ADYA PRASAD
I work on machine learning research, with a long-term goal of contributing to LLM inference research, LLM alignment, and machine learning systems design. My current experience include model evaluation, post-training, reinforcement learning, and multi-agent systems. At BaseThesis Lab, I study chain-of-thought unfaithfulness using residual-stream activations and linear probes, exploring how models reason internally and whether their verbal explanations faithfully represent that process.
I also bring 4 years of experience building Python backend systems. This engineering background helps me turn research questions into reproducible experiments, build reliable training and evaluation pipelines, and take technical work from an initial hypothesis to analyzed results. I am currently seeking a research internship or job at a startup where I can contribute across ML research, research engineering, and system design.
in my intermediate schooling I publish a STEM website it name was Indiecore where I provided educational material, quizzes and PYQ. I built the website with a CMS and scaled it to 1 million plus impressions and 200k traffic. It Shutdown in May 2019 - Dec 2022
Currently as a side hustle I am building a lab Creators AI Lab where I am innovating and developing tools and SaaS for Content Creators.
♡ A soul, committed to the intersection of backend dev, ML research, and startup culture ♡
EXPERIENCES
Research Fellow | Sentient Future Org
Septemeber 2026 - Present
- Problem Statement: LLM-driven causal modeling of AI loss-of-control risk.Mentor: Swaptik Chowdhury
- Using a custom built library of AI loss-of-control scenarios. And converting it into causal loop diagrams by extracting decision-relevant variables and the relationships between them, using an LLM pipeline grounded in established literature-extraction methods
- Role: Building the technical pipeline end to end. This includes implementing the literature-extraction methods to pull decision-relevant variables and relationships from the scenario library, building the LLM pipeline that converts these into causal loop diagrams, and testing the SaferAI quantitative risk estimation methods on top of the resulting diagrams
Applied AI Research Intern | BaseThesis Lab
May 2026 - Aug 2026
- Working on research: Evaluating Chain-of-Thought Unfaithfulness from Residual Stream Trajectories and linear probe
- Initiated and co-led an empirical research project investigating whether a model's internal residual-stream representations faithfully align with its verbalized chain-of-thought during reasoning task.
- Designed and curated a diverse evaluation dataset spanning mathematics, reasoning, history, and safety to improve statistical robustness and probe generalization. Scored AUROC 0.65, demonstrating residual-stream trajectories provide useful signals for detecting unfaithful reasoning
Full Stack Developer | NowTreat Healthcare
Jun 2023 - Nov 2025
- Ownership of official website and integrate a patient post treatment care and booking service, along with feedback loop using WordPress CMS, and JavaScript to support post-treatment guidance, and improve patient engagement.
- Small team, high ownership! Along with marketing team and founder, designed an marketing funnel for better user experience and engagement and converting online traffic to potential patents. Collectively achieve 38% more engagement and 26% more footfall
Founder: Creators AI Lab
2026
An applied AI lab for innovating and developing ai applications and SaaS for media research, generation, and automation. Focus audience: content creators and influencers.
- Cupid AI Suite: An app with Swarm of AI agents for diverse and user
personalized research and content creations using LangGraph, FastAPI, API Services,
PostgreSQL, Redis, ChromaDB, and Docker.
Designed a multi-stage safety and response evaluation pipeline combining heuristic validation, policy checks, and quantitative monitoring metrices to filter user inputs and validate LLM outputs before responding to user. - Luffy Create:Luffy Create is a free and open source AI video editing, video
creation, image editing and graphic designing web app.
Luffy Create has been built with the principle of providing a lightweight and fast editing experience. Goal: Our aim is to make 'Luffy Create' the best editing app for animations, transitions, and effects.
Freelance Works (Developer)
2022 - present
- Yatidhara [FMCG]:Work with founder in supply chain of Ayurvedic and healthy (gift style) products. I work on designing premium healthy products (snacks) packaging and end to end ecommerce web development.
- Focus Point [Digital Library and Study Space]: Along with founder Mukesh Sonkar (a frontend engineer)and I (as a full stack developer) build platform (library marketplace). The Focuspoint serve as marketplace for both library owner and students. Library owner can list their library and students can find the best library as per their need in nearby
OPEN SOURCE CONTRIBUTIONS
Hugging Face: Peft
1. Resolve issue #2944: ensure_weight_tying respects config + improved tests {PR
#3171}
updated src/peft/utils/other.py in the set_additional_trainable_modules() and _get_module_names_tied_with_embedding() functions to respect the model config and also create test file
updated src/peft/utils/other.py in the set_additional_trainable_modules() and _get_module_names_tied_with_embedding() functions to respect the model config and also create test file
2. Experiment method comparison #2310 : adaptation_prompt {PR #3301}
Conduct experiement with finetuning, benchmarking the LLaMA-Adapter (adaptation_prompt) for Conversation AI and Math reasoning. I configured custom adapter config and training parameter with the Llama-3.2-3B model, ran multiple training pipelines on the MetaMathQA dataset, and evaluated its memory efficiency, accuracy, loss and reasoning retention against baseline methods to provide benchmarked result to community.
Conduct experiement with finetuning, benchmarking the LLaMA-Adapter (adaptation_prompt) for Conversation AI and Math reasoning. I configured custom adapter config and training parameter with the Llama-3.2-3B model, ran multiple training pipelines on the MetaMathQA dataset, and evaluated its memory efficiency, accuracy, loss and reasoning retention against baseline methods to provide benchmarked result to community.
ARTICLES
1. AushadiNet-GATv2: Graph Attention Network for Predicting Drug-Drug Interactions in Polypharmacy
-I researched and presented AushadhiNet-GATv2, a graph neural
network
architecture for predicting drug-drug interactions in polypharmacy. System addresses
critical
limitations of existing DDI prediction methods through multi-view molecular representation
learning,
hierarchical GATv2 message passing with residual connections, dual-head classification for
interaction detection and mechanism prediction, and integration with patient-specific
cardiovascular
risk profiling
CREATORS AI LAB, PROJECTS
Cupid: Agentic Social Media Content Generation Platform
A multi-agent system to generate platform-specific social content from
natural-language prompts. Four LangGraph agents coordinate the pipeline: supervisor
agent, personalization agent, research agent, and composer agent. The personalization agent
decomposes user intent into five orthogonal search queries, the research agent runs parallel web
retrieval through DuckDuckGo with trafilatura-based content extraction, the composer agent
distills sources into atomic facts and generates three angle-diverse variants in parallel, and
the trends service handles scheduled background ingestion from Google News RSS ranking articles
per-user with BM25 and recency decay, hybrid memory layer and LLM chain for fallback: Groq,
HuggingFace and local heuristics.
multi-agent orchestrationMCPRAGLangchain
Luffy Create: AI Video Creation App
Everything! You need to create a professional videos Luffy Create is a free, open-source video editor built for creators who want real control with both online and offline editing, AI-assisted scene edits, automatic captions, automatic subtitle translation, inbuilt Latex, Python & Manim animation, Code sandbox, and faster multiformat export. No cloud, no account, no watermark.
Version 1.3.5 ships builds for Windows, macOS, and Linux. Users can also run from source on any platform with Node.js installed.
Version 1.3.5 ships builds for Windows, macOS, and Linux. Users can also run from source on any platform with Node.js installed.
desktop appmultimedia processingstream management
Audio Augment | Audio Changer Studio
A very simple and fast audio editing studio that works with great precision. It has been designed and developed for content creators to edit their audio at their fingertips without any hassle.
User upload an MP3 or WAV and get professional cleaned and edited audio. It provide two independent processing stages: Noise Removal and Audio Augmentation
real time audio editlocal whisperaudio encoding
RESEARCH PROJECTS
AushadhiNet-GATv2 DDI Model
AushadhiNet-GATv2 is a graph neural network model that predicts drug-pair
interactions with probability scores and classifies interaction types (potential adverse drug
reactions) across 86 mechanisms. The system alerts physicians and patients before prescribing or
taking medications, preventing harmful drug combinations.
Dataset use: Drug Interaction Data (Training): ddis, drug_smiles, drug_names, ddi_type_mapping CVD Patient Data (Validation): cardio_base 70,000 patient record
Dataset use: Drug Interaction Data (Training): ddis, drug_smiles, drug_names, ddi_type_mapping CVD Patient Data (Validation): cardio_base 70,000 patient record
#Research
ProjectGNNDrug
MoleculesDrug Interaction Prediction
Research CLI Agent (Hybrid RAG)
A production-oriented
The agent work in CLI, receive a input with different
command-line research agent I built to tackle the engineering problems of reliable agentic systems. The project intentionally emphasizes runtime correctness, observability, failure containment, and evaluation over frontend development. The agent work in CLI, receive a input with different
--flags commands and spawns workers bounded loop and ultimately researched and provided relevant response.
Flow:model proposal → typed decision parser → tool registry and schema validation → policy and budget checks → tool execution → structured observation → validation → result OR next bounded decision
Agent HarnessHybrid RAGReal-time Research
Provenance: Production ML Agent to tests upstream data changes in pipeline
A producer or monitoring agent emits a structured, allowlisted change
signal. Provenance reads the source asset and complete downstream lineage from DataHub.
It selects only the affected model and constructs a bounded experiment.
The runner mutates an isolated evaluation copy and compares model behavior. Numeric policy—not an LLM—returns PASS, WARNING, or BLOCK. Provenance saves JSON and Markdown evidence and publishes the result to DataHub.
The runner mutates an isolated evaluation copy and compares model behavior. Numeric policy—not an LLM—returns PASS, WARNING, or BLOCK. Provenance saves JSON and Markdown evidence and publishes the result to DataHub.
#Research
ProjectMLOpsML TestingData Provenance
GLC Platform | Loan Assesement
Green Loan Cycle platform streamlines ESG-linked loan assessments automatically, identifying and measuring key metrics, making data capture on performance 30% easier with LMA framework and AI agents.
The "GLC Platform" provides a uniform platform for both borrowers and lenders, incorporating a compliance/monitoring framework that applies from loan origination throughout the entire life cycle of Green Loan Principles (GLP) and the LMA framework. It also offers services such as loan data resolution, loan assessment, sustainability checks, key stakeholder evaluation, loan balance sheet generation, and AI-assisted advice. This helps both borrowers and lenders reduce pressure, streamline processes, improve communication, and maintain compliance + reporting.
The "GLC Platform" provides a uniform platform for both borrowers and lenders, incorporating a compliance/monitoring framework that applies from loan origination throughout the entire life cycle of Green Loan Principles (GLP) and the LMA framework. It also offers services such as loan data resolution, loan assessment, sustainability checks, key stakeholder evaluation, loan balance sheet generation, and AI-assisted advice. This helps both borrowers and lenders reduce pressure, streamline processes, improve communication, and maintain compliance + reporting.
#Research
ProjectMLOpsML TestingData Provenance
SKILLS AND TECH STACK
LANGUAGES:PythonTypeScriptR
langJavaScriptSQL
FRAMEWORKS:FlaskDjangoFastAPIReact
jsNext js
TOOLS:Apache Ecosystem (learning: Kafka, Spark, and
Hadoop)Hugging Face
EcosystemOllamaMongoDBPostmanPrometheusn8nCanva
DB:PostgreSQLSQLiteMongoDB
RedisChromaDB
LIBRARY /
MODULES:PyTorchTensorFlowNLTK
(NLP)NumpyPandasMatplotlibPEFTTransformersOpenaiLangchainLlamaIndexSQLAlchemyChromaDB
PLATFORMS:DockerKubernetesAWS
Ubuntu OSKali OS
EXTRACURRICULAR ACTIVITIES & ACHIEVEMENTS
-
Deloitte Australia Technology Job Simulation on Forage
- Develop a program to normalise different telemetry JSON data of IOT machine (Simple ETL pipelines)
- Craft a complete "software development proposal" for a real-time monitoring live dashboard
-
Founder (Developer and Author) of a STEM educational website: indiecore.in
- Run and manage a STEM educational website, specially focused on high school and intermediate students to provide free of cost study material for their exam preparation, computer skills, and reasoning skills development
- Create 500+ resources and materials: blogs, webstories, quizzes, mock tests, pdfs and flashcards. Achieved 1 million plus impressions and than 200k traffic
- Teaching assistance for freshers, in fundamentals of computer application subject. Office hours work on managing course materials and leading discussion group.
- Hackathon Winner: Bye to Beat on Devpost | Project: AushadhiNet-GATv2 DDI Model | Predicting Drug-Drug Interactions
- 200+ Submission Competitive Programming: HackerRank, LeetCode, CodeForces, and TensorTonic
LEARNINGS (EDUCATION)
St Thomas International School:
High School (9 and 10)Intermediate (11 and 12).
VBSP University:
B.Tech(Computer Science & Engineering)Harvard CS50x:
Computer Science, Cybersecurity, Artifical Intelligence and R*.Master Diploma in Computer Application
from Amrtikosh Info-tech institute (duration: 2 years). Fundamentals of Computer programming, C, JavaScript, TypeScript, Bash Scripting, Web Development, Office Automation.Certifications:
- YouTube full course playlist I take: Standford C229 Machine Learning (Autumn 2018), Stanford CS230 Deep Learning (Autumn 2025), Standford CS224R Deep Reinforcement Learning (Spring 2025)
- Amazon SageMaker AI Getting Started by AWS Training - [Oct 2025]
- Build Your Own Chatbotby IBM SkillsBuild- [Jan 2025]
- Market research by Great Learning - [May 2024]
- Google Technical Writing Course One and Two by Google Developers - [June 2022]
- Content Creation, Optimization, SEO and Strategies by HubSpot Academy