Available for AI/ML collaboration

Building intelligent systems
that reason, see & scale.

Applied AI engineer & team lead who takes frontier research to production — RL post-training (GRPO) of SOTA LLMs & VLMs, multimodal pipelines fusing 5+ models, and architecture-level tuning for fast inference. Full stack: from model weights to Kubernetes.

▋
View Work → ↓ Download Résumé Get in touch
✉thegigasurgeon@gmail.com 📞+91-9915503368
Delivered Value

Outcomes, not just models.

0
PDFs processed daily by AI platform I lead
0
Document extraction accuracy — up from 35%
0
Developer-hours saved weekly via LLM orchestration
0
Inference pipeline speedup from optimization
0
FPS gain — real-time video analytics on the same GPUs
0
Post-processing time eliminated in geospatial CV
0
Engineers led across validator jurisdictions
0
SOTA models fused in one production pipeline
About

Research-grade AI, engineered for production.

I'm an AI engineer working across NLP, vision, and everything multimodal — at the intersection of deep-learning research and large-scale systems. I pretrain and fine-tune state-of-the-art LLMs and VLMs (LLaMA 3.3, Qwen 3, Gemma, Gemini 3.5 Pro, GPT-4o) using LoRA/QLoRA and RL post-training with GRPO to build reasoning and thinking capabilities into domain models.

Today I lead the development of an enterprise-scale, agentic AI document intelligence platform processing 80K+ PDFs a day — multi-agent text + visual reasoning that fuses 5+ specialized models (detection, OCR, VLMs, reasoning LLMs, embeddings) in a single microservices-based application, with tool-calling, subagents, and skills. Along the way I've built RAG systems, multi-LLM orchestration infrastructure, and extraction pipelines that took accuracy from 35% to 99.85%.

I also go below the framework line — tinkering with model architectures, quantization, and graph-level optimization to hit aggressive inference targets. Those instincts come from years of computer vision: real-time sports analytics, satellite segmentation, pose estimation, and edge deployment. I care about the whole stack, from a novel architecture to a quantized engine shipping 3× the FPS.

At a glance

RoleAI/ML Engineer
LocationDelhi, India
FocusNLP · Multimodal · CV
TeamLed 8 engineers
EducationB.Tech CSE · 8.28 GPA
Featured Work

Systems I've shipped in production.

// Work is proprietary — the visuals below are architecture illustrations, not source or screenshots.
AI document intelligence architecture diagram
Real Brokerage · Lead

Agentic AI Document Intelligence Platform

An enterprise-scale agentic platform that reads unstructured real-estate documents through multi-agent text + visual reasoning — 5+ specialized models (detection, OCR, vision VLMs, reasoning LLMs, embeddings) coordinated with tool-calling and subagents across microservices. Processing 80K+ PDFs a day with structured, auditable output.

VLMsMulti-agentTool-callingYOLOv9Doctr OCR
read the case study
Problem — Real-estate transactions hinge on dense, wildly inconsistent documents. Manual review was slow, error-prone, and impossible to scale with brokerage growth.
Approach — Decomposed review into cooperating agents over 5+ specialized models in one pipeline — object detection, OCR, a vision VLM for layout & visual grounding, reasoning LLMs, and embedding models — coordinated through tool-calling and subagents. Fine-tuned open models (LLaMA 3.3, Qwen, Gemma) on the real-estate domain with LoRA/QLoRA and GRPO-based RL post-training, alongside hosted Gemini 3.5 Pro / GPT-4o.
Deployment — Microservices on Kubernetes across AWS and GCP, with GPU-backed inference using continuous batching; self-hosted GPUs serve the fine-tuned open models.
Impact — Extraction accuracy lifted from 35% to 99.85%; the platform now processes 80K+ PDFs every day in production.
Multi-LLM orchestration router diagram
Real Brokerage

Multi-LLM Orchestration Infrastructure

An intent-routing layer that dispatches each request to the right model — Gemini 3.5 Pro, GPT-4o, LLaMA 3.3, Qwen 3, Gemma — behind one interface. Fine-tuned for robust agentic workflows and RAG, it automates internal processes and saves thousands of developer-hours every week.

Intent routingRAGFine-tuningGemini 3.5 ProGPT-4o
read the case study
Problem — Every internal automation was hand-rolling its own LLM calls: duplicated prompt logic, no routing, no fallback, no shared evaluation.
Approach — A central intent-routing layer over Gemini 3.5 Pro/Flash, GPT-4o, LLaMA 3.3, Qwen 3, and Gemma behind one interface — LoRA fine-tunes per task family, graceful model fallback, and structured outputs by default.
Deployment — Runs on Kubernetes with autoscaling across AWS and GCP (Vertex AI); hosted and self-hosted models mixed per cost/latency profile.
Impact — Became the org's default path to LLM automation, saving thousands of developer-hours every week.
Real-time sports analytics detection and pose diagram
Streamingo AI

Real-Time Sports Analytics Engine

A live video pipeline combining object detection, multi-object tracking, pose estimation, and court localization — built on Swin, VideoMAE, and ViT, then optimized with TensorRT and ONNX to deliver 3× the frame rate for real-time analysis.

Object detectionTrackingPoseTensorRTONNX
read the case study
Problem — Frame-accurate sports analytics needs detection, multi-object tracking, pose estimation, and court localization running together at full frame rate — on a fixed GPU budget.
Approach — Built the pipeline on Swin, VideoMAE, and ViT backbones, then compiled every stage into TensorRT engines with graph fusion, INT8/FP16 quantization, and batched inference across streams.
Impact — 3× the frames-per-second on identical hardware, unlocking true real-time analysis instead of offline processing.
Retrieval-augmented generation pipeline diagram
Real Brokerage

End-to-End RAG Pipeline

A precise retrieval-augmented generation system: intent routing, semantic chunking, embedding generation, and hybrid vector + keyword retrieval, fused into context-optimized prompts that drive LLM reasoning and structured output across multiple sources.

Semantic chunkingHybrid retrievalEmbeddingsRerankingStructured output
read the case study
Problem — Naive RAG retrieves plausible-but-wrong context: bad chunk boundaries, single-mode search, and prompts stuffed past the model's attention budget.
Approach — Intent routing decides whether and where to retrieve; semantic chunking keeps passages coherent; hybrid dense + keyword retrieval with reranking; context optimization fuses multi-source evidence before the LLM reasons over it.
Impact — Grounded, structured answers over heterogeneous document corpora — the retrieval backbone for the document-intelligence platform's 80K+ PDFs a day.
Claude Agent SDK automation loop diagram
Real Brokerage

Claude Agent SDK Automation

An agentic automation layer built on the Claude Agent SDK — triggered by PRs, incidents, and webhooks. It reads diffs, runs tests, spawns subagents, and posts reviews or fixes in a tool-calling loop, removing hours of manual engineering toil.

Claude Agent SDKTool-callingSubagentsPR reviewWebhooks
read the case study
Problem — Code review, incident triage, and routine ops were consuming senior-engineer attention on work that is 80% mechanical investigation.
Approach — An agentic layer on the Claude Agent SDK, triggered by PRs, alerts, and webhooks. The agent reads diffs, runs tests, inspects logs, spawns focused subagents, and posts its findings — a full tool-calling loop, not a one-shot prompt.
Impact — Automated first-pass PR reviews and incident investigations, returning senior-engineer hours to feature work.
Satellite image segmentation input and mask diagram
Attentive AI

Satellite Image Segmentation

Owned two CV products (Falcon & Respod) end-to-end, training state-of-the-art segmentation networks — HRNet and Vision Transformers — on satellite imagery. Lifted model accuracy by 12% while cutting post-processing time by 90%.

HRNetVision TransformerSegmentationGeospatial
read the case study
Problem — Extracting property features from satellite imagery demanded pixel-precise segmentation, but model output still needed heavy manual post-processing before it was usable.
Approach — Trained HRNet and Vision Transformer segmentation networks end-to-end and owned both CV products (Falcon & Respod) across the company — data, training, serving, and the post-processing stack.
Impact — +12% model accuracy and 90% of post-processing time eliminated, turning a bottleneck into a throughput multiplier.
Experience

Where I've built.

AI/ML Engineer · Real Brokerage
Oct 2024 — Present
  • Pretrained & fine-tuned SOTA LLMs and VLMs with reasoning capabilities (LLaMA 3.3, Qwen 2.5 72B, Qwen 3, Gemma) using LoRA/QLoRA and GRPO-based RL post-training, tailored for real-estate auto-review.
  • Lead development of an enterprise agentic AI document intelligence platform processing 80K+ PDFs daily — 5+ models fused per pipeline via multi-agent text + visual reasoning with tool-calling, subagents & skills, deployed as microservices.
  • Built scalable Multi-LLM Orchestration Infrastructure, saving thousands of developer hours weekly.
  • Fine-tuned Gemini 3.5 Pro/Flash and GPT-4o for robust multi-LLM agentic workflows.
  • Engineered an end-to-end RAG pipeline: intent routing, semantic chunking, hybrid vector retrieval, multi-source context aggregation & structured output.
  • Integrated the Claude Agent SDK for automated PR reviews, incident investigation & webhook-driven workflows.
  • Boosted extraction accuracy from 35% → 99.85% with YOLOv9, Doctr OCR & custom VLMs; achieved 200% pipeline speedup via graph optimization, batched inference & quantization.
Deep Learning Engineer · Streamingo AI
Nov 2022 — Oct 2024
  • Developed temporal event-understanding systems with Swin, VideoMAE, MVD, ViT, BEiT.
  • Built a real-time sports analytics engine — object detection, multi-object tracking, pose estimation & court localization.
  • Designed zero-shot & few-shot classification with CLIP, Detic & Prototypical Networks for rapid scaling to unseen classes.
  • Deployed optimized inference with TensorRT & ONNX, delivering 3× FPS.
Computer Vision Research Engineer · Attentive AI
Apr 2021 — Nov 2022
  • Built & trained SOTA networks for satellite image segmentation (HRNet, Vision Transformer).
  • End-to-end ownership of CV products (Falcon & Respod) across the company.
  • Improved model accuracy by 12% and cut post-processing time by 90%.
Software Engineer · Nokia
Jul 2020 — Mar 2021
  • Created a meta-learning platform for dynamic model selection & hyperparameter optimization.
  • Integrated Explainable AI (SHAP, LIME) for model interpretability & compliance.
Under the Hood

How I take models to production.

🚀 Deployment

Multi-cloud and self-hosted — each model lands where its cost, latency, and data-privacy profile fits best.

AWS · EC2 / ECS / SageMakerGCP · Vertex AI / GKE Kubernetes + DockerMicroservice architecture AutoscalingSelf-hosted GPU clusters

⚡ Inference Optimization

The 200% pipeline speedup and 3× FPS came from stacking these — measured at every step, not guessed.

INT8 / FP16 quantizationAWQ / GPTQ TensorRT enginesONNX + graph fusion Continuous batchingKV-cache reuse

🎯 Model Adaptation

Right-sizing intelligence: RL post-training for reasoning, low-rank adaptation for the domain, and architecture surgery when off-the-shelf isn't fast enough.

GRPO (RL post-training)LoRA / QLoRA DistillationPruningArchitecture modification LLaMA / Qwen / GemmaGemini 3.5 Pro / GPT-4o tuning
Toolkit

Technologies I work in depth.

🧠 LLMs, NLP & Post-Training

LLaMA 3.3Qwen 3Gemma Gemini 3.5 ProGPT-4o GRPO (RL)LoRA / QLoRA Reasoning modelsMultimodal fine-tuning

🤖 Agentic AI

Multi-agent systemsTool-callingSubagents Claude Agent SDKAgent skillsLangChain Webhook-driven agents

🔎 RAG & Retrieval

Intent routingSemantic chunkingHybrid retrieval Vector storesEmbeddingsContext optimization Structured output

👁️ Vision & VLMs

YOLOv9Doctr OCRSwinVideoMAE ViT / BEiTCLIPDeticHRNet Detectron2Pose estimation

⚡ Inference & MLOps

TensorRTONNXQuantization Batched inferenceGraph optimizationMicroservices Kubernetes / DockerEdge / Raspberry Pi

💻 Languages & Frameworks

PythonC++JavaScript PyTorchTensorFlowOpenCV SLAM / ORBv3
Selected Projects

Things I've built for fun & research.

01

Basketball Video → 3D Mapping

Reconstructed player trajectories in 3D from monocular sports footage using deep video transformers and classical SLAM. Added temporal event summarization via transformer-based description generation.

PyTorchVisual SLAMORBv3BlenderVideoMAE
02

Tekken with Pose

A camera-driven interface that maps real-world body actions into virtual gameplay using pose estimation, latency-aware gesture recognition, and real-time control mapping.

YOLO-NASPose EstimationRaspberry Pi
03

Charlie — AI Virtual Assistant

An edge-deployable assistant with multi-modal recognition: speaker ID, facial login, offline speech-to-text, and contextual response generation — all running on-device.

Raspberry PiSTTFaceNetPython
Contact

Let's build something intelligent.

Open to conversations about AI/ML engineering, applied research, and hard problems in LLMs, VLMs, and computer vision.

✉ Email me in · LinkedIn 📞 Call me
Email thegigasurgeon@gmail.com
Phone +91-9915503368