Generative AI engineer building efficient multimodal and on-device systems.
I work across models and the systems that ship them: diffusion transformers, LLM adaptation, retrieval, and inference optimization. My focus is making capable models faster, cheaper, and reliable enough to use outside the lab.
Portfolio · LinkedIn · Email · Bengaluru, India
Core contributor to a ~30B-parameter foundation model that generates 2K video and synchronized stereo audio in a single forward pass. Reduced sampling from 50 steps to ≤8 using rectified flow and step distillation.
PyTorch Diffusion Transformers Rectified Flow Step Distillation
Sigil — private, on-device AI
Founder and builder of an AI-native mobile app that turns spoken agreements into signed, tamper-evident micro-contracts. Speech recognition and language-model inference run entirely on-device; the complete workflow works offline after setup.
Private source · Flutter whisper.cpp On-device SLMs Ed25519 SQLite
SemantiCache — semantic LLM caching
A lightweight Python library that caches LLM responses by meaning rather than exact prompt matches, with local or OpenAI embeddings and in-memory or SQLite storage. On its included 60-query benchmark, it avoids 47% of API calls at threshold 0.85 with zero false positives.
Benchmarks · Python Sentence Transformers SQLite
33 merged upstream pull requests across 12 projects, focused on correctness, reliability, and regression-tested fixes. View contribution history.
| Project | Merged PRs | Areas improved |
|---|---|---|
| BlockNote | 11 | Editor behavior, clipboard handling, DOCX and email exports |
| folium | 5 | GeoJSON and TopoJSON popups, grouped layer controls |
| Feast | 4 | Redis event ordering, registry lookups, materialization reporting |
| statsmodels | 3 | SARIMAX, prediction validation, covariance scaling |
| DVC | 2 | Run-cache hashing, mixed pull and checkout targets |
| Hydra | 2 | Dictionary overrides, user-facing error behavior |
Also merged in k6, matplotlib, Gymnasium, Recharts, astroid, and dulwich.
- AI Research Engineer / ML Engineer — Stealth AI startup · 2026–present
Multimodal generation and inference optimization for AURA. - Founder — Sigil · 2026–present
Privacy-first, on-device AI for spoken agreements. - ML Engineer & Generative AI Specialist — Fiverr · 2024–present
LLM fine-tuning, retrieval pipelines, and computer vision for client systems. - Project Engineer — Centre for AI & Robotics, DRDO · 2023–2026
Edge ML, real-time analytics, and threat detection for military IoT. - Senior System Engineer — Infosys · 2021–2023
Production observability, telemetry analysis, and incident response.
- Generative AI: PyTorch, diffusion transformers, rectified flow, latent consistency models, step distillation, cross-modal attention
- Language systems: LLM fine-tuning, RAG, FAISS, Pinecone, on-device SLMs
- Engineering: Python, TypeScript, Dart, C++, Flutter, React, Next.js
- Cloud and edge: Azure, AWS, IBM Cloud, MQTT, Raspberry Pi, Arduino, ESP32, observability
Open to collaboration on multimodal generation, inference optimization, on-device AI, and impactful open source. Get in touch.


