01
Foundation Models
Pre-training, post-training, long context, multimodal reasoning, and task-balanced VLM curricula.
building multimodal systems
About
Anirudh
Srinivasan
Senior Research Engineer · Vision-Language Models Team Lead
I build sovereign vision-language foundation models from scratch — spanning data, pre-training, post-training, evaluation, and the ML systems that make them scale.
Vision-Language Models
Document AI
ML Systems
Visual Grounding
Multilingual AI
01 · What I work on
From pixels to useful systems.
My work sits where multimodal modeling, document intelligence, and large-scale ML infrastructure meet.
02
Document Intelligence
OCR, parsing, layout understanding, grounding, VQA, reconstruction, and structured extraction.
03
ML Systems
Distributed training, large-scale generation, data systems, efficient storage, and multimodal evaluation.
02 · What I've been building
The Patram ecosystem.
A family of sovereign document VLMs, agents, evaluation systems, and infrastructure — from task-specialists to multi-task foundation models.
Sovereign Document AI
PATRAM
VLM ecosystem
Task-specific VLMs
PatramVQA
Patram-GrounderGrounded VQA
Patram-LayoutLayout 2.0
Patram-OCROCR 3.0
Patram-CoTGrounded CoT VQA
Multi-task Foundation VLMs
Patram 1.5
Patram 2.0
Evals
PatramBench
Patram-Eval
DarshinAI
Infrastructure
Patram-Core
Patram-Ingest
Patram-DB
Patram-FS
Patram-Infer
Agents
PatramParse
PatramRAG
The common thread: build the models, agents, evals, and systems needed for sovereign multimodal intelligence end-to-end.
Explore Patram →03 · My journey
A lot of exploring, then a lot of building.
I started by trying very different corners of ML. That breadth eventually converged into multimodal foundation models and systems.
IIT Hyderabad
B.Tech · Computer Science
Institute Gold Medal for Excellence in Academics and Co-curricular Activities · Rank 1 / 1103.
Exploration
Across the ML stack
Speech, quantum optimization, autonomous driving, neuromorphic AI, vision, motion, and compiler representations.
Hackathons
Inter-IIT + build sprints
Bosch, ISRO, DevRev, MPhasis, and Milan — rapid prototyping, competition ML, and a lot of late-night debugging.
Research
Documents + multimodality
Research spanning document understanding, visual grounding, layout analysis, reconstruction, and multimodal evaluation.
Today
BharatGen
Building sovereign VLMs from scratch across data, training, post-training, evaluation, inference, and deployment systems.
#1
IEEE Signal Processing Cup 2024
CVPR · ECCV · ICDAR
Research venues
Institute Gold Medal
IIT Hyderabad · Rank 1 / 1103
04 · Earlier explorations
I like learning by building.
Before document VLMs became the center of my work, I explored problems across speech, vision, quantum ML, neuromorphic AI, compilers, motion, and rapid-build competitions.
Research explorations 9 projects · speech, vision, quantum ML, neuromorphic AI, compilers, motion Explore
Motion · Generative AI
MotionGPT for Yoga
Motion-as-language modeling and training with in-house yoga MoCap data.
Open project → Neuromorphic AINeuromorphic AI Survey
From Hebbian learning and Hopfield networks to SNNs and cellular neural networks.
Open project → Quantum MLQuantum Spherical K-Means
Qiskit-based spherical q-means applied to multi-depot vehicle routing.
Open project → Optimization · ClusteringNeuro-Fuzzy GMMs
ANN reformulations of EM and neuro-fuzzy clustering for Gaussian mixtures.
Open project → SpeechROBOVOX
Far-field speaker recognition for noisy robotic environments · IEEE SP Cup winner.
Open project → 3D VisionLiDAR Segmentation
Semantic segmentation and upsampling pipelines for autonomous-driving point clouds.
Open project → Quantum OptimizationQuantum VRP
D-Wave CQM and QUBO formulations for vehicle-routing problems and variants.
Open project → Computer VisionSingle-Image HDR
Three-stage CNN reconstruction for clipping, non-linearity, and quantization reversal.
Open project → Compilers · RobustnessIR2Vec Robustness
Adversarial training and program-level attacks for learned compiler representations.
Open project →Hackathons & competitions 5 builds · Inter-IIT Tech Meets, Milan, agents, quantum, vision Explore
Inter-IIT Tech Meet 12.0
DevRev · Agent 007
Tool-selecting LLM agent with retrieval, ranking, and open-model generation.
Open project → Inter-IIT Tech Meet 12.0MPhasis · Quantum
D-Wave DQM business-rules engine for airline schedule-change optimization.
Open project → Silver · Inter-IIT 11.0ISRO · Moon Mapping
GAN-based super-resolution for Chandrayaan TMC-2 imagery and lunar-atlas generation.
Open project → Silver · Inter-IIT 10.0Bosch · Age & Gender
Age and gender estimation from low-resolution surveillance feeds.
Open project → Gold · MilanMovie Review System
Full-stack movie platform with opinion summarization and sentiment-based review clustering.
Open project →05 · Community
Teaching, mentoring, reviewing.
I enjoy explaining difficult ideas, helping people build, and contributing back to research communities.
Teaching
Head TA · Data Structures
TA · Introduction to Programming
Mentoring
Epoch · AI/ML & Data Science Club
Sunshine UG mentorship
Reviewing
Document AI · VLM benchmarks
ICDAR · AAAI
06 · Beyond the lab
The rest of the context window.
Models are fun. So are cricket, concerts, films, travel, and disappearing into a topic until I understand it properly.
Cricket
Huge Virat Kohli fan. RCB heartbreak is part of the training schedule.
Racket sports
Table tennis, badminton, pickleball—and just enough competitiveness.
Travel
New places, new food, new activities, and an unreasonable number of random facts.
Music
Playlists, concerts, and letting the soundtrack decide the mood.
Films
If there is a good release playing, there is a decent chance I am in a theatre.
The Backward Pass
I enjoy understanding ideas from first principles and explaining
them clearly—especially deep learning, foundation models, and ML systems.
Read the blogs →
I like understanding how intelligent systems work all the way down—and then building them all the way up.