About
I'm Vicky Suman, a Senior Data Science Engineer based in Gurugram, India. Over nearly a decade of building AI systems across risk intelligence as sass and enterprise platforms, I've learned that the most important skill in this field isn't knowing the best algorithm, it's knowing how to turn a messy, high-stakes problem into a system that reliably delivers value at scale.
Not in notebooks. Not in demos. In production.
At Ontic Technologies, I lead the architecture and delivery of enterprise AI platforms that security teams depend on every day. The problems I work on don't have clean datasets or obvious solutions. They have noisy signals, tight latency budgets, real clients, and real money on the line. That's the environment I thrive in.
Cutting Through the Noise — Risk Event Intelligence
Security teams at enterprise companies were drowning. Thousands of raw threat signals flooding in every hour, no way to tell signal from noise, and clients paying for intelligence that wasn't actually intelligent.
I was tasked with architecting a platform from the ground up that could ingest, deduplicate, cluster, classify, and geo-locate security events — in real time, at enterprise scale, with enough precision to actually change how analysts made decisions.
I built a real-time embedding pipeline on Kafka and Triton Inference Server communicating over gRPC, storing vectors in Qdrant and MongoDB to power semantic search and trend clustering. On top, I engineered an LLM-based contextual location extraction layer with reverse geocoding, and a hypothesis-driven severity classification system using few-shot prompting via DSPy — with prompt optimization running through GEPA, SIMBA, and Decision-Tree optimizers to stay reliable at scale.
The platform launched as a revenue-generating product at ~$7.5K per client, processing 10,000+ messages per hour. Alert noise dropped significantly — analysts went from drowning in raw signals to working with actual intelligence.
Seeing Clearly — Denoising Face Detection at Scale
Our computer vision pipeline was generating too many false-positive image matches. Beyond the accuracy problem, every false positive was triggering a costly third-party API call — quietly compounding into a serious monthly burn.
I needed to rebuild the pipeline in a way that was both smarter and leaner — reducing errors without increasing infrastructure complexity.
I designed a two-stage denoising system combining YOLOv11 ONNX with a fine-tuned Vision Transformer (ViT), deployed on Triton Inference Server with batched inference. The architecture meant we only escalated confident detections — dramatically reducing noise before it ever hit the API layer.
False-positive matches fell sharply, and by eliminating redundant third-party calls, we saved $5K/month in infrastructure costs. Better engineering, directly mapped to business impact.
Teaching AI to Evaluate Itself — GenAI Metrics Platform
As LLM usage expanded across our product suite, we had a growing blind spot: no consistent, structured way to measure whether our models were actually performing well across different modules and tasks.
I set out to build an internal evaluation platform that could generate structured, backend-compatible metrics automatically — without requiring manual review at every step.
I fine-tuned LLMs using QLoRA and served them via TGI and DSPy, designing the system to output structured evaluation data in API-compatible format — plugging directly into existing backend workflows without adding friction to the development cycle.
Teams could now monitor LLM performance across modules systematically. What was an invisible risk became a measurable, manageable signal.
Building the Foundation — Data Infrastructure That Doesn't Break
Great AI models sitting on fragile data infrastructure will always eventually fail in production. Across multiple projects, I kept running into the same structural problem: pipelines that couldn't scale, monorepos that couldn't be maintained, and data lakes that couldn't support real-time use cases.
I took on the responsibility of building the data foundation that the rest of the AI stack could rely on.
I designed and deployed AWS-based data lake pipelines using Lambda, Glue (PySpark), and Hudi on S3 for real-time event processing. I also established a Monorepo architecture using Pants, enabling modular builds, dependency isolation, and faster multi-service development workflows — and stood up production-grade LLM inference on Kubernetes using TGI to serve multiple modules at scale.
The infrastructure became the platform — stable, observable, and modular enough that new AI capabilities could be shipped without rebuilding the foundation each time.
Where It Started — Cyber Risk and the Cost of Being Wrong
At RMS (Moody's Analytics), I spent years building causal models for cyber risk — estimating Business Email Compromise likelihood by company size, sector, and jurisdiction; building multilingual NLP pipelines using M2M-100 and mBART-50 to extract threat intelligence across languages; and designing custom NER systems with CRF and SpaCy to surface affected organizations, attack types, and financial impact from open-source feeds.
The difference was the stakes. These weren't product features — they were models informing financial decisions for insurers and risk analysts worldwide. Being wrong had real-world consequences. That pressure sharpened something in me that hasn't dulled since: a deep, almost stubborn respect for the gap between a model that performs on a benchmark and a model that's genuinely trustworthy in production.
Even earlier, at Phenom People, I was building AI chatbots with RASA and Dialogflow before conversational AI was mainstream — learning early that understanding user intent is as much a design problem as a technical one.
What Drives Me
I hold an M.Sc. in Applied Statistics and Informatics from IIT Bombay and ranked All India 60 in IIT JAM. There's a mathematician underneath all the engineering, and that foundation quietly shapes how I think about every model, every tradeoff, every system design decision.
I've led teams, defined technical direction, and shipped products that clients pay for and depend on. I take ownership seriously — not because I'm told to, but because I genuinely care whether the thing I build actually works for the person using it.
My core stack spans Python, Docker, Kubernetes, AWS, MongoDB, Kafka, Triton, DSPy, Hugging Face, Dagster, and TGI/vLLM — but I think of tools as means, never as identity. The right tool is the one that solves the problem cleanly and holds up six months later.
My long-term ambition is to keep pushing at the frontier of agentic AI and LLM systems — not just deploying what exists, but contributing to what comes next.