Saksham Parashar

Searching huge metadata repositories used to mean knowing exactly what to type. Now you just ask.

• Delivered a RAG platform with Sentence Transformer embeddings that answers natural-language questions about CDC metadata repositories, cutting search time by 60–80%.
• Re-architected the LLM integration with streaming responses to cut perceived latency.
• Shipped a Streamlit interface whose answers stay grounded in the retrieved metadata.

• Azure OpenAI
• LangChain
• FAISS
• Pinecone
• Streamlit

Code review that shows up on every pull request, without anyone having to ask.

• Three agents split the work: one fetches the pull request, one gathers context around it, one analyses the code.
• Parses PR diffs, builds contextual prompts and gives feedback on correctness, readability and best practices.
• A multi-step evaluation workflow with cost optimisations cut API calls by 57%.
• Wired into the pull-request workflow, so reviews run automatically.

• Azure AI Foundry
• Agent Framework
• Prompt engineering
• Azure DevOps

A writer and a critic, both language models, running on a local GPU.

• A generator model reads a dataset's metadata and writes insights the way a data scientist would.
• A second model plays critic: it scores each insight out of 10 and says what to fix.
• They loop until the score reaches 9 (or the retries run out), then the agent commits the result to its own git branch.

• Open-source LLMs
• Hugging Face
• PyTorch + CUDA
• Git automation

Brands write briefs; creators write bios. This reads both and finds the matches.

• Matches brand campaign briefs to the most relevant content creators using semantic search over sentence embeddings (BGE-M3).
• Ranks creators by similarity, with filters to narrow the shortlist.
• An interactive Streamlit app with CSV export of the results.

• Sentence Transformers
• BGE-M3
• Streamlit
• Pandas