PortfolioVol. 012026Bengaluru, India
Open - AI Engineer roles

SP Devanandhan.

Production AI agents and voice automations - not fragile demos.

I build production AI agents and voice automations that replace 40+ hours of manual work per week - not fragile demos that break the first week. Last build: 847 outbound dials, 47 booked appointments, and 3 funded loans in week one for a US mortgage brokerage. 93% Job Success across 29 contracts on Upwork.

Open to Agentic AI Engineer, Forward Deployed Engineer, and Founding AI Engineer roles — remote, on-site in Bengaluru, or relocation to EU / US-aligned hours.

Download résuméEmail meUpwork — 93% JSS
LinkedIn ↗GitHub ↗+91 96330 20011
RoleClaude Code & Agentic AI Engineer
FocusClaude Code · LangGraph · Voice · RAG
StackPython · TS · FastAPI · n8n · MCP
BasedBengaluru / UTC+5:30
Upwork93% · 29 jobs · 479 hrs
PubIEEE · ICCCNT 2025
The ledger - what's shipped

  • 93%

    Job Success Score on Upwork

    29 contracts · 479 billable hours

  • 847 → 3

    Outbound dials → funded loans

    Voice agent, week one, US mortgage brokerage

  • $10K+

    Earned on Upwork

    Direct & studio contracts on top - since 2024

  • IEEE

    Published - ICCCNT 2025, IIT Indore

    Healthcare AI compliance & evaluation

§I. Work

Three things live under my name.

The studio · the outbound product · the vertical

01
Live · 2024 →

Studio - LLM, RAG & agent systems for B2B

Hollerlabs.

40+ production AI deployments across healthcare, insurance, SaaS, real-estate, and cybersecurity - earning $10k+ on Upwork alone plus direct contracts since 2024.

Stack

GPT · Claude · LangGraph · FAISS · FastAPI · Docker · n8n

The studio I run as co-founder. We ship multi-agent orchestration, RAG pipelines, conversational AI, OCR/ASR workflows, and FastAPI microservices - wired with structured logging, retries, drift detection, and audit trails so they survive enterprise QA.

Selected production engagements include StellaInsurance (OCR + NLU for policy & claims), Pug.ai (LLM lead scoring + ICP verification at scale), Mastracorp (OCR for property/leasing docs powering a subscription product), and a multi-client voice agent that ran 800+ outbound calls in its first week.

Visit hollerlabs.tech ↗GitHub ↗
SPDPlate 01
Live · 2024 →
Hollerlabs.

Visit

hollerlabs.tech

Fig. 01 - hollerlabs.techStudio
02
Shipping 2026

Product - AI desk for B2B outbound

ChaseDesk.

Multi-agent outbound platform that researches accounts, drafts personalized sequences, and works leads through reply, qualification, and CRM handoff - without a human re-typing anything.

Stack

Python · LangGraph · Claude · Postgres · Next.js · Vercel

Three agents share a single workspace - Research (account + persona + recent signal), Draft (sequence + variant generation grounded in ICP), and Reply (intent classification, qualification, calendar handoff). Every action is logged with citations so a human can audit any send before it goes out.

Built to replace the stack of Apollo + Clay + Smartlead + a VA - one product, one bill, one set of evals.

Visit chasedesk.app ↗
SPDPlate 02
Shipping 2026
ChaseDesk.

Visit

chasedesk.app

Fig. 02 - chasedesk.appProduct
03
Live · v1

Product - Refinance pipeline OS

RefinanceFlows.

Vertical AI for US mortgage brokerages - multi-source lead intake, enrichment, loan-size routing, and templated follow-up that cut a broker's daily routing work from ~4 hours to ~15 minutes.

Stack

n8n · OpenAI API · HubSpot · Mailgun · Google Sheets · Render

Lead intake from forms, partner spreadsheets, and inbound email lands in one normalized queue. An LLM step enriches each row with loan-size bracket, intent score, and a one-line summary the broker reads before opening anything. Routing rules push qualified leads into HubSpot with the right owner and kick off a templated Mailgun follow-up.

A nightly audit catches anything the model was unsure about and surfaces it for human review before morning - the bar is broker-of-record trust, not chatbot demo.

Visit refinanceflows.com ↗Loom walkthrough →
SPDPlate 03
Live · v1
RefinanceFlows.

Visit

refinanceflows.com

Fig. 03 - refinanceflows.comProduct
§II. Capabilities

What I'll bring on day one.

Six surfaces - already shipped in production, not learned from a tutorial.

S.01

Agent systems.

Multi-agent orchestration with tool use, memory, structured outputs, function calling, and graceful fallback. Claude Code skills, MCP servers, LangGraph plan-then-verify loops, custom agents wired into real product flows.

  • Claude Code skills
  • MCP servers
  • LangGraph
  • Tool design
  • Memory + guardrails
S.02

RAG & retrieval.

Domain-grounded retrieval that earns the right to answer - hybrid search, citation-anchored generation, retrieval evals, refusal under low confidence. FAISS, Chroma, LlamaIndex over proprietary client corpora.

  • FAISS
  • Chroma
  • LlamaIndex
  • Embeddings
  • Semantic search
  • Citation grounding
S.03

LLM ops & evaluation.

Production rigor - code, model, and human graders; precision / recall / F1; A/B evaluation; drift detection; structured logs; cost & latency budgets; prompt caching; regression suites that gate every deploy.

  • Evals (code · model · human)
  • Drift detection
  • Prompt caching
  • Cost / latency tuning
  • Audit trails
S.04

Voice & conversational AI.

ASR → NLU → RAG → fallback pipelines with conversational memory, dynamic prompting, and intent classification. One deployment ran 800+ outbound calls in week one with measurable conversion.

  • ElevenLabs
  • Realtime API
  • Whisper
  • Conversational memory
  • Intent classification
S.05

Document AI.

OCR pipelines for insurance, leasing, and claims documents with field validation, anomaly detection, and compliance-aware audit trails. Powers subscription products at production scale.

  • OCR
  • Layout-aware parsing
  • Field validation
  • Anomaly detection
  • Compliance audit trails
S.06

Production stack.

FastAPI microservices with Dockerized deployment, retry logic, structured logging, and monitoring. Next.js front-ends. Postgres / Supabase. n8n where it earns its keep - orchestration glue, not core IP.

  • Python
  • FastAPI
  • TypeScript
  • Next.js
  • Docker
  • Postgres
  • n8n

A sample skill I'd ship

I write Claude Code skills, MCP servers, and LangGraph agents the same way: small contract at the top, eval at the bottom, the whole thing readable in 30 seconds.

---name: rag-evaluatordescription: Run citation-grounded RAG with eval gates.model: claude-sonnet-4-6tools: [Read, Grep, Bash, Inference]---## When to useQuestion over proprietary corpus where the answermust cite a source and refuse under low confidence.## Workflow1. retrieve(k=8) → rerank → top-32. answer → require ≥1 citation per claim3. eval: faithfulness, context-precision, refusal4. log to MEMORY/EVALS/{slug}.jsonl
§III. Selected work

Named clients. Real deployments.

A slice of the 40+ production systems shipped through Hollerlabs.

  • C.01

    StellaInsurance

    Insurance

    OCR → NLU → rule-based triage for policies and claims

    Field-level validation, anomaly detection, and compliance-aware audit trails on every inference.

  • C.02

    Pug.ai

    SaaS · Sales intelligence

    LLM lead scoring & ICP verification at scale

    AI-driven scoring, segmentation, and enrichment with rule-based logic, retries, and orchestration monitoring.

  • C.03

    Mastracorp

    Real-estate · PropTech

    OCR extraction powering a subscription product

    Document parsing for property and leasing docs with validation rules and reusable FastAPI microservices.

  • C.04

    Voice agent (multi-client)

    Healthcare · Services

    800+ outbound calls in week one, measurable conversion

    ASR + NLU + LLM with FAISS retrieval, dynamic prompting, conversational memory, and graceful fallback.

  • C.05

    IEEE Healthcare AI

    Research · Healthcare

    Three-model architecture for clinical compliance

    Published at ICCCNT 2025 (IEEE / IIT Indore). Probability-based evaluation and compliance-aware inference with audit-ready outputs.

§IV. Reviews

What clients have actually written.

Five recent 5-star reviews from 29 completed contracts. Verifiable on Upwork.

  • T.01★★★★★
    Dec 2025 - Apr 2026
    “Devan and his team are amazing. They work fast and have bent over backwards to help us get it right. They are very knowledgeable as well.”

    Build and deliver a fully functional AI agent with complete workflows / integrations

    5.0 · $900 · fixed

  • T.02★★★★★
    Jul 2025 - Feb 2026
    “Built a Slack-based project called Notty on top of n8n. Overall, very solid experience. Automations run reliably, logic is flexible, and integration with Slack feels natural.”

    Enhancement of Slack Bot Integration in n8n

    5.0 · $2,275 · 152 hrs

  • T.03★★★★★
    Feb 2026 - Mar 2026
    “Great experience working with Devan and the team at Hollerlabs. They helped migrate our workflows from Make to n8n and set up the automation for our AI voice agent system.”

    AI Voice Agent Fixing

    5.0 · $200 · fixed

  • T.04★★★★★
    Nov 2025 - Feb 2026
    “Devan had been really helpful and communicative throughout the project, he answered the questions and explained properly especially on those that we do not have expertise on.”

    AI Automation Specialist - Zapier + AI + OneUp + SEO stack

    5.0 · $250 · fixed

  • T.05★★★★★
    Mar 2026 - Apr 2026
    “Amazing as always!”

    Repurpose PocketBarber Slideshow Workflow into Ready-to-Fly Daily Content Engine

    5.0 · $86 · fixed

SourceUpwork profile - Devanandhan S. ↗93% JSS · 29 jobs · 479 hrs
§V. About

The short version.

I'm Devanandhan - most people call me Devan. I'm 23, based in Bengaluru, and I build production AI systems for B2B teams. Since 2024 I've shipped 40+ deployments across healthcare, insurance, SaaS, real-estate, and cybersecurity - including $10k+ earned on Upwork plus direct contracts on top.

I co-founded Hollerlabs as my studio surface for client work and turned three of those builds into products of my own - ChaseDesk for outbound, RefinanceFlows for US mortgage brokerages, and a few more in the oven. My day job is LangGraph, Claude Code skills, MCP servers, FastAPI microservices, RAG pipelines, and the evals that keep them honest.

On the research side, I co-authored a paper on healthcare AI compliance - published at ICCCNT 2025 (IEEE, IIT Indore) - on a three-model architecture for clinical compliance and cost-aware inference.

Right now I'm looking to join a SaaS or YC-funded team as an Agentic AI Engineer, Forward Deployed Engineer, or Founding AI Engineer - somewhere I can ship flagship AI features instead of demos. n8n is a tool I keep in the belt for orchestration glue; my core is LLM + agent + RAG engineering at production scale.