I build reliable AI systems with Python, LangGraph, RAG, tool-using agents, evaluation pipelines, and FastAPI. My work focuses on making LLM applications observable, testable, secure, and useful in production-like environments.
I am a 2026 B.Tech graduate in Electrical Engineering from REC Sonbhadra (AKTU) and an applied AI engineer. I build production-style LLM systems: agent harnesses, RAG pipelines, memory, evaluation, observability, and security boundaries. I care about reproducible results, offline tests, and explicit failure handling.
I am open to Applied AI Engineer, LLM Engineer, AI Engineer, and AI Evaluation roles in India, with preference for Bengaluru, Hyderabad, Mumbai, Delhi-NCR, Pune, and Noida.
Currently building Argus — a small, inspectable agent harness with zero runtime dependencies.
Argus — inspectable agent harness
Connects a language model to a focused set of workspace tools, keeps every action inside a path jail, records durable sessions, and turns safety decisions into readable code.
- Scale: 2,500+ source LOC, 1,547 test LOC, 105 offline tests, 0 runtime dependencies, 6 model providers, 4 interface languages
- Safety: deterministic deny list, path jail, approval flow, secret redaction, optional semantic review
- Proof: Live site · Docs · CI
nano-AgentRL — GRPO for search agents from scratch
Multi-turn GRPO trainer that teaches an LLM to interleave reasoning with Wikipedia search (BM25 over 21M passages), with retrieved-token loss masking and a KL penalty.
- Results: GSM8K 32.4% → 47.9%; search-agent exact match 0% → 13.0% (NQ) and 19.6% (HotpotQA)
- Engineering: 180 CPU tests, RAG baselines, pass^k reliability analysis, silent-RL-bug fixes
- Proof: Repository · CI
MCP Sentinel Lab — runtime security gateway for MCP agents
Manifest scanning, tool pinning, policy enforcement, privacy guards, and CI-backed evaluation for tool-using agents.
- Results (7 bundled demo cases): attack success 100% → 0% against an allow-all baseline; these are deterministic regression checks, not a general security benchmark
- Controls: workspace escape, credential exfiltration, prompt injection, memory poisoning, cross-tool leakage
- Proof: Live demo · Repository
Agent Reliability OS — reliability and evaluation layer
Trace collection, runtime tool-risk policies, secret redaction, and baseline-vs-protected evaluations for tool-using LLM agents.
- Demonstrates: observable decision traces, CI-backed policy checks, regression evaluation, FastAPI and Streamlit surfaces
- Proof: Live demo · Repository
DevMind — terminal-native AI coding agent
LangGraph and Anthropic coding agent with streaming, six security-hardened tools, plugins, and atomic persistent sessions.
- Engineering: 170 offline tests covering security, p50/p95/p99 metrics, rate limiting, retries, and session persistence
- Proof: Repository · CI
EvoCode Scientist — evolutionary coding agent
AlphaEvolve-inspired agent using LLM/local mutations, sandboxed execution, deterministic scoring, and population selection.
- Results: demo bin packing passes all 7 candidates and scores about 0.96 (observed 0.958–0.965; the small drift comes from the measured runtime penalty)
- Engineering: AST validation, subprocess timeouts, SQLite lineage, leaderboards for TSP, bin packing, compression, and knapsack
- Proof: Repository · CI
| Project | What it does |
|---|---|
| ContextOps Agent | Typed long-term memory, context compression, privacy firewall, and evaluations for long-horizon agents |
| OpenAI AutoData | Budget-aware challenger/solver/judge pipeline for hard research QA with fail-closed validation |
| Secure RepoPilot | Issue-to-PR coding agent with baseline verification, command guardrails, and privacy auditing |
| TrustDI Agentic RAG | Agentic RAG for enterprise schema matching with evidence-backed decisions and RAG metrics |
| WishTrip Shillong | Deterministic, explainable trip planner with a live demo |
| Atomity Workload Forecasting | Leakage-safe Azure VM workload forecasting with 24-step predictions and tests |
| SHL Grammar Scoring | Grammar scoring of spoken English from audio (Kaggle SHL assessment) |
- FiniteFree #39 — merged: runnable PyPI quick example and checker regression tests
- paypal/agent-toolkit #111 — open: migrate the Python OpenAI chatbot example to the Responses API (fixes #93)
- paypal/agent-toolkit #112 — open: fix Python example setup docs and environment variable names
- crewAI #8095 — open: opt-in task guardrail for tools
- langchain-ai/deepagents #6020 — closed: rewrite routed glob filters in composite grep
LLM and agents: OpenAI API, Anthropic API, LangGraph, LangChain, tool calling, agent memory, prompt engineering, MCP
RAG and evaluation: vector search, embeddings, hybrid search, corrective RAG, reranking, citation grounding, precision/recall/F1, regression testing, LangSmith, RAGAS
Backend and cloud: Python, FastAPI, Pydantic, SQL, SQLite, Streamlit, Docker, GitHub Actions, AWS SageMaker, pytest
- Build the smallest reliable system that proves the idea.
- Test failure paths, not only happy paths.
- Make cost, state, and model behavior visible.
- Keep claims aligned with reproducible code and results.
- Document limitations clearly instead of overstating benchmark performance.