Software Engineer, CVE Lead
March 2025 - [Present], Bangalore
Tune AI was acquihired by Ema Unlimited in March, 2025.
- • Leading post sales implementation for several F50 clients. End to end lifecycle from discovery to delivery.
- • Built 2+ internal tools. Reduced effort by multiple hours/week/employee.
After moving to Post Sales team, I realised the challenges of project management. Led building CVE-One AI for post sales team to automate project management. Used by multiple teams to track updates for a project.
Head of Research
Dec. 2020 — Feb. 2025, Chennai, Bangalore & San Francisco
GenAI for Enterprises
At Tune I have seen every part of the startup journey from ideation to discovering PMF to failing in monetization to eventually landing contracts with some of the best organisations in their market. Backed by the best: Accel, Together Fund, Techstars, Venture Catalysts, Cornerstone Venture Partners, Chennai Angels, and Astarc Ventures.
Key Achievements
- • Successfully delivered multiple enterprise projects from ideation to production combined revenue of $140K+
- • Pioneered the implementation of Chain of Thought (CoT) and several other context engineering techniques for Tune Chat
- • 500K+ users on Tune Chat and cutomers finetuned 100+ models on Tune Studio
- • Among the first companies to deploy Meta Llama 2 in production within 24 hours of its release
- • Deployed Meta Llama 3 in production within 1 hour of its release. Recognized by Meta as one of their India partners.
- • Did several events for Tune AI helping build the Bangalore AI/ML community.
AI Research
Led AI solutioning working directly with Abu Dhabi F1 organizer (Ethara), world’s largest scientific contents product (Clarivate), and Intel. Projects became biggest revenue drivers for Tune AI.
- • Led two teams totaling 8 people
- • Developed AI Agents reducing sales TAT from 14 days → 5 minutes by auto generating 200+ slide long PPT presentation and is highly personalized for each prospect and potential event, following the design guidelines. Works from inside MS Teams to answer any question via chat interface.
- • Architected large data processing pipeline to run inference on 100K+ documents/day for extraction task with 96% accuracy.
- • Context engineering systems to ensure 100% grounded AI results
- • Build a novel transformer model to run AlphaGo style Monte Carlo Tree Search based algorithm for autoregressive tasks. It was trained to perform both classification and regression. Implemented by writing custom kernels for Nvidia Triton & vLLM.
- • Solutioned LLM training with 375GB+ data
- • With Intel we delivered the whitepaper on OpenVino, delivering 20x faster Mask-RCNN
The achievements might be mild by world standards. But for a team of 20 young hackers this was a wild success after 4+ years of hard work. Truth is, just like any other startup, no one told us any process and we discovered / built our own, all the way 0 to 1. For a 24 year old, this was the FAFO learning phase of life. I worked on everything I could get my hands on.
Product developement for features in Tune and NimbleBox:
- • Blob : Client facing agent and configurable assistant in Studio
- • ChainFury : Backend tool for chain of thought prompting
- • Koro : Experimental Llama 2 derived architecture for unified model for Next token generation (NTG), Embedding model (LM-Head removed), Regression Model for linear prediction.
- • Silk : Distributed code storage and execution engine that could run any DAG workflow of arbitrary python functions. Ensured retries and rollbacks on errors.
- • Saturn : A layer on jupyterkernelgateway that could be connected with K8S to run code as a service. This was our take on AWS lambda.
- • LMAO : General purpose logging and rule based alerting system with UI rule builder.
- • Vedang : Remote python code parsing engine, used to determine how to run a python code.
- • Armoury : RAG system backend written completely in go
- • NN-One : Our attempt to train an LLM. Curated 375GB+ of data and attempted to train a model on it. Learnt a lot about hard problems in software, the hard way.
- • Lot of eventually forgotten code
Things I'll remember:
- • Built an nbox.Operator python toolkit that could be used to deploy code as a job/service on CPU/GPU. This was eventually perfected by Modal.
- • We shipped a version of NimbleBox with album style cover image for each Project.
- • The joy of people when they discovered something about the product.
- • Travelled to US twice and spoke to Jeff Dean about our work.
- • Pleasure of no shame after loosing fear of rejection in sales.
Other than the points mentioned above, in the startup, I've helped with sales, design, branding, customer success, etc.
Mentor & Judge - Hack MIT 2024 & PennApps XXV
September 2024, MIT, Cambridge & University of Pennsylvania, Pennsylvania
- • Gave a technical workshop on Tune's AI research
- • Mentored teams building AI products for first responders, fashion designing, etc
AI Consultant
NPAW, Spain
Dec. 2020 — March 2021, Remote
Research and develop a Grafana plugin agent that converts user input in natural langauge to charts. The novel solution used a decision tree to parse the query parameters based on prompting. Deployed model sharded GPT-2 1Bn on 2 Nvidia-3090 GPUs to maximise the context length for each input query.
ML Engineer
July 2019 — Nov. 2020, Ahmedabad
- • Built ML solution to convert unstructured business data like documents (scanned, digitised) to structured knowledge using supervised and unsupervised machine learning algorithms
- • Built services on top of this extracted data like rules management, abnormality detection along with a full learning system
- • Worked with planets largest supply chain companies (Maersk & CMA-CGM) to deliver PoCs, clients based in Europe, APAC and LATAM regions
- • Involved in product design, development and customer interaction with multiple clients
Summer Intern
Kaaenaat
April 2018 — Oct. 2018, Bangalore
- • Upgradation of Kount, a product which resulted in improved results by performing live traffic analysis like Dynamic Trajectory Clustering and Anomaly Detection by looking at raw footage using machine learning, was implemented on an embedded device for on-the-edge application
- • Designing and making of an in-house application for faster Image Segmentation compared to classical methods
Machine Learning Intern
May 2017 — July 2017, Pune
Worked on Connecticus' cognitive platform NESSA to develop machine learning based approach to making a FAQ module. Implemented Facebook's Memory Networks and a variety of NLP tasks such as POS tagging, stemming and lemmatisation of language.
Summer Intern
May 2017 — July 2017, Nagpur
Designed and developed a python toolkit for rapid deployment of infographics in company's E.R.P. Solutions. Toolkit was made in python3, used several external libraries like, ggplot2, pandas, numpy etc.
Projects apart from work
Project अर्थ (Artha)website
Started July 2025 [Ongoing]
Building world's largest digital enclyclopaedia for ancient Indian literature.
This is my personal project I have worked on for a while. I developed the backend and vibe-coded the frontend. Spend time reading, curating and digitising books, editing and compiling the digital Encyclopaedia. First time in 10 years I'm making software just for me.GitHub
AI Researcher
Nov. 2020 — Aug. 2021, Remote, India
I spent a lot of time in-between jobs working on AI application research.
- • Research on RL agents that do not need the perfect board state to play superhuman chess. This would demonstrate that NNs have internal representation capacity to solve complicated problems giving only traces of information.YouTube
- • New research directions for weather modelling that uses ground based sensor data instead of solely relying on weather satellites, which causes issues like cold-bias and wrong temperature prediction.GitHub
tuneapiA swiss knife python package for building application with LLMs. Very opinionated. Available in Python and Typescript
vrikshamTree based conversation storage engine interface and Cypher implementation in go.
asteaSimple pure-python AST engine with lazy lookup and code traversal
ChainFury🦋 Production grade chaining engine behind TuneChat. Self host today!
nboxThe official python package for NimbleBox. Exposes all APIs as CLIs and contains modules to make ML 🌸
yQLprotobuf powered RPC but REST instead sockets, creates client & server stubs. Used in production at NimbleBox.
general-perceiversgperc or How to general purpose perceivers! Train models by just throwing data at it. gperc simplifies using PerceiverIO an architecture by DeepMind
chess_lmSupervised Pre-training a chess engine on moves only, can it surpass me? Can it learn board representation internally? Can this learned vector be used with tree search?
dall-e-babyOpenAI's dall-e is a kick ass model that takes in a natural language prompt and generates an images based on that. Now I cannot recreate the complete Dall-E so I make the baby version of it trained in CIFAR10-100 dataset. If Dall-E is picasso this is well ... baby.
spyqlQuery data on the command line with SQL-like SELECTs powered by Python expressions
vaayuvidhaCode to run a novel recurrent-GNN model for weather prediction.
freeciv-pythonPython agent to play Freeciv game.
Education
B. Tech. in Electronics and Telecommunication
National Institute of Technology Raipur
May 2015 — May 2019, Raipur
Activities and societies: Training and Placement Cell NIT Raipur. Entrepreneurship Cell of NIT Raipur. Photography Club of the NIT Raipur. Manager at the Entrepreneurship Cell (E-Cell) of the NIT Raipur.
Was part of a team that organised E-Summit 2016, 2017 and 2018 Central India's largest Entrepreneurship Event. Responsible for the conduction of the Wall Street Event in the Annual Techno-Management Fest of the College (Aavartan 2016).
Worked on building AI powered Indian sign language detector for Texas Instruments' challenge. Implemented transformer network for speech to text for Microsoft's challenge (MSAIC).GitHub