-tuning and GRPO reinforcement learning on Qwen2.5-7B, on six H200 GPUs with DeepSpeed ZeRO-2.
- Used a 72B model served on vLLM as a judge that gives partial credit.
- Designed task-specific reward functions and difficulty-based data filters across four CTIBench benchmarks.
Results
- Training steps with no learning signal fell from 80% to under 30%.
Problem
Enterprise clients want generative AI on their own infrastructure, with safety controls, not only a public API.
Approach
- Deployed production LLMs (Qwen-72B, Foundation-Sec-8B) on Red Hat OpenShift with vLLM tensor parallelism, GPU-aware scheduling and autoscaling.
- Red-teamed the endpoints against jailbreaks with Cisco AI Defense.
- Delivered a multimodal RAG shopping assistant with guardrails, a catalog-enrichment pipeline for product text, images and 3D models, and document and video Q&A assistants.
- Built a LangGraph network-compliance agent for about 10,000 managed devices, and a real-time voice agent.
Results
- Automated guardrail tests for the production LLM endpoints.
Problem
Enterprises need a governed way to build, deploy and monitor AI agents across SAP, ERP and CRM systems.
Approach
- Built the ML lifecycle tools of the platform: data ingestion, training, experiment tracking, model registry, deployment and live monitoring.
- Delivered them through a web UI and an open-source CLI, with CI/CD and role-based access control.
Problem
Interview practice with a person is expensive and hard to schedule; text chatbots do not feel like an interview.
Approach
- Built a real-time voice pipeline: streaming speech-to-text, an LLM and text-to-speech over WebSockets.
- Wrote 15 drill-specific LLM evaluators with strict JSON-schema outputs.
Results
- Live at prepvue.ai.
Problem
Two long-running projects to learn reinforcement learning and representation learning from first principles.
Approach
- montecarlo-zero: an AlphaZero-style chess engine in PyTorch with policy and value heads, trained by Monte Carlo tree search self-play.
- alphatft: a Transformer that scores Teamfight Tactics boards, trained on placement data from the Riot API.
- The chess network starts from human games and then improves by self-play.
02About
From the model to the system around it.
I build AI systems for enterprises, and I focus on the step that many AI projects never reach: production.
My work covers the full path: data, training and evaluation, model serving, and the product, security and monitoring around the model. At Velocis Systems, I build forecasting, agentic and LLM systems for enterprise clients on NVIDIA GPU infrastructure. Before that, I was a data scientist at CoreOps.ai, on a platform for enterprise multi-agent systems. I also run PrepVue, my own AI voice app for interview practice.
I studied Economics and Mathematics at Boston University. That background shapes how I work: I start from the business decision that a model supports, and I measure the result against a baseline.
How I work
- 01
Measure against a baseline
A model earns its place only if it beats the method it replaces. Foresight was benchmarked against 15 statistical and deep-learning models before it shipped.
- 02
Ground every answer in evidence
An AI answer that nobody can check is not useful to an operations team. My systems cite the evidence behind each claim and refuse figures that no tool returned.
- 03
Build security in from the start
Enterprise AI needs identity, guardrails and an audit trail on the first day, not after the pilot. I design with access control, red-teaming and monitoring from the start.
Experience
Velocis Systems
AI Engineer · Joined through its subsidiary Agilus Technologies
Sep 2025 – now
CoreOps.ai
Data Scientist
Sep 2024 – Sep 2025
Mphasis
AI/ML Intern · Client: Hewlett Packard Enterprise
Jun – Sep 2022
Education
Boston University
Bachelor of Arts, Economics and Mathematics
2026
Capabilities
- Models
- LLM fine-tuning (SFT, LoRA, GRPO, DPO)Agentic systemsRAGTime-series forecastingEvaluation
- Infrastructure
- vLLMNVIDIA NIMDeepSpeedMulti-GPU trainingKubernetes / OpenShiftDocker
- Product and data
- PythonFastAPIReactPostgreSQLMongoDBRedisVector databases
- Security and operations
- Cisco AI DefenseCisco DuoGuardrails and red-teamingSplunkCI/CD
03Writing
Notes from production
- Accuracy is not the finish line: taking a forecasting model to productionSix lessons from moving a spare-parts forecasting model from a benchmark to a system that planners use every day.3 min read
- Evidence first: designing an AI agent that operations teams can trustDesign rules from an agent that finds the root cause of enterprise network problems, and why each claim must point to its evidence.3 min read