AI/ML Engineer | GenAI & Agentic AI Developer | Building AI Agents, Local LLM Applications, RAG Systems & Workflow Automation
AI/ML Engineer who has independently shipped 4+ production-grade AI systems, from a self-healing agentic RAG system to an offline desktop AI assistant. Active open-source contributor with 10+ merged pull requests to a LangGraph-based Playwright self-healing engine — including three maintainer-rated "Core Feature" implementations.
Interested in: Generative AI Engineering · Applied AI · Agentic AI · RAG Systems · AI Automation · Forward Deployed Engineering
Fresher | Open to Full-Time Opportunities | Open to Relocation
10+ merged pull requests into a public production codebase, progressing from test coverage to core-feature architecture — promoted from Contributor to Collaborator by the maintainer.
- Built a deterministic MatchScorer (Python) for network-mock replay using weighted similarity scoring across URL, headers, query params, and body, with tie-breaking logic for reproducibility
- Built a persistent SnapshotStore (Pydantic schemas, structlog logging, custom exception hierarchy) for Shadow Runtime state — rated "Core Feature" by the maintainer
- Built the Playwright Mock Injector for the Shadow Testing system, handling sync/async Playwright runtimes via route interception — also rated "Core Feature"
- Wired end-to-end Shadow Runtime replay orchestration (
run_shadow()), launching headless Chromium and connecting Playwright via a remote-debugging websocket — called a critical feature by the maintainer - Added a shadow-verify node gating patches through deterministic offline replay before live test runs, wired into the core self-healing LangGraph workflow — rated "Core Feature"
- Replaced a regex-based JSX diff parser with a tree-sitter AST parser, with graceful fallback, to make selector-diff analysis robust against complex nested JSX
- Added 70+ pytest unit tests across CLI, selector verification, and sandboxed runner modules, including a cross-platform Windows symlink fix
Repo: Lee-Dongwook/E2E-Self-Heal
Built a self-healing agentic RAG system that automatically detects and corrects its own failed answers instead of returning them silently, as measured by an automatic-retry loop triggered whenever an answer fails a faithfulness/relevance check, by architecting a LangGraph state machine wrapping a pgvector retrieval pipeline with a Redis semantic cache.
- Designed a LangGraph agent (retrieve → reason → verify) with a self-healing loop that automatically retries with an altered retrieval strategy when an answer fails evaluation, instead of returning it unflagged
- Built a Redis semantic caching layer in front of a pgvector HNSW retrieval pipeline to cut repeat-query latency and eliminate redundant LLM spend
- Reduced ungrounded responses via a Ragas-style faithfulness and relevance evaluator, with prompt-injection guardrails gating every answer before it reaches the user
Tech Stack
Repo: Avaneesh635/DocOps-Agent
Consolidated 5+ business functions into a single zero-intervention pipeline, as measured by elimination of separate tools for lead intelligence, support triage, and meeting intelligence, by architecting a unified AI platform using n8n Switch nodes with LLM-structured outputs feeding live Google Sheets dashboards.
- Reduced manual reporting time by ~70%, as measured by time-to-dashboard dropping from 2+ hours to under 5 minutes, by automating lead scoring, sentiment classification, and priority detection end-to-end
- Processed 100+ requests per workflow run with zero manual intervention, as measured by end-to-end automation rate, by designing a fully orchestrated pipeline with no human touchpoints
Tech Stack
Delivered a privacy-first desktop assistant running entirely offline, as measured by zero external API calls across chat, memory, and automation features, by architecting an Electron + React app powered by a local Ollama LLM daemon with a secure ContextBridge IPC layer.
- Enabled safe hands-off OS automation, as measured by every system-level action (email drafting, file organizing, batch renaming, app launching) requiring explicit click-to-grant consent, by building sandboxed automation modules gated behind consent security modals
- Built a self-updating personal memory layer, as measured by automatic extraction of user preferences and facts during natural conversation, by implementing agentic memory extraction with a searchable, editable local memory dashboard
Tech Stack
Delivered a high-performance, asynchronous in-memory key-value store, as measured by successful command execution (GET, SET, DEL, EXISTS, PING) and seamless compatibility with the official redis-cli tool, by architecting a multi-threaded TCP server in Rust powered by the Tokio runtime.
- Eliminated lock contention under heavy concurrent read/write traffic, as measured by independent lock acquisition across partitions, by implementing a database sharding engine splitting data store state across 16 memory shards protected by separate RwLocks
- Prevented thread starvation and request blocking during active key eviction, as measured by zero-overhead key cleanups, by designing a background TTL garbage collection worker that utilizes non-blocking try_write locks to prune expired keys
Tech Stack
- Google Cloud Certified Professional Machine Learning Engineer
- AWS Cloud Practitioner Certified
- Oracle Agentic AI Certified
LinkedIn: https://linkedin.com/in/avaneeshjoshi18
Email: avaneeshjoshi18@gmail.com
GitHub: https://github.com/Avaneesh635
Portfolio: https://avaneeshportfolio.netlify.app