Engineering Leader · Technical Co-Founder
Carson Rodrigues
I'm a product engineer — I architect and own features end-to-end. Production AI (voice agents, LLM pipelines), full-stack platforms, and mobile apps — and research papers on the systems I build (eighteen in total: one published at ICANN 2026, six more under review).
Impact at a glance
1st place · 18 August 2026
Won Adaption Labs's AutoScientist Challenge in Data & Visualization
PolyChart scores how much a chart lies, giving every distortion an exact lie factor rather than a yes/no label. Built on a controlled study of more than 200 fine-tuning jobs, driven through a headless client I wrote before the platform's official API shipped. Entered in all seven categories; placed first in this one.
Featured · funded · international · awarded
Research
Eighteen papers on the systems I ship
Voice AI latency, intent detection, MCP architecture, multi-agent reliability — measured on production workloads, not benchmarks. One published at ICANN 2026 (Springer LNCS), 7 more under peer review.
Published at ICANN 2026 (peer-reviewed, Springer LNCS) — “Latency Optimization in Production Voice AI Pipelines: A Systems Approach,” −41.8% end-to-end latency. doi:10.1007/978-3-032-38404-1_36
Latency Optimization in Production Voice AI Pipelines: A Systems Approach
A systems-level latency study of a production voice-AI platform (Anthropic Claude intent detection + ElevenLabs TTS over a NestJS WebSocket pipeline). The central finding is that running intent detection and TTS concurrently — rather than shaving any single stage — is the highest-leverage optimization, cutting median end-to-end latency from 3,277 ms to 1,909 ms.
MCP Server Architecture Patterns
A pattern catalogue for production Model Context Protocol (MCP) servers — a Gamma-format taxonomy, anti-patterns, and cross-cutting concerns — validated with a real inter-rater reliability study (kappa = 0.76). Proposes a scoped Proxy-Aggregator pattern once tool counts exceed a practical threshold.
When Do LLMs Replace Fine-Tuned NLU? A Decision Framework for Intent Detection
A decision framework for choosing between LLM classifiers and fine-tuned NLU on noisy production transcripts. Shows that full-data TF-IDF still reaches 95.2% on ATIS, and maps the regimes where an LLM-based intent classifier is — and is not — worth its cost and latency.
Labeled Examples Decide the Gap Between Prompted LLMs and Fine-Tuned Intent Encoders: An Audit of Five LLMs and Three Decision Models
Should a fine-tuned intent classifier be replaced by a prompted LLM? This study compares a tuned RoBERTa, SetFit and simpler encoders with five prompted LLMs and three decision models (which return a probability per label) on ATIS and CLINC150. Zero-shot, every LLM trails the encoder by 13.4 to 27.5 points on ATIS; with 20 retrieved training examples an LLM comes within 1.6 points on ATIS and exceeds the encoder on CLINC150. From label names alone, the best decision model matches the encoder's CLINC150 in-scope accuracy (95.8 vs 96.3) but trails it by 6.4 points on ATIS. A validation-tuned threshold lets the encoder beat every zero-shot LLM on both in-scope accuracy and out-of-scope recall.
What I do
Four archetypes I get hired for
Voice AI, LLMOps, distributed backends, and engineering leadership — usually two or three at the same time.
Real-time voice AI
What it covers (6)
End-to-end voice agents and avatars: ASR, intent, TTS, transport, latency optimization. Production-grade across 40k+ locations.
End-to-end voice agents and avatars: ASR, intent, TTS, transport, latency optimization. Production-grade across 40k+ locations.
→ −41.8% E2E latency · paper
LLM pipelines & MCP
What it covers (6)
RAG, agent orchestration, evals, MCP server design. Anthropic SDK + multi-model routing. Author of MCP Server Architecture paper.
RAG, agent orchestration, evals, MCP server design. Anthropic SDK + multi-model routing. Author of MCP Server Architecture paper.
→ MCP architecture paper · 5+ AI systems shipped
Distributed backends
What it covers (6)
Event-driven NestJS / FastAPI / Hono backends, geospatial pipelines, queues, observability. The plumbing that keeps AI products up.
Event-driven NestJS / FastAPI / Hono backends, geospatial pipelines, queues, observability. The plumbing that keeps AI products up.
→ 40,000+ locations · 99.9% uptime
Engineering leadership
What it covers (6)
Lead 6+ engineers through architecture reviews, hiring, and shipping cadence. Cut release cycles weeks → hours via infra automation.
Lead 6+ engineers through architecture reviews, hiring, and shipping cadence. Cut release cycles weeks → hours via infra automation.
→ Team of 6 at Ôdasie · founder × 2
Speaking
Talks, workshops and live demos
A live OmniDirector demo at Google's AI Day for Startups India 2026, a 3-hour AI for Marketing masterclass at BITS Law, Mumbai, and guest sessions back at my alma mater.
Featured
Production systems shipped
AI Career Prep Platform
Senior Software Engineer, Forward Deployed & Team Lead · VoiceQube
- Real-time AI voice mock interviews — Pipecat + LiveKit + Claude + ElevenLabs + Deepgram
- LLM-driven resume tailoring per JD with RAG over candidate documents
- Job matching, application tracker, interview prep, upskilling paths
- Resume parsing pipeline: Tesseract OCR + pdf-parse + mammoth
- Multi-payment integration: Razorpay, Stripe, Apple Pay
Academic Operations Platform
Senior Software Engineer, Forward Deployed & Team Lead · VoiceQube
- 447 of the repo's 2,688 merged PRs are mine; most commits of any contributor
- Admissions CRM with an AI voice agent that calls new enquiries in five minutes (~220k LoC)
- IPE: AI grading of handwritten answer sheets, 12,847 marks reviewed by faculty in prod
- Hono API (152 route modules), Drizzle over Postgres — 323 tables, 216 migrations
- 125,299 questions in the bank, 78% AI-drafted, none published without faculty approval
AI Voice Caddie ft. Andy North (2× US Open Champion)
Senior Software Engineer, Forward Deployed & Team Lead · VoiceQube
- Real-time voice caddie via LiveKit + WebRTC stack
- MongoDB geospatial indexing for live GPS course position
- AWS Lambda + MediaConvert for shot video capture and processing
- Production-grade across 40,000+ golf courses (event-driven backend)
- Voice + intent stack with Anthropic Claude in the loop
- Architected end-to-end across iOS, Android, web, vendor ecosystem
- ML-powered planning workflows and intelligent vendor matching
- Viral QR photo-sharing for live event capture
- Live on App Store and Google Play; multi-market customer base
- Established engineering culture, CI/CD automation, technical roadmap
Mobile apps
Live on the App Store and Play Store
React Native / Expo across health-tech and consumer products. Multi-market customers downloading them today.
CELABE
AI-Powered Wedding & Events Platform
Co-founded multi-market wedding / events platform. iOS, Android, web; ML-powered planning, vendor marketplace, viral QR photo-sharing. Founding engineering leader.
SmokeMukti
ICMR-Funded Tobacco Cessation App
Behavior-change app for tobacco cessation funded by the Indian Council of Medical Research (ICMR). Personalized AI chat coach, progress tracking, goal setting, and 24/7 motivation pathway. Built under guidance of Dr. Anil V. Ankola; ICMR STS-backed research project.
Experience
Where I've shipped
- Co-founded and architected an AI-powered wedding / events platform
- Led full lifecycle from concept to production launch (iOS, Android, web, vendor ecosystem)
- ML-powered planning workflows, viral QR photo-sharing, real-time vendor coordination
- Established engineering culture, CI/CD automation, and the technical roadmap
- Led a team of 6 engineers through architecture reviews, code reviews, and technical mentorship
- Architected and shipped 5+ production AI systems: LLM pipelines, voice agents, MCP servers, geospatial processors
- Real-time AI avatar with LiveKit + WebRTC (bi-directional voice + video + data)
- Heavy n8n automation — LLM agent orchestration, API integrations, approval workflows
- Adopted Claude Code in production engineering workflow; cut release cycles weeks → hours
- Three years leading engineering teams — 30–40 engineers on and off across four product lines, through architecture reviews, code reviews, and mentorship
- Forward deployed: discovery, scoping and delivery run directly with overseas clients, remote and on-site
- Shipped four production AI products end-to-end: Phiny.ai, Crosslane, Northway Tech, Brilliance
- Architected NestJS + AWS backend serving real-time interactions across 40,000+ locations
- Live AI avatars and voice agents — LiveKit, Pipecat, ElevenLabs TTS, Deepgram ASR, Anthropic intent
- Brilliance: 447 of 2,688 merged PRs and more commits than any other contributor on a 13-app platform
- Engineered enterprise IoT desktop applications — 20k+ LoC, 1,000+ Jest tests
- Architected real-time MQTT communication and concurrent schedulers — ~3× performance improvement
- Stack: Angular, Electron, RxJS, NestJS, RxDB, Python, MQTT pub/sub
Stack
What I build with
Voice AI and LLMOps day-to-day. Heavy lean on Claude, MCP, and the agentic stack.
Voice AI
AI / LLMs
AI Dev Tools
Backend
Frontend
Cloud / DevOps
Data
Observability
Writing
Latest articles
Available · London / UK · senior AI, FDE, contract
Building something with AI?
Voice agents, MCP servers, LLM pipelines, agentic workflows — pick a slot, drop a message, or send your email and I'll reply within a day.
Replies within ~24 hours · London, UK · on-site, hybrid or remote