Skip to main content
MZ
MD Ziauddin
Forward Deployed AI & Systems Engineer • India

Turning complex AI & distributed architectures into low-latency production systems.

I'm MD Ziauddin - a Senior Software Engineer and Forward Deployed AI Engineer with 5+ years of experience. My work sits at the intersection of deep backend engineering (TypeScript, Fastify, Python), production Voice AI & hybrid RAG orchestration, and high-stakes enterprise delivery across OmniDimension (1M+ calls), Growwth Partners, and VMware.

Explore System ArchitecturesVoice AI Latency Lab
ARCHITECTURAL OPERATING PROFILE5+ Years Exp
Primary Focus:Voice AI, Hybrid RAG & Distributed Backends
Core Runtime:TypeScript • Fastify • Python • React
AI & Data Stack:LangGraph • Qdrant • Redis • PostgreSQL
Production Scale:1M+ Voice Calls • -90% Sync Latency
How I Work: Tablet + Excalidraw System Sketch -> Benchmark & Telemetry Spec -> Production TypeScript / Python / Fastify
LIVE SYSTEM TRACE • OMNIDIMENSION SCALE (1M+ CALLS)

Enterprise Voice AI + RAG Conversational Loop (Click any stage to inspect telemetry)

Caller Perceived Response Wait Time (Turn-Taking Dead Air):380ms • Natural Human Cadence (-79% Latency)
STAGE 03: DYNAMIC CONTEXTUAL FILLERSub-120ms Intent Classifier + Pre-warmed Audio
Click any stage card above to inspect optimization details
⚡ Agent (380ms Bridge): "I can certainly help with cardiology - let me check tomorrow afternoon’s slots right now..."

Injects context-aware, natural conversational bridges ("Let me check that schedule for you right now...") while heavy LLM & CRM lookups execute asynchronously.

Zia's FDE Optimization: Zia’s signature GenAI optimization: reduces perceived conversational dead-air from 1,850ms down to ~380ms during live enterprise calls.

VERIFIED PRODUCTION TELEMETRY & ENTERPRISE IMPACT
OMNIDIMENSION • PRODUCTION SCALE
1,000,000+
99.4% Uptime
Enterprise Outbound Voice AI Calls

Architected and scaled autonomous healthcare & e-commerce Voice AI workflows from design to high-volume production.

GROWWTH PARTNERS • FINTECH CORE
-90%
2,400ms -> 240ms
Platform Application Latency Reduction

Re-architected static request-response workflows into a low-latency bidirectional WebSocket AI assistant architecture.

SWE - GENAI -> FORWARD DEPLOYED ENG
3 Months
End-to-End Ownership
Fast-Track Promotion to FDE

Promoted within 90 days to spearhead enterprise solution architecture, customer CRM integrations, and live launch triage.

SYSTEM STANDARDIZATION & MENTORSHIP
+40%
4 Engineers Mentored
Engineering Release Velocity

Created a standardized React Google Apps Script (GAS) framework and modular monorepo patterns across product teams.

01 • PRODUCTION VOICE & GENAI

Sub-400ms Voice AI & Empirically Benchmarked RAG

I build AI systems that survive real enterprise traffic, from Dynamic Contextual Fillers masking 1.8s LLM latency across 1,000,000+ calls to MLflow-benchmarked Qdrant + Bedrock medical RAG pipelines (0.9471 NDCG@10).

02 • DISTRIBUTED SYSTEMS & FULL-STACK

90% Latency Cuts & High-Throughput Primitives

Whether replacing polling bottlenecks with multiplexed WebSockets + Redis Pub/Sub, securing VMware APIs with OAuth 2.0, or building standardized React GAS frameworks, I focus on tail latency and developer velocity (+40%).

03 • FORWARD-DEPLOYED LEADERSHIP

Whiteboard Architecture to Live Production Triage

Promoted to FDE in 3 months at OmniDimension. I work directly with enterprise customers to translate business SOPs into Excalidraw blueprints, integrate Odoo/LeadSquared CRMs, and lead live launch incident resolution.

SIGNATURE WORKFLOW: EXCALIDRAW TABLET SKETCH → PRODUCTION TELEMETRY

Flagship System Architectures & Proofs

Every system starts on my drawing tablet in Excalidraw before a single line of TypeScript or Python is written. Toggle between ⚡ Live Production Metrics and ✏️ Hand-Drawn Architecture Diagrams, or click Expand Diagram on any card to inspect the full-screen blueprint.

Switch All Cards:
ENTERPRISE VOICE AI & QA INFRASTRUCTURE
OmniDimension•Nov 2025 - Oct 2026

Scaling Autonomous Voice AI to 1M+ Calls with Sub-400ms Perceived Latency

The Engineering Problem: Enterprise healthcare and e-commerce voice campaigns suffered from 1.8s+ awkward conversational silences during LLM/CRM lookups, while manual QA teams could only audit <1% of call recordings against strict clinical and operational SOPs.

Outbound Volume
1,000,000+ Calls
Perceived Latency
~380ms TTFA
First Month Conversions
50+ Consultations
QA Audit Coverage
100% Automated SOP
Automated LLM Call QA & Compliance Pipeline SampleAudit Score: 97.6% SOP
HIPAA / Patient Identity VerificationVerified DOB & consent within first 18s
98/100 • PASS
Turn-Taking Latency (<500ms threshold)Contextual filler masked 640ms Odoo slot lookup
95/100 • PASS
Appointment Qualification & CRM SyncLeadSquared webhook confirmed slot #LS-8841
100/100 • PASS
TypeScriptPythonVoice AI PipelinesSTT/TTS BenchmarkingWebSocketsOdoo & LeadSquared CRMTwilio / Ozonetel / ExotelEvent-Driven Alerts
Read 1M+ Call Architecture Post ->
HEALTHCARE AI SIDE PROJECT • MLFLOW RAG BENCHMARKING
MedSearch AI & K-Search Engine•Empirical MLflow Suite

Empirical Medical RAG Evaluation Pipeline: Qdrant, Bedrock Titan v2 & LangGraph

The Engineering Problem: Standard AI dogma assumes naive 50/50 Hybrid BM25 + Dense search always beats dense retrieval. Across 180 MLflow evaluations on PubMedQA (100-doc & 300-doc scale sweeps), naive BM25 matched generic clinical tokens ("patient", "cell"), degrading MRR by ~40% until rescued by an Agentic LangGraph reflection loop.

Dense NDCG@10 (100 / 300 Docs)
1.1417 / 0.9471
Agentic Hybrid Rescue (Exp 2)
+15.3% NDCG / +13.1% MRR
Peak Answer Relevance
0.8094 (Dense Agentic)
Scale Sweep Latency (300 Docs)
~240ms - 252ms / query
MLflow Benchmark (PubMedQA: 100 & 300 Doc Sweeps)Verified MLflow Commit

Exp 1 (100 Docs, ~500 Chunks, 15 Queries): Dense NDCG@10 = 1.1417 | MRR@10 = 0.4989 | Answer Relevance = 0.8094 (~500ms/query)

Run 1: Dense + Linear (Qdrant + Titan v2)NDCG@10: 0.9471 (Exp 1: 1.1417) | MRR@10: 0.4450 (Exp 1: 0.4989)
Answer Relevance: 0.7955Scale Sweep Latency: ~251ms/query
Run 2: Dense + Agentic (LangGraph Reflection)NDCG@10: 0.9471 (Exp 1: 1.1417) | MRR@10: 0.4450 (Exp 1: 0.4989)
Answer Relevance: 0.8094 (Exp 1 Peak)Scale Sweep Latency: ~252ms/query
Run 4: Hybrid + Agentic (LangGraph Rescue)NDCG@10: 0.6441 (+15.3% vs Linear) | MRR@10: 0.2922 (+13.1% vs Linear)
Answer Relevance: 0.7929Scale Sweep Latency: ~246ms/query
Run 3: Hybrid + Linear (Naive 50/50 BM25 + RRF)NDCG@10: 0.5587 | MRR@10: 0.2583 (Lexical Noise)
Answer Relevance: 0.7853Scale Sweep Latency: ~240ms/query
Python 3.11FastAPIQdrant (v1.12.4)AWS Bedrock (Titan v2 + Claude Sonnet 4.5)LangGraph (Agentic RAG)BM25 + RRF (k=60)MLflow (v2.17.2)ranx IR Evals
Read Full MLflow Benchmark Report ->
REAL-TIME FINTECH & REACT GAS FRAMEWORK
Growwth Partners & RyzUp Sheets•Jul 2024 - Nov 2025

Cutting Fintech Latency by 90% & Embedding AI Assistants Inside Google Sheets

The Engineering Problem: Financial analysts and CFOs were stuck between a high-latency web platform (2.4s+ polling delays) and fragmented Google Sheets workflows that lacked real-time AI modeling or standardized frontend engineering.

Application Latency
-90% (2.4s -> 240ms)
Release Velocity
+40% Faster Cycles
Team Leadership
4 Engineers Mentored
Core Ownership
AI Assistant & Referral
RyzUp Sheets • Standardized React GAS Add-On RuntimeWebSocket Live Sync
RangeFormula / Assistant PromptAI Output (240ms)
A1:D1=RYZUP_AI("Analyze Q3 SaaS burn vs ARR")Net burn improved 18.4%; CAC payback 11.2 mos
A2:D2=RYZUP_FORECAST(B2:M2, "runway_months")22.4 Months Runway (+3.2 mos optimized)
A3:D3=RYZUP_AUDIT("Flag anomalies > 15% MoM")2 cloud infra spikes flagged ($4,820 variance)
React.jsTypeScriptWebSocketsGoogle Apps Script (React GAS)Node.jsPostgreSQLRedisTurborepo
Visit RyzUp AI Platform ->
REAL-TIME VOICE & HIGH-THROUGHPUT BACKEND PRIMITIVES
Articulate AI & Core Micro-Systems•Open Architecture Project

Articulate AI Mock Interviewer & Sub-Millisecond Fastify + Redis + Sqids Engine

The Engineering Problem: Building responsive consumer voice coaching and high-throughput link/telemetry primitives requires zero-bloat backend engineering where every millisecond of serialization and cache lookup counts.

Cache Hit Latency
0.42ms Redis Lookup
ID Obfuscation
Collision-Free Sqids
API Framework
Fastify JSON Schema
Persistence
Supabase PostgreSQL
Live Micro-Demo: Fastify + Sqids + Redis Shortlink EngineTry generating a link
ziadev.in/s/fde-hire-9kM2
0.38ms (Redis HIT)
ziadev.in/s/voice-ai-7pX4
0.41ms (Redis HIT)
FastifyRedisSqidsSupabasePostgreSQLNext.jsWebRTC / Voice AISystem Design
Read Articulate AI Deep-Dive ->
PUBLISHED VENDOR BENCHMARKS VS. FDE PIPELINE OPTIMIZATION

Voice AI STT / LLM / TTS Latency Calculator Lab

Grounded in published third-party benchmarks (Pipecat STT TTFS, Artificial Analysis LLM TTFT, Coval/Gradium TTS TTFA). Inspect the raw unoptimized vendor latency first, then toggle our Tuned Streaming Stack and Dynamic Contextual Filler to see how we achieved sub-400ms perceived turn-taking.

Compare Scenarios:
Vendor Standard:160ms
FDE Optimized:95ms (-65ms)

Standard multi-hop WSS relay + default jitter buffer (160ms)

Pinned regional edge WSS termination + 20ms mulaw frame pacing (-65ms)

Pipecat Median TTFS:247ms
FDE Tuned VAD:165ms (-82ms)

Published Pipecat median TTFS (Time-to-Final-Segment): 247ms (TTFT ~91-131ms)

Persistent WSS + custom 220ms VAD endpointing & interim token trigger (-82ms)

Vendor TTFT + RAG:460ms
FDE Cached + HNSW:295ms (-165ms)

Published median TTFT ~220ms (Artificial Analysis) + 240ms uncached RAG = 460ms

Prompt prefix caching (145ms TTFT) + HNSW RAM-pinned Qdrant (150ms) (-165ms)

Published TTFA:190ms
FDE WSS Chunked:95ms (-95ms)

Published independent median TTFA: ~190ms (Server-only TTFB ~80-90ms)

Persistent WSS + raw 8kHz PCM/mulaw byte-stream chunking (-95ms)

Latency Progression:Vendor Standard: 1057ms→FDE Tuned Pipeline: 650ms→Perceived TTFA: 380ms
OPTIMAL HUMAN CADENCE (<500ms)
Tel 95ms
STT 165ms
LLM+RAG 295ms (Masked)
TTS 95ms
Zia's FDE Benchmark Note:Every provider metric above starts with published industry median benchmarks (Pipecat STT TTFS, Artificial Analysis LLM TTFT, and Coval/Gradium TTS TTFA). By combining WebSocket connection pooling, custom 220ms VAD endpointing, RAM-pinned Qdrant HNSW, and our <120ms Dynamic Contextual Filler, we cut perceived caller wait time from 1057ms down to 380ms.
ENTERPRISE TRACK RECORD • INSPIRED BY ANTV G6 DEPENDENCY GRAPHS

Career Evolution: Frontend → Full-Stack → Forward Deployed AI

Over 5+ years, my work has evolved from building enterprise React UIs and AntV G6 infrastructure graphs (VMware, Ariksa) to leading fintech platform modernization and owning 1M+ call Voice AI deployments.

Feb 2026 - Oct 2026 • India

Forward Deployed Engineer (FDE) @ OmniDimension

PROMOTED IN 3 MONTHS

Spearheaded enterprise Voice AI deployments across 1,000,000+ outbound healthcare & e-commerce calls.

Verified Production Impact & Ownership:

Promoted from Software Engineer within 90 days to own end-to-end enterprise solution architecture from Excalidraw blueprint through production launch.
Profiled and benchmarked multi-vendor LLM, STT, and TTS matrices to hit sub-400ms perceived conversational turn-taking.
Engineered bidirectional integrations with Odoo and LeadSquared CRMs plus hospital scheduling engines, driving 50+ qualified patient consultations in the first month.
Acted as primary engineering liaison for enterprise rollouts, leading live production triage, root-cause validation, and cross-team incident resolution.
Connected Architecture & Skill Nodes:
● Voice AI● RAG Pipelines● STT/TTS Benchmarking● Odoo / LeadSquared● Production Triage
VERIFIED ENGINEERING & LEADERSHIP ENDORSEMENTS (LINKEDIN)
From Directors, MTS & Senior Backend Engineers @ VMware

“Zia has always been very meticulous in his approach and hard-working. I collaborated with him on multiple assignments and it was always good to work with him. He will add great value to any engineering team.”

Arnab Khound
Director UX @ Netskope (Ex-VMware)

“Ziauddin is proactive, result-oriented, responsible, and technically sound. His development skills are excellent and he is always ready to put in the effort to get mission-critical work done. A great asset to any company.”

Arfa Bakhtiar
Member of Technical Staff (MTS) @ VMware

“For me and many of us from the backend team, Zia was the go-to person for frontend architecture. His due diligence to understand the end-to-end use case enabled us to deliver performance-optimized frontend and backend integrations on time with high quality.”

Sanket Kadarkar
Sr. Software Engineer @ Spotnana (Ex-VMware, IITB)
ENGINEERING BLOG & SYSTEM WRITE-UPS

Engineering Field Notes & Deep-Dive Articles

Detailed architecture breakdowns covering sub-second Voice AI orchestration, hybrid Qdrant RAG evaluations, and low-latency backend primitives.

Full Project & Blog Archive
READY FOR HIGH-IMPACT AI & FDE CHALLENGES

Need an engineer who can design the architecture, benchmark the models, and lead the customer launch?

Whether you're scaling production Voice AI to millions of calls, eliminating hallucinations in clinical or enterprise RAG pipelines, or modernizing high-throughput backend systems, let's connect directly on LinkedIn.

Connect on LinkedInExplore GitHub & Benchmarks
How I Can Help Your Engineering TeamDirect DM on LinkedIn
Forward-Deployed Voice AI & GenAI: Taking enterprise voice agents from 1.8s dead air to sub-400ms natural conversations with 100% automated SOP QA.
Empirically Benchmarked RAG Systems: Replacing guesswork with MLflow evaluation matrices across Qdrant, Bedrock, and self-correcting LangGraph agents.
High-Velocity Full-Stack Architecture: Cutting platform latency by 90% with bidirectional WebSockets, Fastify, Redis, and React.
Message MD Ziauddin on LinkedIn →