Ankish Nayak

Ankish Nayak

LLM / AI Engineer
India
Full-time Hybrid

About Me

AI/LLM Engineer with 2+ years of experience building practical Generative AI systems and backend services. I work primarily with Python, FastAPI, LangChain, LangGraph, RAG, vector databases, LLM evaluation, and MLOps. At…

Experience

Digital Specialist Engineer (GenAI & LLM Agents)

Infosys · Pune, India
Aug 2024 - Present · 2 years 2 months

<p>Architected a modular ReAct loop engine (Thought → Action → Observation → Answer) with JSON step parsing, loop breakers, and a 20-iteration cap — preventing agent deadlocks and runaway iterations.</p><p> Built an LLM-agnostic client layer (DeepSeek, OpenAI, Anthropic, Gemini, Groq, Ollama, custom) via LiteLLM with tenacity-based 5× exponential-backoff retries and mock fallback, enabling zero-code model swaps.</p><p> Implemented SSE streaming (6 typed events, per-step duration/iteration metadata) via async generators on FastAPI with a hand-rolled ReadableStream client in a React frontend, delivering real-time step-by-step visibility.</p><p> Designed an extensible tool registry (10 pluggable tools + MCP): web search with Tavily/Google/DuckDuckGo fallback, sandboxed Python execution, AST-safe calculator, HTTP, file ops (path-traversal & symlink-escape protection), and episodic + semantic memory via ChromaDB.</p><p> Hardened security (AST-whitelist evaluation, method/scheme allowlists, sandbox path resolution, rate/size limits) and exposed FastAPI endpoints (chat + SSE, health, tools, sessions) with Loguru logging and Pydantic settings.</p><p> Wrote a deterministic offline test suite (36 pytest tests + 2 integration suites via a scripted-LLM mock) plus a 174-question stress harness; verified the full agent loop E2E on Groq and local Ollama models.</p>

Software Engineer

Pageup Software Services Pvt. Ltd. · Remote, India
Jan 2024 - Aug 2024 · 7 months

<p>Built an LLM-powered internal assistant using OpenAI GPT-4 for document summarization and Q&A over company knowledge bases.</p><p> Implemented a RAG pipeline using OpenAI text-embedding-3-small embeddings and semantic retrieval to ground LLM responses and reduce hallucinations.</p>

PROJECTS

MRKL ReAct Agent Framework

Infosys
Duration : 01-Aug-2026 - 01-Aug-2026

<p>Built a modular, production-oriented ReAct agent framework that supports multi-step reasoning, tool calling, memory, and real-time streaming.</p><p></p><p>• Developed a Thought → Action → Observation → Answer loop with JSON step parsing, loop breakers, and a 20-iteration limit to prevent runaway agent behavior.</p><p>• Created an LLM-agnostic integration layer using LiteLLM for OpenAI, Anthropic, Gemini, DeepSeek, Groq, Ollama, and custom models, with retry and fallback handling.</p><p>• Implemented FastAPI and Server-Sent Events (SSE) for real-time agent-step streaming to a React frontend.</p><p>• Built a pluggable tool registry with web search, sandboxed Python execution, safe calculator, HTTP tools, file operations, MCP support, and ChromaDB-backed semantic memory.</p><p>• Added guardrails, path-traversal protection, allowlists, rate limits, structured logging, and automated tests.</p><p></p>

SnackGPT — Natural Language to SQL Platform

Infosys
Duration : 01-Sep-2024 - 01-Jan-2026

<p>Worked on a GenAI platform that lets business users ask questions in natural language and receive validated SQL-based insights.</p><p></p><p>• Built an NLP-to-SQL workflow covering intent classification, LLM SQL generation, schema-aware validation, query execution, retries, and few-shot prompting.</p><p>• Achieved 95% SQL translation accuracy across 500+ evaluation queries.</p><p>• Migrated LLM and embedding model serving to Databricks Mosaic AI and supported DEV/QA/PROD CI/CD workflows.</p><p>• Implemented LangGraph-based orchestration for modular LLM calls, product mapping, and state management with frontend streaming.</p><p>• Added LLM-as-a-Judge validation that reduced manual E2E testing effort by 90%, reducing validation time from about one day to one hour.</p>

Skills

Python TypeScript JavaScript React LLM Agents RAG Prompt Engineering MCP Vector DBs ChromaDB FastAPI LangGraph LangChain LiteLLM Databricks Mosaic AI GCP AWS Lambda Amplify Docker CI/CD DynamoDB
Report this Profile?