Mumbai, India · open to opportunities

Shubh
Sareen.

I engineer production-ready AI & inference pipelines.

IIT Bombay engineer designing reliable machine learning systems — from zero-cost local LLM inference and multimodal ingestion to wrapping probabilistic outputs in deterministic programmatic constraints.

View Projects Get in Touch
0M+

orders analyzed in production analytics

0k+

students served by campus platform

0%

faster page loads on institute portal

₹0

LLM API cost after local migration

Featured build

Hierarchical Video RAG Pipeline

Turns multi-hour YouTube playlists into structured notes and a queryable knowledge base — batched hierarchical summarization with a retrieval layer on top.

Try live All projects
Now
Autonise

Research engineering intern — production AI pipelines & LLM reliability

Seasons of Code

Mentoring Stable Diffusion: probability → VAEs → DDPM → CLIP

How I work

Probabilistic models fail in production unless you constrain them. My default playbook: shift work off the model onto deterministic code, self-host when the economics demand it, and design evaluations that can prove failure — not just suggest success.

// AI reliability is a design problem, not a model problem.

Education
IIT Bombay

B.Tech Chemical Engineering
Minor — CMInDS

8.72/ 10 CGPA
01.

Where I've Worked

RESEARCH ENGINEERING INTERN · AUTONISE

PRODUCTION AI GENERATION PIPELINES & RELIABILITY RESEARCH

APPLIED ML // BENGALURU
ActiveMAY 2026 — JULY 2026

Working across the full stack of a production AI system — generation pipelines, LLM reliability research, evaluation methodology, and inference infrastructure.

[01]

PROGRAMMATIC SVG & ORGANIC MOLECULAR COMPILERS

PROGRAMMATIC SVG & ORGANIC MOLECULAR COMPILERS
Identified through iterative testing that raw SVG generation from autoregressive LLMs consistently fails because models lack spatial lookahead and absolute coordinate positioning l...
[02]

INTRANSITIVE LLM COMPARATORS & DETERMINISTIC PAIRWISE RANKING

INTRANSITIVE LLM COMPARATORS & DETERMINISTIC PAIRWISE RANKING
Investigated optimizing comparison complexity from O(N²) round-robin overhead down to O(N log N) using merge-sort and ELO ranking systems powered by LLM comparators. Our initial a...
[03]

DocLayout-YOLO REGION SEGMENTATION & MULTIMODAL INGESTION

DocLayout-YOLO REGION SEGMENTATION & MULTIMODAL INGESTION
Architected a custom multimodal text and layout extraction pipeline to automate question ingestion from complex exam papers. Integrated DocLayout-YOLO to perform initial region se...

AI reliability is a design problem, not a model problem.

ML RESEARCH INTERN · INFO-ME

DWELL-TIME RECOMMENDATIONS & EMBEDDING ALIGNMENT METRICS

RECOMMENDATION SYSTEMS
JAN 2026 — FEB 2026

Built and evaluated a two-stage content recommendation pipeline with a rigorous focus on honest evaluation methodology.

[01]

BEHAVIOR TELEMETRY RETRIEVAL & Llama 3B RERANKING

Engineered an embedding-based content recommendation pipeline by designing a dynamic user-profile embedding mechanism built on real-time behavior telemetry. Formulated a dynamic '...
WEB CONVENOR · INSTITUTE SPORTS COUNCIL

INSTITUTE SPORTS WEBSITE

IIT BOMBAY // STUDENT COUNCIL
Active2025 — PRESENT

Developed and maintained the official institute sports council website to serve as a centralized hub for sporting events and information.

[01]

WEBSITE DEVELOPMENT

WEBSITE DEVELOPMENT

Built a responsive sports website that helps students stay updated with sports events, tournament schedules, and relevant contacts.

[02]

PORTAL MAINTENANCE

Handled regular content updates and ensured seamless portal operation to provide an accessible interface for institute residents.

ML MENTOR · SEASONS OF CODE 2026

STABLE DIFFUSION MENTORING PROGRAM

IIT BOMBAY // SUMMER PROGRAM
ActiveMAY 2026 — PRESENT

Selected as mentor for IIT Bombay's flagship summer ML program, leading a 12-week deep-learning implementation track.

[01]

PROBABILITY FOUNDATIONS TO DDPM & CLIP CONDITIONING

End-to-end Stable Diffusion implementation project. Guiding students through: Probability Foundations → VAEs → DDPM → CLIP Conditioning. Building onboarding resources and weekly co...
02.

Skills & Tools

ML SYSTEMS

LLM Systems & PipelinesProgrammatic IR (LLM → Code)Evaluation MethodologyInference OptimizationRanking Systems (Swiss-Tournament)Recommendation SystemsDiffusion Models (DDPM, VAEs)

ENGINEERING

Pipeline ArchitectureHybrid Local/Cloud InferenceEmbedding-based RetrievalAPI IntegrationsDeployment (Vercel, Railway)Frontend (React / Next.js)

STACK

PythonMatplotlib · RDKitvLLM · LoRA · MoETypeScript (Working Knowledge)Linux

MINDSET

Systems ThinkingFailure Mode AnalysisMetric Design & CritiqueBenchmarking & ExperimentationResearch ↔ Engineering Bridge