AI engineer working on agentic systems, retrieval and LLM applications for healthcare and fintech. I build production pipelines end to end: data extraction, RAG, multi-agent orchestration, evaluation and deployment. I also contribute bug fixes to open-source AI/ML libraries.
Currently
- Data Scientist at Martlet AI, building clinical document classification and an AI assistant for medical coders and physicians.
- Co-founder and CTO of Causal Markets, building AI analytics and agent tooling for prediction-market traders.
- Contributing to open-source AI/ML projects: docling, sentence-transformers, MTEB, txtai, PEFT and more (full log).
36 merged, 44 in review across 34 repositories. Full log with commits and diffs: MaiAditya/oss-contributions
| Project | PR | What I fixed |
|---|---|---|
| medplum/medplum | #11539 | Mapped Act status through PROCEDURE_STATUS_MAPPER and mapped active to in-progress, with tests. |
| neuml/txtai | #1624 | Set the configured name and description on tools built by smolagents.tool, keeping the function values if unset. |
| crewAIInc/crewAI | #9595 | Opted SaveArtifactTool out of caching with _never_cache_function, so each call saves a new version. |
| reflex-dev/reflex | #7880 | Added _forwarded_client_ip() to strip ports and brackets, normalise addresses and fall back to the peer address if invalid. |
| Arize-ai/openinference | #4416 | Mapped cached prompt and cache creation tokens to llm.token_count.prompt_details.cache_read and cache_write when non-zero. |
| huggingface/sentence-transformers | #4410 | Replaced constructor asserts with ValueErrors for empty input, length mismatch and non-0/1 binary labels. |
| neuml/txtai | #1621 | Caught the smolagents exception in ToolFactory.createtool so the docstring fallback runs, and added tests. |
| Arize-ai/openinference | #4367 | Recorded cached_tokens and reasoning_tokens as span attributes, matching the groq instrumentor. |
| huggingface/sentence-transformers | #4354 | Used the untruncated embedding dimension in MatryoshkaLoss and EmbedDistillLoss. |
| neuml/txtai | #1602 | Returned an empty result per query in the Faiss ANN backend's search when the limit is 0, and added dense and hybrid tests. |
| embeddings-benchmark/mteb | #5929 | Materialized models and tasks into lists once at entry so iterators and empty inputs work. |
| PostHog/posthog-python | #1277 | Matched the project root on directory boundaries: the path equals the root or starts with the root plus os.sep. |
| embeddings-benchmark/mteb | #5906 | Added an evaluated_splits argument to save_to_cache so only subsets that were actually run get run settings. |
| huggingface/sentence-transformers | #4397 | Resolved the default similarity function names on each call and rebuilt the CSV headers when they changed. |
| neuml/txtai | #1592 | Pinned llama-cpp-python==0.3.36 in the Windows dependency step of the build workflow. |
| neuml/txtai | #1579 | Added Cluster.limit to slice merged results by the query's LIMIT, falling back to the request limit. |
| neuml/txtai | #1577 | Added Indexes.implements and checked subindex ANNs up front so these calls raise NotImplementedError first. |
| neuml/txtai | #1571 | Added case keywords as conditionals and stopped treating the token after and/or as an alias. |
| neuml/txtai | #1569 | Moved null from Token.OPERATORS to the literals, next to true and false. |
| PostHog/posthog-python | #1267 | Marked the resolved exc_info as captured and skipped already-marked ones in Client and AsyncClient. |
| huggingface/sentence-transformers | #4383 | Added the main similarity to the STS and triplet evaluators, and made IR evaluators raise a ValueError at construction. |
| PostHog/posthog-python | #1256 | Appended output blocks and tool results per generation and promoted queued input at the next message_start. |
| huggingface/peft | #4157 | Fixed BOFT kernel shapes and made OFT Conv2d rotate the weight in forward, matching its merged output. |
| neuml/txtai | #1561 | Routed a [ after an operator to bracket() in compound() so these expressions translate correctly. |
| reflex-dev/reflex | #7866 | Split strings by Unicode code points when the separator is empty, including runtime separators. |
| embeddings-benchmark/mteb | #5880 | Returned an empty ModelResult with the model name and revision, without loading the model. |
| embeddings-benchmark/mteb | #5879 | Filtered cached and merged results to the task's splits and subsets after saving, leaving the cache file intact. |
| docling-project/docling | #5244 | Joined the section label and title into one heading, e.g. ## 1. Introduction. |
| neuml/txtai | #1550 | Counted nested parentheses in function arguments so each function ends at its own closing paren. |
| neuml/txtai | #1535 | Split signs from operators when tokenizing SQL. |
| embeddings-benchmark/mteb | #5870 | Added euclidean and manhattan support to AbsEncoder and a memory-bounded chunked manhattan_sim. |
| docling-project/docling | #5283 | Added <preformat> and <code> blocks as code items, keeping line breaks and the code language. |
| embeddings-benchmark/mteb | #5864 | Delegated pairwise similarity to the sentence-transformers model's own similarity_pairwise. |
| huggingface/sentence-transformers | #4313 | Scored paraphrase mining with the model's own similarity function. |
| docling-project/docling | #5078 | Kept image-only tables as pictures with their labels and captions. |
| neuml/txtai | #1523 | Made each similar() clause keep its own number of results when clauses are batched. |
Data Scientist, Martlet AI · Jun 2026 – present
- Production agentic architecture combining classical ML and deep learning to identify and classify medical documents for patient charts (93.58% accuracy).
- AI assistant for medical coders and physicians to explore patient cohorts, clinical histories and coding workflows in natural language.
- Both systems are deployed with US pharmaceutical companies and healthcare organisations.
Co-founder and CTO, Causal Markets · Jan 2026 – present
- AI analytics, agentic automation and trading-intelligence products for prediction-market traders: 100+ DAU and 524+ users within weeks of launch.
- Three AI microtools (insider-trading detection, copy-trading intelligence, wallet analytics) serving 1,500+ users in the Polymarket community.
- Selected for Entrepreneur First, Antler founder cohorts and the Polymarket Builder Program.
Jr. Data Scientist, John Snow Labs · Jan 2025 – Jan 2026
- Multimodal agentic pipeline extracting images, tables and text from medical research PDFs for downstream NLP, retrieval and QA.
- LLM-powered SQL QueryBot over the OMOP clinical database using MCP, for natural-language querying of patient data.
- Reinforcement-learning-based guardrails framework against prompt-injection attacks, shipped as a private installable module.
- AI orchestration with Pydantic-AI, OpenAI APIs and Hugging Face models for RAG and structured generation.
Co-founder and CTO, SummonLM · Aug 2025 – Dec 2025
- AI-powered GEO automation and agentic content intelligence; 15+ product development contracts in the first months.
ML Engineer Intern, Nurix AI · Jul 2024 – Dec 2024
- LLM-powered agentic chatbot with Temporal and Conductor orchestration (20% faster responses).
- RAG-based agentic framework for API management (95% accuracy) using better chunking and re-ranking.
- Custom Airbyte source connector used in 10+ projects; copilot tools handling 80% of customer queries.
Indian Institute of Technology, Roorkee · 2020 – 2024 B.Tech, Metallurgical and Materials Engineering · CGPA 8.00/10
- Languages: Python, C++, SQL
- AI/ML: PyTorch, TensorFlow, Hugging Face, LangChain, Pydantic-AI, RAG, agentic and multi-agent systems, LLM evaluation and guardrails
- Models and APIs: OpenAI, Claude, Gemini, Mistral, Llama
- Infrastructure: AWS, Google Cloud, Docker, Redis, Temporal, Airbyte

