VIT Bhopal '27 · AI/ML Engineering Intern @ AmberFlux EdgeAI · India
Portfolio · LinkedIn · Email · Hugging Face
I build the boring-but-hard parts of AI products: retrieval that is measured, agents that are observable, and memory that survives beyond one conversation.
NOW AI/ML Engineering Intern @ AmberFlux EdgeAI
FOCUS Production RAG · Multi-agent systems · LLM evaluation
SEEKING 2027 full-time AI Engineer · Applied AI · SDE + AI roles
EDUCATION B.Tech CSE, VIT Bhopal · 2023—2027 · CGPA 8.61If you have two minutes:
01 DoCopilot → production RAG, hybrid retrieval, evaluation, guardrails
02 Argus → LangGraph orchestration, HITL review, LangSmith tracing
03 ContextCore → FastMCP, persistent memory, agent tool execution|
Multi-tenant RAG platform
Document Q&A with tenant-scoped hybrid retrieval: BM25, dense vectors, RRF fusion, and cross-encoder reranking. Includes async PDF/TXT ingestion, prompt-injection detection, PII redaction, and source-cited streaming. 89.2% correctness
90.5% relevance
2.86s average latency
40 evaluated queries |
Multi-agent research engine
Supervisor-led research workflow coordinating planner, researcher, critic, writer, and supervisor agents. Built for explicit routing, human review, streamed execution, and traceable decisions. 5 specialist agents
3 research-depth tiers
30–90s report turnaround
HITL review via LangGraph interrupts |
|
Stateful memory agent
A dual-memory architecture that separates conversation checkpointing, semantic recall, and persistent user profiles. A FastMCP server exposes task and note tools directly to the agent workflow. 17/17 evaluation tests passing
3 memory layers
MCP callable tool server |
Parameter-efficient summarization
LoRA fine-tuning of FLAN-T5-base on SAMSum, updating only 2% of the model parameters. 49.01 ROUGE-1
72.25 BERTScore F1
42.51 METEOR
2% trainable parameters |
BUILDING Agent Guard — static analysis for safer agent workflows
HARDENING Evaluation harnesses for retrieval, tool calling, and routing
INTERESTED IN AI systems that fail visibly, recover predictably, and improve from tracesPython AST CLI GitHub Actions · in development
A static-analysis tool for identifying common failure patterns in agent code:
→ unbounded retry loops
→ unsafe shared state in concurrent fan-out branches
→ LLM / HTTP calls without explicit timeouts
→ missing failure boundaries in agent workflowsagents LangGraph · LangChain · MCP / FastMCP · Tool Calling
retrieval Qdrant · FAISS · BM25 · Hybrid Search · RRF · Reranking
backend Python · FastAPI · PostgreSQL · Redis · Celery
delivery Docker · Azure Container Apps · GitHub Actions · Linux
evaluation LangSmith · Ragas · LLM-as-Judge · ROUGE · BERTScore
ml PyTorch · Transformers · PEFT / LoRA · scikit-learn
languages Python · C++ · TypeScript · SQLoffline: DSA · gaming · sci-fi / thrillers · manhwa · EDM / lo-fi / metalBuilding systems with evidence, not buzzwords.