Chirag Saraswat
Trained a 7 billion parameter language model from scratch. Now builds agentic AI systems for enterprise clients at production scale.
One of the few engineers in India who has trained a large language model from scratch, now applying that depth to production AI for enterprise clients.
Specializes in production grade LLMs, agentic AI systems and MLOps infrastructure. Background spans foundational model training, real time computer vision at broadcast scale, enterprise AI consulting on Azure, and teaching AI to more than 150 students across two institutions. International research exposure from Hanyang University, Seoul, and IIIT Delhi.
Download CV
BharatLLM
A 7 billion parameter Hindi-English-Hinglish language model, architected and trained from the ground up rather than fine-tuned from an existing base, a capability more commonly found at senior or staff engineering level.
- 01 Trained using PyTorch FSDP across 8 A100 GPUs, with custom tokenization, positional encodings and attention mechanisms
- 02 End to end data pipeline over 45TB+ of multilingual corpus via Apache Spark, with quality filtering and deduplication
- 03 Production inference via vLLM with TensorRT optimization: 100+ requests per second at 120ms P99 latency with dynamic batching
- 04 18% performance improvement on Indic language benchmarks against existing models
Projects worth a closer look
Vorra
SIGNATUREReal-time voice AI platform: one voice engine, a swappable reasoning layer per use case, across inbound, outbound, and support.
- Incremental low-latency speech, holds conversational flow instead of waiting on full responses
- Three reasoning layers per deployment type: sales, support, and domain-specific
- Live supervisor tools: real-time listen-in, whisper guidance, per-call confidence scoring
Virtual AI Office
SIGNATUREMulti-agent orchestration platform: one orchestrator directing a team of specialized agents, each with a defined role.
- Orchestrator routes and sequences tasks across specialized agents automatically
- Each agent scoped to a defined role and its own tool access
- Built for real operational workflows, not just a chat interface
Prompt-to-Dataset Engine
SIGNATUREMulti-modal synthetic data platform: text, image, audio, and tabular training data from a single prompt.
- Custom LoRA adapters fine-tuned on Flux and SDXL for image generation
- Distributed Apache Spark pipeline for quality filtering and deduplication at scale
- Built for Indian-language data, where usable training sets are thin across most modalities
IntelliAgent
Multi-agent financial analysis platform with coordinated research, analysis and trading roles.
- 28% improvement in trading signal accuracy
- Hybrid RAG (BGE-M3 + BM25) over 1M+ financial documents, 94% relevance
- Kafka pipeline ingesting 20K+ market events/min across 15+ exchanges
Broadcast Computer Vision
Real-time sports analytics pipeline built at Quidich Innovation Labs for live ICC cricket broadcast.
- 91% ball tracking precision across 45+ live matches
- 4K video at 30FPS with GPU-accelerated multi-object tracking
- Active-learning annotation pipeline, cut manual labeling by 75%
Enterprise AutoML Platform
No-code AutoML platform built at Dataviv Technologies, democratizing model training for non-technical users.
- 90% automation across CV, deep learning and NLP pipelines
- 50+ model architectures with distributed GPU allocation
- 500+ concurrent requests, sub-100ms, 98.9% uptime
Trains models from nothing. Ships the agents that run on them in production.
Where the work happened
Enterprise agentic AI on Azure: multi-agent framework design, deployment and scaling for client engagements, alongside deep AI research.
Worked across dataset generation, synthetic testing and end-to-end AI build tooling, including fine-tuning Flux and SDXL with custom LoRA adapters and a LangGraph research pipeline benchmarking 500+ ML papers weekly.
Architected a no-code enterprise AutoML platform and its MLOps backbone, and served as AI trainer at Mayo College, Ajmer.
Built an AI hiring platform serving 2,000+ interviews, fine-tuned Mixtral 8x7B with QLoRA, and shipped a real-time 3D lip-syncing interview avatar.
Mentored 150+ students in ML and NLP, and built an ML-based resume evaluation platform used by 500+ students.
Researched small object detection with YOLO-NAS and DETR for sports analytics, cutting training time from 60 to 8 hours via distributed training on NVIDIA DGX.
Built the real-time ball-tracking pipeline for live ICC broadcast, detailed under Selected Work above.
MERN and PERN stack development, React front ends over Node and Express, with Docker and CI/CD pipelines.
Designed and taught a hands-on AI course at Mayo College, Ajmer.
Designed and led 15+ cross-domain AI projects with a team of students under agile methodology, spanning computer vision, NLP, reinforcement learning, drones, and robotics, and built the AI course curriculum for Mayo College end to end.
Mentored 150+ students in ML, NLP and CV.
Led hands-on sessions on unsupervised learning and clustering, and built an automated resume-evaluation platform used by 500+ students, reaching an 85% completion rate in the specialized AI track.
One connected stack, applied AI down to the infrastructure underneath it.
B.Tech, Computer Science Engineering
Specialization in AI & Research
Exchange, Advanced AI Specialization (CV, NLP, RL). One of a small, competitively selected cohort.
Exchange, Machine Learning & Deep Learning Specialization. Selected among top 5%.