Adarsh S Mcold start
000
resolving runtime01/05
Open to workBengaluru, India

I build generative systems that ship — from 30B-parameter video models to language models small enough to run on your phone.

Generative AI / Deep Learning Engineer · 5+ yrs
epoch3/3loss0.214
Selected work
  • PyTorch
  • Diffusion Transformers
  • Rectified Flow
  • LCM
  • Step Distillation
  • RAG
  • FAISS
  • Pinecone
  • Fine-tuning
  • On-device SLMs
  • CNNs
  • XGBoost
  • Edge AI
  • Azure
  • MQTT
  • Multimodal
  • Cross-attention
  • CUDA

Selected work

Six systems, from a 30B multimodal model to edge inference in the field. Each one opens with the constraint that shaped it and the number that says whether it worked.

Case studies open with the constraints, not the stack.

Tokenizer, live

Type anything. This runs GPT-2's actual byte-level BPE in your browser — same merge table, same pre-tokenisation regex. Colour marks where words begin.

gpt-2 · 50,257 vocab · not loaded
Tokens

Load the tokenizer to see this text split into the units a model actually reads.

Tokens
Characters
103
Chars / token
Word starts

Chars-per-token is why non-English text costs more to serve: the same sentence in a script GPT-2 saw little of fragments into far more tokens, and you pay per token. Try sample 2.

About

I am a generative AI engineer in Bengaluru. Five years in, split between shipping models people actually use and the systems work that makes them affordable to run.

I started in observability at Infosys — telemetry, dashboards, root cause. It is an unglamorous place to learn that a system you cannot measure is a system you cannot improve, and the habit followed me into machine learning: I build the evaluation before I build the model.

Since then it has been military IoT at DRDO, where inference had to happen at the edge because the uplink could not be trusted; two and a half years of freelance LLM and vision work, where a client will not accept a demo; and now a ~30B multimodal model that generates video and synchronised audio in a single forward pass.

The through-line is inference budgets. A model that only works at fifty sampling steps is a research result, not a product — getting it to eight is the part I find interesting. I am also building Sigil, which runs a small language model entirely on-device so a spoken agreement never leaves your phone. Constraints like that make the design decisions legible.

PortraitGenerative AI / Deep Learning Engineer
Bengaluru, India

Skills

  • Daily
  • Comfortable
  • Exploring

Generative & multimodal

What I work on now

  • Diffusion Transformers (DiT)
  • Rectified flow matching
  • Latent Consistency Models
  • Step distillation / Reflow
  • Cross-modal attention
  • Audio-visual synchronisation

Language models

Adaptation, and making them small enough to ship

  • Fine-tuning (GPT, BERT)
  • Transfer learning
  • Prompt engineering
  • On-device SLMs

Retrieval

Opinionated: evaluate against the manual process

  • RAG pipelines
  • FAISS
  • Pinecone

Vision & classical ML

Where a smaller model is the right answer

  • CNNs
  • Medical imaging
  • Scikit-Learn
  • XGBoost

Edge & IoT

From DRDO and IBM — inference where connectivity is not assumed

  • Edge AI
  • MQTT
  • Raspberry Pi / Arduino / ESP32
  • IBM Watson IoT

Platform & observability

Enough to own a system end to end, and to prove it improved

  • Python
  • PyTorch
  • Azure Monitor / Log Analytics
  • Application Insights
  • Power BI
  • AWS / Azure / IBM Cloud

Experience

  • Apr 2026 — now

    AI Research Engineer / ML Engineer

    Stealth AI startup · Remote

    Core contributor to AURA, a ~30B dual-branch flow-matching DiT generating 2K video and synchronised stereo audio in one forward pass. Cut sampling from 50 steps to ≤8 via step distillation.

    • DiT
    • Rectified flow
    • LCM
  • May 2026 — now

    Founder

    Sigil · India · Remote

    Building an AI-native mobile app that turns oral agreements into binding digital micro-contracts, using on-device SLMs so voice never leaves the phone.

    • On-device SLM
    • Privacy-first
    • Product
  • Jan 2024 — now

    ML Engineer & Gen AI Specialist

    Fiverr · Freelance · Remote

    Fine-tuned LLMs for code generation and intent recognition (+30% contextual accuracy) and built RAG pipelines over financial data (−40% manual effort) with FAISS and Pinecone.

    • RAG
    • Fine-tuning
    • CNNs
  • Dec 2023 — May 2026

    Project Engineer

    Centre for AI & Robotics, DRDO · India · Remote

    Handpicked to design Military IoT systems with edge computing and ML threat detection, enhancing battlefield intelligence by 40%. Led 10+ engineers to a fielded deployment at 100% standards compliance.

    • Edge AI
    • Real-time analytics
    • IS 15500
  • Mar 2021 — Oct 2023

    Senior System Engineer

    Infosys · Mysore, India · Hybrid

    Real-time observability across production systems: incident response down 35%, uptime up 20%, downtime down 25% through root cause analysis.

    • Azure Monitor
    • Power BI
    • Telemetry
  • Apr 2020 — Aug 2020

    IoT Intern

    IBM · Hyderabad, India · On-site

    IoT solutions across AI, cloud and edge devices — sensor data into AWS, Azure and IBM Cloud, plus predictive maintenance models for real-time analytics.

    • MQTT
    • Edge AI
    • Watson IoT