MR
SYS MR-2026 SENIOR GENAI & SYSTEMS ENGINEER
AVAILABLE

Software Engineer · GenAI Engineer

Manas Rai

Mechanical engineer by training, systems engineer by trade — I build GenAI platforms with the same discipline as a tolerance stack-up: nothing ships until it's measured.

See the work Get in touch
00 · SUPERPOWER

Why Mechanical Engineering Makes Me a Better AI Engineer

Tolerance stacks and control loops are how mechanical engineers keep systems from failing quietly. The same discipline is what makes an AI system hold up in production.

01 · ERROR BUDGETING

Tolerance stack-ups → hallucination control

In a machine, micro-tolerances compound across parts until the assembly seizes. In a multi-agent system, small errors compound across model hops the same way. So I budget for them explicitly: schema validation and eval gates sit between stages, and nothing reaches the output that hasn't cleared them.

How the eval harness works →
02 · CLOSED-LOOP CONTROL

Sensor feedback → agentic self-correction

Open-loop systems drift, because nothing tells them they're off. Closed-loop systems correct, because a sensor does. My LangGraph agents work the second way — they check their own output against a schema and a drift detector before a human ever sees it.

See it in RegLens →
03 · STRESS & LOAD TESTING

Safety factors → high-concurrency RAG

You don't open a bridge without stress-testing it well past its rated load. I hold cloud systems to the same standard: the clinical simulation platform was load-tested to 3,000 concurrent sessions at sub-second latency before it carried a single real user.

Read the load story →
01

Featured projects

  • SYS-01 · MULTI-TENANT RAG ON AZURE · unscrew for spec ⟳

    Healthcare RAG Platform

    A multi-tenant Retrieval-Augmented Generation platform for a healthcare product — tenant-isolated vector stores, secure authentication, and strict data partitioning, taken from zero to production and load-tested for 2,000+ concurrent sessions at sub-second latency.

    STACK: Python · FastAPI · Azure AI Search · Azure OpenAI · Azure AD B2C · Document Intelligence · LangChain · Cosmos DB · RAGAS · Multi-tenant
    Architecture
    Python · FastAPI · Azure AI Search · Azure OpenAI · Azure AD B2C · Document Intelligence · LangChain · Cosmos DB · RAGAS · Multi-tenant
    Scale2,000+ concurrent · sub-second
    IsolationPer-tenant vector · blob · B2C claims
    RetrievalHybrid search + semantic reranker
    EvalRAGAS — 4 metrics
    Deep dive →

  • SYS-02 · REAL-TIME VOICE + TEXT LLM · unscrew for spec ⟳

    Clinical Simulation Platform

    An LLM-powered clinical simulation platform for physician training — prompt-engineered patient personas driven by authored case content, over real-time text and voice, plus an automated feedback report that scores the trainee. Load-tested for 3,000 concurrent sessions at sub-second latency.

    STACK: Python · Go · AWS · EKS · OpenAI · Anthropic · Realtime voice · WebSockets · PostgreSQL · LIT
    Architecture
    Python · Go · AWS · EKS · OpenAI · Anthropic · Realtime voice · WebSockets · PostgreSQL · LIT
    Scale3,000 concurrent · sub-second
    ModalityReal-time text + voice (GPT Realtime)
    ModelsOpenAI + Anthropic, config-driven
    AssessmentAutomated per-session feedback report
    Deep dive →

  • SYS-03 · MULTI-AGENT SDLC AUTOMATION · unscrew for spec ⟳

    DevFlow Kit

    Multi-agent SDLC automation that turns Jira tickets into production PRs with zero added infrastructure. Refinement, implementation, and Jira-sync agents decompose complex tickets into parallel subtasks and cut the ticket-to-PR cycle from days to hours.

    STACK: Python · LangGraph · GitHub Actions · Claude Code · Jira
    Architecture
    Python · LangGraph · GitHub Actions · Claude Code · Jira
    InfraZero added — runs on GitHub Actions
    AgentsRefinement · implementation · Jira-sync
    CycleTicket-to-PR from days to hours
    ModelsProvider-agnostic (Anthropic / OpenAI / Google)
    Deep dive → GitHub ↗ Live site ↗

  • SYS-04 · MULTI-AGENT REGULATORY COMPLIANCE AUTOMATION · unscrew for spec ⟳

    RegLens

    Multi-agent regulatory compliance automation. Feed it a regulatory PDF and your control matrix — a compliance research agent extracts every obligation, a gap analyzer checks each against your policies via RAG and scores the risk, and a report generator produces an audit report with a human-in-the-loop approval gate. Includes a drift-detection evaluation harness.

    STACK: Python · LangGraph · Google ADK · A2A / JSON-RPC · Gemini · Claude · pgvector · FastAPI · RAGAS · Next.js
    Architecture
    Python · LangGraph · Google ADK · A2A / JSON-RPC · Gemini · Claude · pgvector · FastAPI · RAGAS · Next.js
    OrchestrationLangGraph supervisor + A2A/JSON-RPC agents
    ModelsGemini (multimodal) + Claude (structured)
    Statepgvector + Postgres-checkpointed, resumable
    EvalRAGAS · DeepEval · behavioral suite · Presidio
    Deep dive → GitHub ↗ Live site ↗
02

Recent writing