MukktinaadhRaghavarapu
I build AI systems that take real action — not chatbots that just talk. RAG, voice, and inference systems running in production.
// shipped work
Featured Projects
Production systems, open-source contributions, and infrastructure — no demos, no toys.
Full-stack voice agent: Whisper STT (local) -> FAISS RAG -> Gemini 2.5 Flash -> Kokoro TTS (local), streamed over WebSockets. VAD preprocessing reduced false positives by 95%. Runs at ~$0.04/month demo scale, ~$12/month at 10K conversations — deliberately architected around free-tier and local models.
AI Scheduling Agent
Live in ProductionProduction healthcare scheduling — 300+ daily assignments, 90%+ accuracy
MyOnsiteHealthcare
LangChain agent + RAG pipeline auto-assigns field workers to patient visits. Handles natural language commands — override assignments, reschedule visits, update patient data. Built end-to-end on FastAPI + PostgreSQL + Redis + AWS. Cut manual dispatcher workload by ~60%.
Houston Research Agent
Open PRDesigned and shipped a new agent for a YC-backed platform — in a single sitting
Reverse-engineered Houston's codebase conventions and built a complete 9th agent — Research — covering competitive intelligence, market research, prospect dossiers, and due diligence. 8 skills, 2 scheduled routines, full schema docs. 955 lines across 12 files, all passing internal checks.
Distributed task queue from scratch — no Celery, no RQ
Redis-backed distributed task queue with priority queues via sorted sets, exponential backoff retry, dead-letter queue, and live monitoring dashboard. Load tested at 893 req/sec with p99 <170ms under 100 concurrent workers, 0% failure rate.
Privacy-preserving federated learning with zk-SNARK verification
Byzantine fault-tolerant federated learning with zk-SNARK gradient verification. gRPC cuts inter-node overhead 10% vs REST. Maintained liveness under 30% simulated node failure.
// giving back
Open Source Impact
I don't just use AI frameworks—I build the infrastructure powering them. Here are my core contributions to the world's most widely used AI libraries.
Fixed a critical performance bug in langchain-core by lazy-loading the massive transformers library inside the get_tokenizer method instead of at the module level. This surgically cut 300-500ms of dead latency off the app startup time for any application importing BaseChatModel.
Added native msgpack serialization and strict-mode deserialization support for pathlib.PurePath variants (PurePosixPath, PureWindowsPath) and range objects in LangGraph's checkpoint system, fixing a crash during state persistence.
OpenAI Python SDK
Open PRFixed streaming connection destruction regression (TCP FIN after [DONE])
Restored the critical response body drain after SSE [DONE] events in both sync and async streaming paths. Without this, h11 cannot parse the HTTP/1.1 chunked terminator, causing httpcore to destroy connections with TCP FIN instead of returning them to the pool — spiking downstream_remote_disconnect errors on enterprise proxies.
Patched a critical proxy routing failure in LiteLLM where the new OpenAI GPT-5.6 reasoning models crashed when executing function calls. Automatically bridges tool execution requests to the /v1/responses endpoint seamlessly.
Implemented a critical machine-level performance optimization in the Azure AI Agent by replacing sequential loops with asynchronous concurrency (asyncio.gather), significantly reducing execution latency when multiple tools are dispatched simultaneously.
Microsoft AutoGen
Open PRFixed core configuration parsing for Azure deployments and extra_body
Patched a critical bug where OpenAIChatCompletionClient threw ValueErrors on custom Azure deployment names. Also fixed silent dropping of `extra_body` configs when loaded via AutoGen Studio JSON by injecting it through the Pydantic schemas and client initialization pipeline.
Architected `aprepare_kickoff` and refactored the `akickoff` execution engine to properly await asynchronous `before_kickoff_callbacks` and `after_kickoff_callbacks`, resolving a critical concurrency state corruption issue in highly-parallelized agent deployments.
Added undated pricing aliases for azure/gpt-audio-mini and azure/gpt-realtime-mini to the model cost map, resolving silent cost tracking failures for enterprise Azure deployments using audio and realtime endpoints.
Fixed a subtle bug where blank assistant messages were silently injected after tool_calls in the message history, corrupting multi-turn conversations and causing downstream model confusion on providers that validate message ordering strictly.
// publications
Research
Research that directly informs production engineering decisions.
Veriblock-FL: A Blockchain-Based Verifiable Federated Learning Framework Using zk-SNARKs
Privacy-preserving FL framework with zero-knowledge proofs for model aggregation integrity verification.
NeuroChainOps — Privacy-Preserving Blockchain MLOps for Federated NAS
Blockchain-based MLOps framework for federated neural architecture search with privacy guarantees.
// technical stack
What I Work With
AI / ML
Speech / NLP
Languages
Backend / Infra
Cloud / DevOps
Tools / MLOps
// how i work
About Me
Recent CS graduate from Parul University (2026), specializing in Big Data Analytics. I operate at the intersection of AI research and systems engineering — building things that are both theoretically sound and production-ready.
I've shipped a real-time voice agent that responds in under 3 seconds, a healthcare scheduling system handling 300+ daily assignments in production, and a complete agent for a YC-backed platform — built unprompted, in a single sitting.
My NeurIPS 2026 research on latency-constrained LLM inference directly informed my production engineering decisions. I don't separate research from building — one feeds the other.
Ships fast
Built VoiceOps in a week. Went from cloning an unfamiliar repo to an open PR on a YC-backed project in a single sitting.
Production over demos
Both flagship projects are live and actually used — not just portfolio pieces.
Research-grounded
NeurIPS submission on latency-constrained inference directly informed VoiceOps's sub-3-second design.
Cost-conscious
VoiceOps deliberately architected around free-tier + local models — $0.04/mo demo, $12/mo at 50x scale.
// let's talk
Get in Touch
Open to AI/ML engineering, agent engineering, and research roles. Immediate availability, zero notice period. Let's build something that matters.