Available for Full-Time Roles & Freelance AI Projects

Architecting Autonomous Multi-Agent Systems,
Real-Time Vision & Production RAG.

Agentic AI Engineer & Full-Stack Architect with 5 live production systems deployed across multilingual voice AI, enterprise computer vision, and sub-100ms vector pipelines. Built for founders, CTOs, and teams who need hardened, reliable AI infrastructure — not prototypes.

5 Live
Production AI Systems Shipped
40+ APIs
Async Microservice Endpoints
310+ Commits
Production Code Deliveries
92.5% Uptime
Verified Production SLA
AGENTIC ORCHESTRATION ENGINE · LIVE RUNTIME
CLUSTER ACTIVE · 28ms P95
01 · INPUT
Multilingual Voice Stream
Tamil / Hindi / EN · Whisper STT <180ms
02 · ROUTER
LangGraph State Router
Deterministic Fallback & Tool Dispatch
03 · MEMORY
pgvector & ChromaDB
Sub-100ms Cosine Distance Index
04 · OUTPUT
Verified AI Execution
Sub-2s WhatsApp & Web Response
Core Philosophy

Built for Production Realities, Not Prototype Demos.

Bridging autonomous agent orchestration, low-latency vector databases, and enterprise computer vision with hardened infrastructure.

"AI models are commodity endpoints. Real competitive engineering lies in deterministic tool-calling arbitration, stateful memory persistence, sub-100ms vector retrieval, and auditable data governance."

I build AI software that survives production contact. Whether orchestrating a 4-stage multi-agent loop with LangChain & n8n to automate financial planning across 5 Indian languages in Viya AI, or engineering a 5-stage identity arbitration cascade for multi-camera cross-tracking in SmartDetect, my systems prioritize accuracy, determinism, and sub-second latencies.

Every architecture I ship is grounded in full-stack discipline: typed FastAPI microservices, PostgreSQL + pgvector schemas with cosine distance indexing, Redis ephemeral state caching, and responsive frontend dashboards built on Next.js 14 and React 18.

Current Focus Autonomous Agent State Graphs · Real-Time Vision Re-ID · Distributed RAG
Core Strengths Deterministic Tool Arbitration · Low-Latency Vector Queries · Biometric Privacy Compliance
Availability Open to Full-Time Roles (Remote / Relocation) & Freelance AI Contracts
Engineering Blueprint

Production AI System Architecture

How autonomous agents, vector memory, computer vision, and cloud infrastructure wire together in production systems.

INPUT ORCHESTRATION INTELLIGENCE DATA LAYER DELIVERY Voice / Text Input Tamil · Hindi · English Camera Feeds Multi-camera · RTSP User Queries API · WhatsApp · PWA LangGraph State Graphs · Tool Routing n8n Orchestration Workflow Automation Agent State Mgr Fallback · Retry · Schema LLM Routing Gemini Pro · Claude · GPT-4 · Whisper STT Vision Pipeline YOLOv8 · ByteTrack · ArcFace 512-d · OSNet Re-ID Cascade 5-Stage Arbitration · DPDP/GDPR Governance pgvector Cosine Search · <100ms ChromaDB Vector Memory · RAG Redis Cache Ephemeral State · Pub/Sub FastAPI 40+ Async Endpoints Next.js PWA React 18 · TypeScript WebSocket Real-Time · Low Latency End-to-end latency target: sub-2s voice · sub-50ms vision · sub-100ms vector retrieval
Shipped Products

6 Real Projects. Real Users. Proven Outcomes.

Every system here is live, tested, and built from scratch — with real metrics and intuitive interfaces.

01
Flagship Product ● Live on WhatsApp & Web

An AI that speaks your language — manage your money naturally through voice in Tamil, Hindi, or English.

Viya AI — Your Personal AI Financial Advisor

Viya makes personal finance effortless. Instead of typing into complex budgeting apps, you simply talk or send voice notes on WhatsApp. It categorizes expenses, tracks budgets, and gives instant savings advice in under 2 seconds.

🎙️ 5 Indian Languages: Speaks & understands Tamil, Hindi, English, Telugu, Kannada.
💬 WhatsApp Integrated: Zero app downloads needed — manage money directly via WhatsApp.
⚡ Sub-2s Speed: 121+ production updates shipped with 95%+ intent accuracy.
Next.js PWA FastAPI LangChain Gemini Pro Supabase ChromaDB Whisper STT
🔒 heyviya.vercel.app
● WhatsApp Synced
🎙️ "இந்த மாசம் பட்ஜெட் எவ்ளோ மிச்சம் இருக்கு?"
Tamil Voice Input · Whisper STT (180ms)
LK

You have ₹12,400 remaining in your discretionary budget. You're on track to save 22% of your target this month! 🎯

₹43,600 spent of ₹56,000Healthy Status ✓
📊 Monthly Breakdown 🔔 Set ₹500 Alert ⚡ 1.8s Response
02
Security AI ● Production Engine v2.0

A security camera system that never confuses two people — even across 10 different cameras at the same time.

SmartDetect — Multi-Camera Person Tracking

Detects, tracks, and confirms identities in under 50ms per frame. Standard camera systems mix people up when they cross paths; SmartDetect locks onto face identity first so it never makes mistaken identity guesses.

👁️ Zero Identity Confusion: Strict face-anchored matching eliminates false positive merges.
⚡ Sub-50ms Video Speed: Real-time YOLOv8 + OSNet Re-ID across multiple RTSP streams.
🔒 Biometric Privacy Law (DPDP & GDPR): Auto-purges footage with verifiable deletion receipts.
Python 3.12 FastAPI YOLOv8n InsightFace ArcFace OSNet Re-ID pgvector
📹 SmartDetect v2.0 · Multi-Camera Cluster
● LIVE · 60 FPS
CAM 01 · LOBBY ENTRANCE60 FPS
TARGET #0842 · CONF 99.2%
CAM 02 · 2ND FL CORRIDOR60 FPS
RE-ID MATCH: 98.6% · ID #0842
03
Services Marketplace ● Live

Find verified plumbers, electricians, and technicians near you in under 15 minutes.

Kaizy — India's On-Demand Services App

Connecting customers to nearby skilled workers across 35 categories with instant location dispatch, live tracking, and payments held safely in escrow until the job is done.

⚡
Ramesh K. · Verified Electrician Rating: 4.9★ (124 jobs completed) 📍 Arriving in 14 mins
🛡️ ₹850 Secured in Razorpay Escrow · Cashless
04
Career Intelligence ● Live

Tells students exactly which jobs they qualify for based on a real coding test.

Mentixy — AI Career Guidance Platform

Candidates solve real coding problems in a live browser arena. Claude AI analyzes their performance, builds a 4D skill profile, and maps them to their dream engineering jobs.

Python / FastAPI
95%
Multi-Agent Systems
90%
pgvector Vector RAG
92%
✨

Claude AI Verdict: Top 2% Match for Senior AI Systems Engineer.

05
Voice Commerce AI ● Open Source

A WhatsApp AI assistant for small grocery stores — handles billing & stock by voice in Tamil or English.

KadaiGPT — Smart Kirana Store Assistant

Shop owners speak their daily sales. The AI creates the invoice, updates inventory, and generates a GST bill sent via WhatsApp in under 10 seconds.

KADAI STORE #42 GST INVOICE #882
Toor Dal (1kg) x 2₹240
Sunflower Oil (1L) x 1₹145
Total + GST (5%)₹404.25
🎙️ Generated by Voice in 8.4s · Sent to WhatsApp PDF
06
National Hackathon ● UIDAI 2026 Finalist

Analyzed nearly 5 million government records to find hidden population migration trends.

AadhaarAnalytics 360 — Government Data Intelligence

Led a national hackathon team to process 4.94 million records without crashes, discovering a 28% surge in urban registrations missing from public reports.

Q1
Q2
Q3
Q4 (+28%)
● 4,942,668 Records Streamed · Urban Migration Spike Confirmed
Technical Arsenal

What I Build With & Master

Categorized by engineering capability — from autonomous multi-agent state loops to low-latency computer vision and cloud infrastructure.

AUTONOMOUS WORKFLOWS

AI Agents & Orchestration

Tools to build autonomous systems that reason, call external APIs, maintain memory, and execute multi-step workflows deterministically.

LangChain LangGraph (State Graphs) n8n Orchestration CrewAI Model Context Protocol (MCP) Claude 3.5 & Gemini Pro Tool-Calling Loops Prompt Engineering
VIDEO & OBJECT DETECTION

Computer Vision & Video AI

High-speed visual perception for real-time person detection, face recognition, and cross-camera tracking under 50ms.

YOLOv8 Detection InsightFace (ArcFace 512-d) OSNet Person Re-ID ByteTrack Multi-Camera PyTorch OpenCV 4 RTSP Video Streaming Face Recognition
VECTOR SEARCH & MEMORY

AI Memory & Vector Databases (RAG)

Sub-100ms vector search and knowledge retrieval pipelines that connect LLMs to private databases safely.

pgvector (PostgreSQL 16) ChromaDB Supabase Redis Ephemeral Cache Semantic Caching Cosine Similarity Search Hybrid Dense/Sparse RAG
INFRASTRUCTURE & WEB

Full-Stack & Cloud Infrastructure

Hardened, typed backend microservices, real-time WebSocket streams, and responsive web applications built for high traffic.

FastAPI (Python 3.12) Next.js 14 (App Router) React 18 & TypeScript Docker Containers WebSockets (Real-Time) DPDP & GDPR Compliance CI/CD GitHub Actions
Track Record

What I've Built & When

A timeline of shipped systems, hackathon wins, and live AI deployments — with real outcomes.

Now · 2026 Lead AI Architect
● Production Live

SmartDetect v2.0 — Security Camera AI

Built a camera system that correctly identifies the same person across 10+ cameras without ever mixing up identities. Built with privacy law (DPDP & GDPR) compliance from day one.

🎯 Zero Identity Mix-ups ⚡ <50ms Frame Rate 🔒 7-Day Auto Privacy Purge
YOLOv8 InsightFace OSNet Re-ID ByteTrack FastAPI pgvector
2026 Team Lead & Data Architect
● National Finalist

UIDAI National Hackathon 2026 — Led Team to Finals

Led a 4-person team that processed nearly 5 million official government records in a national competition, discovering hidden 28% urban migration patterns that were missing from published reports.

📊 4,942,668 Records Handled 🏆 National Finalist ⚡ Zero Memory Leaks
Python 3.12 Pandas Plotly FastAPI PostgreSQL
2026 Full-Stack AI Engineer
● Finalist Award

VEL IDEAFORGE 2026 — Mentixy AI Career Platform

Built Mentixy — an AI platform that gives students live coding tests, evaluates submissions with Claude AI, and gives them a personalized career roadmap to their target engineering job.

💻 Judge0 Live Sandbox 🤖 Claude AI Evaluation 🎯 Vector Skill Matching
Claude AI pgvector Judge0 Next.js 14 FastAPI
2025 Creator & Lead Engineer
● 24/7 Live

Viya AI — Multilingual Voice Financial Assistant

Shipped a conversational AI financial assistant that speaks Tamil, Hindi, English, Telugu, and Kannada. Users talk to it like a friend — tracking expenses, setting budgets, and getting savings tips on WhatsApp.

⚡ Sub-2s Voice Latency 🗣️ 5 Indian Languages 🚀 121+ Shipped Commits
LangChain n8n ChromaDB Whisper STT FastAPI Next.js 14
Direct Answers

Frequently Asked Questions

Straightforward specifics on multi-agent architectures, computer vision cascades, SLA guarantees, and engagement models.

Have a specific architecture question?

Ask the AI assistant or reach out directly for system reviews.

Email Direct ↗
Basic wrappers simply feed a prompt into an LLM and stream the text. My systems employ stateful, deterministic multi-agent graphs (LangChain, LangGraph, n8n) equipped with fallback routing, strict schema validation, vector database memory (pgvector, ChromaDB), and sandboxed execution. For example, Viya AI utilizes a 4-stage arbitration pipeline that routes voice intents to specific financial tools with sub-2s latency across 5 Indian languages.
SmartDetect enforces a rigorous 5-Stage Arbitration Cascade:
  • Stage 1 (Primary Anchor): InsightFace ArcFace 512-d embedding matching (cosine similarity ≥ 0.56). Under the Face-Anchor principle, if a face is clearly seen but doesn't match, fallback Re-ID is blocked to eliminate identity confusion.
  • Stage 2 (Torso Metric): Torso HSV color distance ≤ 30 within a 10-minute temporal window.
  • Stage 3 (Deep Re-ID): OSNet 512-d body feature cosine similarity ≥ 0.68 within a 12-hour window.
  • Stage 4 (ID-Switch Guard): ByteTrack tracklet drops cached matches upon 2 consecutive contradictions (< 0.35).
Yes. I am actively interviewing for Full-Time AI Engineering roles (open to remote worldwide or on-site relocation) and available for select freelance AI MVP engineering contracts. Inquiries sent via lokiiii1211@gmail.com or WhatsApp receive responses within 24 hours.
All architectures follow strict data minimization and regulatory compliance (DPDP Act India & GDPR). For instance, SmartDetect implements automated biometric retention purging (7-day unconsented / 365-day consented), verifiable Right-to-be-Forgotten erasure receipts that physically delete snapshot files and purge embeddings from pgvector, backed by an append-only audit trail.
Direct Reach

Let's build something that matters.