Anubhab Pradhan | AI Engineer

AI Engineer focused on building reliable intelligent systems across LLM applications, RAG, evaluation, production ML, cloud platforms, and developer tooling. I work at the intersection of AI systems and software engineering, turning experimental ideas into measurable, maintainable, and production-ready software. Outside the terminal, chess, binge-watching, and travelling are my favourite ways to disappear — always curious about new ideas, unfamiliar places, and things I haven't figured out yet.

More about me

WHAT I DO

LLM Applications & RAG Systems

Designing retrieval-augmented generation pipelines and intelligent language-model applications that handle real-world input, context, and failure modes.

Evaluation Pipelines & ML Infrastructure

Building structured evaluation frameworks that measure model behaviour, retrieval quality, and system reliability across repeatable test scenarios.

Production AI Workflows & Cloud

Taking AI prototypes to production — event-driven pipelines, failure handling, monitoring, and cloud deployment on AWS and similar platforms.

Developer Tools & Platforms

Creating tools, dashboards, and platforms that help other engineers build, evaluate, and ship AI systems with confidence.

EXPERIENCE

Lamatic.ai | Applied AI Intern

Jan 2026 – May 2026

Engineered AI/ML product workflows and evaluation pipelines for LLM applications, structuring repeatable evaluation paths to assess model behavior and reliability.
Built event-driven integrations between AI workflows and application systems, automating multi-step execution paths and reducing manual intervention across workflows.
Developed production-oriented components for AI applications with an emphasis on reliable execution, failure handling, and maintainable integration boundaries.
Contributed fixes and feature improvements to LangChain and Metaflow, extending practical experience from production AI systems into open-source infrastructure.

Hasprana Health Care Soln Pvt Ltd | AI Engineer

Jul 2026 – Present

Engineer AI systems spanning RAG retrieval, model evaluation, and production ML components, taking features from implementation through production-ready integration.
Design evaluation workflows for LLM applications that assess retrieval quality, model behavior, and system reliability across repeatable evaluation scenarios.
Build and refine AI service components with attention to reliability, maintainability, failure handling, and practical production constraints.
Audit and review AI-service code for correctness, maintainability, and production readiness, identifying implementation issues before they propagate into downstream workflows.

SKILLS

Python
Java
JavaScript
TypeScript
Rust
C
C++

Python
Java
JavaScript
TypeScript
Rust
C
C++

PyTorch
TensorFlow
scikit-learn
NumPy
Pandas
LangChain
LangGraph

PyTorch
TensorFlow
scikit-learn
NumPy
Pandas
LangChain
LangGraph

Transformers
Hugging Face
MLflow
OpenCV
FastAPI
Next.js
Streamlit

Transformers
Hugging Face
MLflow
OpenCV
FastAPI
Next.js
Streamlit

PostgreSQL
SQL
Redis
Temporal
Docker
AWS
Git
GitHub
GitHub Actions
Linux
REST APIs
pytest
Ollama
Jupyter
OpenCode

PostgreSQL
SQL
Redis
Temporal
Docker
AWS
Git
GitHub
GitHub Actions
Linux
REST APIs
pytest
Ollama
Jupyter
OpenCode

PROJECTS

Kairos

IN-PROGRESS

Open-source platform for transparent RAG development, evaluation, experimentation, and explainable AI.

TypeScript
RAG
Explainable AI

RedOps

IN-PROGRESS

Production-grade LLM evaluation and red-teaming platform covering hallucination, jailbreak resistance, latency, cost, and token usage across providers.

Python
LLM Evaluation
Docker

More Projects

RESEARCH & PUBLICATIONS

RESEARCH / 2026

Benchmarking Small Language Models for Domain-Specific Question Answering

A Comparative Study of Phi-3-mini, Mistral-7B, and Gemma-2 on SQuAD v2

SLMQUESTION ANSWERINGSQuAD v2PHI-3-MINIMISTRAL-7BGEMMA-2

Benchmarks small language models for domain-specific question answering. Compares Phi-3-mini, Mistral-7B, and Gemma-2 on SQuAD v2 using Exact Match and F1 evaluation, while examining inference time and the generative–extractive mismatch in instruction-tuned models.

Compared Phi-3-mini-4k-instruct, Mistral-7B-Instruct-v0.2, and Gemma-2-2b-it under identical zero-shot conditions on SQuAD v2.

Evaluated 200 validation examples using Exact Match, F1 score, and average per-query inference time with a fixed seed for reproducibility.

The experiments expose a trade-off between QA accuracy and inference latency across the three SLMs, with no single model dominating every measured dimension.

The study identifies a generative–extractive mismatch: instruction-tuned models can produce semantically correct full-sentence answers that receive poor EM/F1 scores when evaluated against short extractive spans.

View Research Paper

OPEN SOURCE

CONTRIBUTIONS3 PRs

01 Ollamaollama

OLMo3 Terminal-Chunk Tool Call ParsingPR #18677 ↗

Fixed OLMo3 streaming tool-call parsing when a complete tool call arrives in the terminal chunk, ensuring final `done=true` responses are parsed instead of being emitted as raw assistant content.

GoLLM InferenceStreamingTool Calling

DeepSeek3 Partial Tool-Call Opening TagsPR #18682 ↗

Fixed DeepSeek3 streaming parser handling for tool-call opening tags split across chunks, preserving partial delimiters until they can be reconstructed into a valid tool call.

GoStreamingTool CallingParser Reliability

02 LangChainlangchain-ai

Post-Generation VerificationPR #36480 ↗

Contributed a post-generation verification component that validates LLM outputs through a Runnable-based fact-checking workflow.

PythonLLM ValidationRunnable

03 MetaflowNetflix

Artifact SerializationPR #3061 ↗

Improved error handling around artifact serialization failures, making failure paths more explicit and robust.

PythonML PipelinesError Handling

GITHUB ACTIVITY

loading…GitHub profile

LessMore

कर्मण्येवाधिकारस्ते मा फलेषु कदाचन।
मा कर्मफलहेतुर्भूर्मा ते सङ्गोऽस्त्वकर्मणि ॥ २-४७

Your duty is to uphold dharma
Never to claim the reward
Let not the promise of victory guide you
The battlefield summons, be relentless in action

- Bhagavad Gita - Chapter 2, Verse 47

AAnnuubbhhaabb PPrraaddhhaann