Skip to content
View priyansh4320's full-sized avatar
💭
open to work
💭
open to work
22 followers · 18 following

Achievements

Achievement: Pair ExtraordinaireAchievement: Pull Sharkx3Achievement: YOLO

Achievements

Achievement: Pair ExtraordinaireAchievement: Pull Sharkx3Achievement: YOLO

Highlights

  • Pro

Block or report priyansh4320

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don't include any personal information such as legal names or email addresses. Markdown supported. This note will be visible to only you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
priyansh4320/README.md

Hi 🐯, I'm Priyanshu Deshmukh

AI Engineer | India 🇮🇳

GitHub Trophies

Priyanshu Deshmukh

AI Engineer & Maintainer @ AG2 (AutoGen)

Building production-scale multi-agent systems and enterprise AI architectures. Contributing to AG2, the leading open-source multi-agent conversation framework serving 20,000+ developers globally.

Contact: [email protected] | Link | Hyderabad, India

Current Work

AG2 (AutoGen) — AI Engineer & Maintainer

Jul 2025 – Present

Core maintainer and AI systems engineer for AG2’s multi-agent orchestration platform. Owned integration of frontier LLM capabilities (GPT-5 series), production-grade tooling for safe code edits, concurrency-safe agent runtimes, and document-scale RAG capabilities. Drove refactors that improved agent correctness, observability, and throughput while producing tutorials and docs that accelerated adoption.

1) GPT-5 series, apply_patch & shell-tool support

  • GPT-5 & GPT-5.2 model enablement — Added full GPT-5.2 model support and reasoning-effort configuration so AG2 agents can use higher-fidelity reasoning modes (including xhigh) and the Responses API when required. This work updated model enums, pricing/effort mappings, and the responses client to accept new parameters. 🔗 PR [#2250]. ([GitHub][1])

  • GPT-5.1 apply_patch tool support (Responses API) — Implemented support for GPT-5.1’s apply_patch tool (V4A diff/patch format) enabling agents to emit structured, actionable multi-file diffs that can be programmatically applied (safe autonomous refactoring, targeted bugfix patches, CI-friendly edits). Also delivered example tutorial/notebook to demonstrate the pattern. 🔗 PR [#2213]. ([GitHub][2]) Tech & skills: tool-calling design, structured output parsing, patch application safety, tests for diff format handling.

  • Shell tool / tool concurrency support — Implemented shell/tool execution improvements so agent tool calls (including shell) are concurrency-safe and return machine-parseable structured outputs for chaining. This reduced deadlocks/race conditions in pipeline runs and made tool outputs deterministic for planners. 🔗 PR reference: shell/tool work collection (see PR list). ([GitHub][3]) Tech & skills: concurrency safety, tool contract design, deterministic output formats.

2) Provider & infra resilience (Bedrock, Gemini, Ollama)

  • AWS Bedrock resilience & structured outputs — Added exponential backoff and retry strategies at the Bedrock provider layer and Bedrock structured-output handling so transient failures don’t cascade into agent crashes and downstream parsers consistently receive JSON/structured responses. 🔗 PR [#2292] (mentioned across PRs/listings). Why it matters: Production readiness for enterprise clouds.

  • Gemini: ThinkingConfig / reasoning controls — Extended ThinkingConfig support to Gemini models so AG2 can control depth/latency of Gemini reasoning and capture "thought signatures" for better traceability in multi-step planning. 🔗 PR [#2254]. ([GitHub][4]) Tech & skills: provider abstraction, reasoning controls, cross-provider parity.

  • Ollama validation fixes — Fixed LLMConfig validation paths that previously caused runtime errors when native_tool_calls were enabled, preventing misconfiguration crashes. 🔗 PR [#1951]. Skills: config validation, cross-provider stability.

3) DocAgent — architecture, dynamic RAG, and high-throughput document workflows

  • DocAgent architecture optimization (threading, concurrent ingest, supervisor) — Simplified and optimized DocAgent internals: added a ThreadPoolExecutor for tool execution, a pseudo-supervisor for task orchestration, concurrent document ingestion, citation support, and flexible inner-agent prompting to greatly improve ingestion throughput and query latency. 🔗 PR [#2097]. ([GitHub][5]) Technical highlights: concurrent ingestion pipelines, thread pools for I/O-bound tasks, scalable supervisor pattern, test scenarios for long-document ingestion.

  • Dynamic RAG (Vector + Graph / Neo4j integration + aggregator) — Extended DocAgent beyond vector RAG by adding Graph RAG support (Neo4j) and a dynamic rag_config to select/aggregate results from vector and graph retrievals. This enables multi-hop, relationship-aware queries and a result aggregator for combined RAG outputs. 🔗 PR [#2105]. ([GitHub][6]) Why it matters: Demonstrates system design for hybrid retrieval (vector + graph) and shows you enabled sub-second, multi-hop retrieval patterns valuable in enterprise search.

4) Agent runtime refactors, networking & message model improvements

  • ConversableAgent refactor (message flow & API consistency) — Refactored ConversableAgent internals to enforce consistent message list APIs, clarify state transitions, and improve extensibility for multi-turn scenarios. 🔗 PR [#2086]. ([GitHub][7]) Skills: API design, state-machine clarity, multi-turn correctness.

  • Agent networking / list[messages] API enforcement — Standardized list[messages] semantics across agent messaging, improving inter-agent communication and easing integration with group chat / networking scenarios. 🔗 PR [#2081]. ([GitHub][8])

  • ParallelAgentRunner / concurrency orchestration — Built ParallelAgentRunner to run agents in parallel safely (thread/process coordination, result aggregation, cancellation semantics), improving throughput for batched workflows. 🔗 PR [#2143]. ([GitHub][9]) Tech & skills: concurrency patterns, cancellation semantics, thread-safe result aggregation.

5) Reliability, fixes & dependency hygiene

  • Memory / long-context stability — Replaced brittle long-context paths and added safer concurrency patterns to avoid context overwrites in document agents (used in the DocAgent refactors above). (See DocAgent PRs #2097 / #2105.) ([GitHub][5])

  • Dependency & resolver hardening — Switched pinned dependencies to ranges where appropriate and fixed async validation issues across providers to reduce release friction and unexpected breakages. (Referenced across PR list.)

6) Docs, tutorials & community enablement

  • Apply-patch tutorial & notebook — Published a tutorial demonstrating GPT-5.1 apply_patch flows in AG2 to help integrators build safe code-editing agents (linked from release notes). 🔗 PR [#2213] (release notes & tutorial). ([GitHub][2])

  • DocAgent & Dynamic RAG notebooks — Added sample notebooks showing concurrent ingestion, dynamic RAG configuration, and example runs — useful in interviews and technical screen walkthroughs. 🔗 PRs [#2097], [#2105]. ([GitHub][5])

  • Developer DX (devcontainer, logging, hooks) — Added a Python 3.14 devcontainer to standardize contributor environments and documented lifecycle hooks (process_message_before_send) so downstream teams can plug custom logic into message pipelines.

Previous Experience

Alvyl - Software Engineer (AI) (Aug 2024 - Jul 2025)

Architected FusionSecurity.ai, an enterprise-grade multi-agent security intelligence platform for Fortune 500 clients:

  • Designed LangGraph-powered architecture with specialized role-based agents processing daily security events with high uptime
  • Implemented hierarchical Supervisor Agent orchestration distributing tasks across 15+ specialized sub-agents in containerized AWS/Kubernetes environments
  • Engineered multi-modal RAG system combining vector search, graph relationships, and structured data analysis
  • Built self-evolving threat intelligence database with dynamic schema expansion
  • Deployed containerized microservices supporting mission-critical security monitoring
  • Integrated Semgrep SAST scanning in CI/CD pipelines
  • Developed multi-cloud metadata analysis agents for AWS/Azure/GCP compliance monitoring

Freelance AI/ML Solutions Architect (2022 - 2024)

Delivered AI transformation solutions for BlackCoffer, Digipplus, PrecilyAI, and UniAcco. Built scalable ML pipelines with end-to-end MLOps integration.

Technical Expertise

Core Competencies

  • AI Pattern Research
  • System Design, Low-level AI, and networking
  • Multi-Threading, Optimisation
  • OpenAI SDK, AWS Converse, anthropic SDK, Gemini API, Ollama, Huggingface, i.e, (Client-Side Development).
  • Multi-Agent Orchestration (AG2/AutoGen, LangGraph)
  • Advanced RAG Systems (Vector, Graph, Structured)
  • LLM Integration & Prompt Engineering
  • Enterprise AI Security & Automation

Infrastructure & Tools

  • Cloud Platforms: AWS, Azure, GCP, Kubernetes
  • Databases: Neo4j, Vector Databases
  • Security: Semgrep, SAST Integration
  • CI/CD & MLOps Pipelines

Architecture

  • Cloud-Native Microservices
  • Production AI Systems
  • Performance Engineering
  • Security Intelligence Platforms

Publications & Impact

  • Published IEEE research on AI systems
  • Contributing to framework serving 20,000+ developers
  • Mentored 50+ community developers on multi-agent architecture
  • Open-source contributions to AG2 core development

Building intelligent systems that bridge cutting-edge AI research with production deployment.

📫 Reach Me At:

🔗 Connect with Me:

LinkedIn

Pinned Loading

  1. ag2ai/ag2 ag2ai/ag2 Public

    AG2 (formerly AutoGen): The Open-Source AgentOS. Join us at: https://discord.gg/sNGSwQME3x

    Python 4k 518

  2. FalkorDB/GraphRAG-SDK FalkorDB/GraphRAG-SDK Public

    Build fast and accurate GenAI apps with GraphRAG SDK at scale.

    Python 541 69

  3. build-with-ag2 build-with-ag2 Public

    Forked from ag2ai/build-with-ag2

    Sample code and application showcases to get you going with AG2 (formally AutoGen)

    Jupyter Notebook 1

  4. Abstractive-Text-Summarization-Enhancing-Sequence-to-Sequence-Models-Using-Word-Sense-Disambiguatio Abstractive-Text-Summarization-Enhancing-Sequence-to-Sequence-Models-Using-Word-Sense-Disambiguatio Public

    This repository contains code and resources for abstractive text summarization (TS) using a novel framework that leverages knowledge-based word sense disambiguation (WSD) and semantic content gener…

    Jupyter Notebook 14

  5. Ice-Sejj Ice-Sejj Public

    Introducing Ice-Sejj: Your Ultimate NLP and Machine Learning Solution

    Python 2 1

  6. sEmantic_sentence_similarity_model_api_with_transformers_and_tensorflow sEmantic_sentence_similarity_model_api_with_transformers_and_tensorflow Public

    an api for the sns model created (semantic sentence similarity model using tansformers ,tenosrflow and fastapi ) : hosted on aws

    Jupyter Notebook 1