# Carson Rodrigues

> Forward Deployed Engineer (FDE) for production LLM systems. Embeds with the customer and takes LLM systems from prototype to production, end to end — agentic AI (tool use, multi-agent orchestration, memory), RAG on Postgres + pgvector, LLM evals, real-time voice agents (LiveKit, Pipecat, Deepgram, ElevenLabs), MCP servers, enterprise system integrations, and the event-driven backends and apps around them, in TypeScript, Python and SQL. Based in London with the right to work in the UK; no sponsorship required. Three years forward deployed at VoiceQube (Sep 2023 – Sep 2026) across six client products (Altfolio, Kissan Pride, Finie, Northway Tech, and the sister products Phiny.ai and Crosslane) and three in-house ones (CueKit.ai, Unifund and Brillianse AI); Brillianse AI was launched publicly by Telangana's IT & Industries Minister in October 2026. Author of twenty research papers on the same systems: one published at ICANN 2026 (Springer LNCS, DOI 10.1007/978-3-032-38404-1_36), nine more submitted, eleven live preprints with DOIs.

Carson Rodrigues is a forward deployed engineer based in London, UK. He is available for Forward Deployed Engineer (FDE), senior AI and staff engineering roles in London and across the UK — on-site, hybrid or remote — and remote-first worldwide. From September 2023 to September 2026 he was a forward-deployed senior engineer and team lead at VoiceQube, running discovery, scoping and delivery directly with US and Indian clients and leading 30–40 engineers on and off across four product lines. There he worked across six client products (Altfolio, Kissan Pride, Finie, Northway Tech, featuring 2× US Open champion Andy North, and the sister products Phiny.ai, AI career prep with real-time voice mock interviews, and Crosslane, a career counselling platform) and three in-house products (CueKit.ai, Unifund, and Brillianse AI, a 13-app academic operations platform live across 12 campuses and launched in Hyderabad on 11 October 2026, with its AI admissions CRM). Other work and side projects: team lead of 6 (contract) at Ôdasie (Water of Asia), with on-site client work in Thailand: an AI itinerary builder, the main website redesign with NFQ (Vietnam), data cleaning and migration, open-source tools self-hosted on Coolify, and managing interns; enterprise IoT at Siemens; a full-stack & AI engineering contract on HotelClaw.ai, a multi-tenant AI platform for hotel operations, owning the PMS integration layer: Oracle OPERA exports and Opera Cloud (OHIP) generalised into one sync runner with Mews, Cloudbeds and Apaleo adapters on top, plus LINE and WhatsApp ingestion feeding an LLM intent router, with RAG on Postgres + pgvector and row-level security (case study: https://carsonrodrigues.com/projects/hotelclaw); co-founder of CELABE, an AI-powered wedding and events platform; and co-founder of V3NEX, a non-profit introducing kids to GenAI.

## Forward deployed engineering

From customer problem to production AI: four forward deployed disciplines in one engagement, with the evidence for each (they often run in parallel, with customer feedback shaping every stage):

- **01 Discover — find the real problem.** Sits in the customer's calls and on their floor and turns a vague ask into a scoped system: US-hours discovery for VoiceQube clients from India, on-site client work in Thailand for Ôdasie (twice), and Brillianse, built around the daily operations of a 12-campus junior-college group whose faculty approve every AI-drafted question.
- **02 Integrate — make it work in their stack.** The enterprise systems an AI product has to live inside: on HotelClaw.ai (a side project), five hotel PMS adapters behind one sync runner (39 fixture tests; OPERA exports live at the pilot resort, unchanged on re-import) and LINE and WhatsApp ingestion; tenant isolation made a CI check whose list of exceptions went from 54 to zero. Prototypes are built against the customer's real APIs, webhooks, enterprise systems and identity boundaries, not mock data: a working demo on the customer's own data in days rather than a slide deck: MCP servers into their systems, Claude Code in the loop, n8n for the glue. Example: the live OmniDirector demo at Google AI Day for Startups India 2026.
- **03 Productionize — ship AI that holds up.** Agents, RAG, real-time voice agents and the backends that keep them up, measured and then made faster: −41.8% median end-to-end voice latency (ICANN 2026 paper), 40,000+ golf courses served by Northway, 4,735 students and 7.6M exam answers on Brillianse.
- **04 Transfer — build the team that owns it.** Runs the team that carries the system (hiring, mentoring, technical roadmaps), writes the evals and docs, and feeds what the field taught back into product: 30–40 engineers across four product lines at VoiceQube; 447 of Brillianse's 2,688 merged pull requests.

Press: [The New Indian Express — "IT & Industries Minister Sridhar launches Hyderabad-developed Brilliance AI" (12 October 2026)](https://www.newindianexpress.com/cities/hyderabad/2026/Oct/12/it-industries-minister-sridhar-launches-hyderabad-developed-brilliance-ai). Product site: https://brillianse.ai.

## Writing on forward deployed engineering

- [What a Forward Deployed Engineer Actually Does](https://carsonrodrigues.com/blog/what-is-a-forward-deployed-engineer): The role from someone who lives it — owning AI deployments from first discovery call to production.
- [My Path to Forward Deployed Engineering](https://carsonrodrigues.com/blog/my-path-to-forward-deployed-engineering): Sales floor, a French agency in Thailand, US startup hours.
- [So You Want to Be a Forward Deployed Engineer (for AI)](https://carsonrodrigues.com/blog/becoming-an-fde): The skill stack and what FDE interviews probe.
- [FDE vs Solutions Engineer vs Consultant vs Product SWE](https://carsonrodrigues.com/blog/fde-vs-solutions-engineer): Why the FDE is a build role, not a support role.
- [Running Discovery for AI Agent Deployments](https://carsonrodrigues.com/blog/requirements-gathering-ai-agents): Finding the real workflow under the stated one and scoping a v1 the customer adopts.
- [Building PoCs and Demos That Win Enterprise Deals](https://carsonrodrigues.com/blog/poc-demos-that-win-deals): What a demo may fake and what must be real.
- [Deploying AI Agents in Enterprise Environments](https://carsonrodrigues.com/blog/deploying-agents-enterprise-environments): Milestones, environment tiers, go-live checklists, hypercare.
- [Evals for AI Agents: Build the Suite Before You Ship](https://carsonrodrigues.com/blog/evals-for-ai-agents): Golden sets from real transcripts and calibrated LLM judges.
- [The FDE Feedback Loop: Turning Deployment Pain Into Product](https://carsonrodrigues.com/blog/field-feedback-product-roadmap): Field reports engineers act on, without forking per customer.
- [Being the Customer's Trusted Technical Advisor](https://carsonrodrigues.com/blog/trusted-technical-advisor): Guiding architecture and saying no without overselling AI.
- [Surviving Enterprise Security Review as an AI Vendor](https://carsonrodrigues.com/blog/security-reviews-enterprise-ai): Data flows, retention, PII, prompt injection, SOC 2.
- [Integrating AI Agents with Legacy Systems](https://carsonrodrigues.com/blog/legacy-systems-ai-integration) and [with CRMs](https://carsonrodrigues.com/blog/integrating-agents-with-crms): SOAP, VPNs, Salesforce and HubSpot in the field.
- [What Actually Matters in Contact-Center AI](https://carsonrodrigues.com/blog/contact-center-ai-lessons): Demos are won on the model; renewals on AHT, CSAT and integrations.

## About

- [Home — what I do](https://carsonrodrigues.com/): Forward deployed engineer value prop, talks and live demos, the app slideshow, featured client projects, the FDE stack map, research, the four FDE disciplines (discover, integrate, productionize, transfer) and experience.
- [Projects — production systems, papers, OSS](https://carsonrodrigues.com/projects): Phiny.ai, Brillianse, Northway Tech, CELABE, Crosslane, SmokeMukti, Oravira, Amorelle, Coco, FoodieMama, Nourish; selected papers; everything-claude-code merged PR; Qubekit Voice AI Server.
- [Research — papers, DOIs, ORCID](https://carsonrodrigues.com/research): Twenty papers, one published at ICANN 2026 (Springer LNCS) and nine more submitted, across voice-AI systems, LLMOps, MCP, multi-agent reliability, agentic safety, human–AI trust, neuro-AI, interpretability, and clinical ML. ORCID 0009-0001-7195-6742; live preprints on arXiv, PsyArXiv, medRxiv, and Zenodo.
- [AutoScientist Challenge — 1st place, Data & Visualization](https://carsonrodrigues.com/adaption-autoscientist): Won Adaption Labs' AutoScientist Challenge Part 2 in Data & Visualization ($4,000 first prize, announced 18 August 2026) with PolyChart — a chart-reasoning model scored against an exact lie factor. Entered all seven categories off one model; backed by a controlled study of more than 200 fine-tuning jobs driven through a headless client written before the platform's official API shipped. Datasets and weights on HuggingFace and Kaggle, seven live demos, open-source client, two Zenodo preprints.
- [Case studies — long-form deep dives](https://carsonrodrigues.com/case-studies): Twelve long-form write-ups on the AI products I shipped end-to-end. Each one walks through Problem → Constraints → Architecture → What I built → Trade-offs → Outcome.
- [Timeline — the path so far](https://carsonrodrigues.com/timeline): 2026 → 2019, founder + papers + employment.
- [Dashboard — career snapshot + live GitHub metrics](https://carsonrodrigues.com/dashboard): 20 papers (1 published at ICANN 2026, 9 submitted), 5+ AI systems, 7 apps live on stores, 30–40 engineers led, 4+ years shipping, BITS Law, Mumbai speaker, Best Outgoing Student (CGPA 9.28).
- [OSS — open-source contributions with live GitHub data](https://carsonrodrigues.com/oss): Merged and open PRs across external repos, pulled live from GitHub.
- [About — bio and how I work](https://carsonrodrigues.com/about): Longer-form background, working style, and current focus.
- [Blog — writing on AI engineering](https://carsonrodrigues.com/blog): Posts on voice AI, LLM pipelines, and production engineering.

## Case studies

- [Phiny.ai — AI career-prep platform with real-time voice mock interviews](https://carsonrodrigues.com/projects/phiny): Smart Recruiter unified backend (NestJS, 516 TS files / ~101k LoC) + FastAPI WebRTC microservice (~17k LoC). Pipecat + LiveKit + Claude + ElevenLabs + Deepgram. Voice E2E p50 under 1.2s.
- [Northway Tech — AI voice caddie ft. Andy North](https://carsonrodrigues.com/projects/northway): Real-time voice caddie over LiveKit + WebRTC with 40,000+ courses geospatially indexed in MongoDB. Player-history loop personalises recommendations across rounds. ElevenLabs voice cloning of 2× US Open champion Andy North.
- [Brillianse — 13-app academic operations platform](https://carsonrodrigues.com/projects/brillianse): Hono API with 152 route modules, Drizzle ORM over PostgreSQL (323 tables), Redis + BullMQ. 13 apps (8 Next.js 15, 5 Expo SDK 55) and 55 packages in one Turborepo monorepo, 6,826 TypeScript files. Claude-drafted questions with a faculty review queue, never auto-published. Live for a 12-campus junior-college group as of September 2026: 4,735 students, 137,821 exam submissions, 7,602,196 answers. Carson merged 447 of the repo's 2,688 PRs. Launched publicly as Brillianse AI by Telangana IT & Industries Minister D. Sridhar Babu on 11 October 2026 ([The New Indian Express](https://www.newindianexpress.com/cities/hyderabad/2026/Oct/12/it-industries-minister-sridhar-launches-hyderabad-developed-brilliance-ai)). Product site: brillianse.ai; live admin: dev.23yards.com.
- [Brillianse CRM — AI admissions CRM with five-second speed to lead](https://carsonrodrigues.com/projects/brillianse-crm): The admissions side of Brillianse, built forward deployed for Indian junior colleges. Leads from Meta and Google ads (connected over OAuth) land in one queue; an AI voice agent calls each within five seconds, qualifies the lead and fills the enquiry form from the conversation; relationship managers click-to-call from the browser or take bridged calls (the agent dials the RM, then the parent, and listens to both) with live coaching and sentiment on screen; paid leads become students in the Brillianse admin panel without retyping. Brand Factory, a Lovable-style editor over each institution's logos and colours, generates landing pages, shareable links and AI ad creatives. Telephony on Twilio and Plivo SIP trunks with numbers per institution; set up for three institutions; 125 merged PRs, ~220k lines.
- [HotelClaw.ai — PMS integrations and tenant isolation for an assist-only hotel AI platform](https://carsonrodrigues.com/projects/hotelclaw): Contract. One PMS sync runner with five adapters: an Oracle OPERA export translator live at the pilot resort, plus Opera Cloud (OHIP), Mews, Cloudbeds and Apaleo, fixture-tested (39 tests) and awaiting their first hotel connection. LINE and WhatsApp webhooks (HMAC-verified) into an event log; an LLM intent router that drafts and routes while staff send every reply, with high-risk intents forced to human review. Tenant isolation as a CI check (row-level security on every table, exceptions 54 → 0). About 600 merged PRs.
- [CELABE — AI-powered wedding & events platform](https://carsonrodrigues.com/projects/celabe): Co-founded; React Native + Next.js across iOS, Android, web. ML-powered planning, vendor marketplace, viral QR photo-sharing. Live on App Store + Play Store.
- [SmokeMukti — ICMR-funded tobacco cessation app](https://carsonrodrigues.com/projects/smokemukti): React Native + Expo. Server-side AI chat (no client-embedded keys). Care-pathway state model. 1st place at AIDEC.
- [Oravira — oral health × HIV care-pathway app](https://carsonrodrigues.com/projects/oravira): Same care-pathway pattern as SmokeMukti applied to a new clinical domain. Local-first storage, server-mediated AI, App Store + Play Store live.

## Research

- [Latency Optimization in Production Voice AI Pipelines: A Systems Approach (Rodrigues & Vas, 2026)](https://doi.org/10.1007/978-3-032-38404-1_36): Systems-level latency study of a production voice AI platform; concurrency between intent detection and TTS is the highest-leverage optimization. **−41.8% E2E latency (p50 3,277 → 1,909 ms).** **Published at ICANN 2026** (peer-reviewed; Springer LNCS, Artificial Neural Networks and Machine Learning – ICANN 2026, pp. 450–461; doi:10.1007/978-3-032-38404-1_36).
- [MCP Server Architecture Patterns (Rodrigues & Vas, 2026)](https://carsonrodrigues.com/research): Design patterns for production Model Context Protocol servers — taxonomy, anti-patterns, cross-cutting concerns — validated with an inter-rater study (**κ = 0.76**). Preprint on arXiv (doi:10.48550/arXiv.2606.30317).
- [When Do LLMs Replace Fine-Tuned NLU? (Rodrigues & Vas, 2026)](https://carsonrodrigues.com/research): A decision framework for intent detection on noisy transcripts; full-data TF-IDF reaches **95.2% on ATIS**. arXiv preprint (doi:10.48550/arXiv.2608.20371); not accepted at AAAI-27 (September 2026). Continued as the sole-authored follow-up below.
- [Labeled Examples Decide the Gap Between Prompted LLMs and Fine-Tuned Intent Encoders: An Audit of Five LLMs and Three Decision Models (Rodrigues, 2026)](https://carsonrodrigues.com/research): A tuned RoBERTa, SetFit and simpler encoders against five prompted LLMs and three decision models 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. Submitted to ACL Rolling Review, October 2026 cycle (target: ACL 2027); continues arXiv:2608.20371 with new experiments and corrected conclusions.
- [Hyperparameter Advisors Built on Three LLMs Did Not Beat Default-Seeded TPE on Small Tabular Tasks (Rodrigues, Vas, DCosta & Prabhakaran, 2026)](https://carsonrodrigues.com/research): Three LLM advisors (Claude Haiku 4.5, GPT-4.1 mini, Llama 3.3 70B) against Optuna-TPE seeded with the same default configuration, under pre-specified hypotheses: **none beat default-seeded TPE; four of fifteen tests passed**, and the comparison depends on TPE's start-up trials. Revising; packaged for Applied Soft Computing, not yet submitted. Data on Zenodo (doi:10.5281/zenodo.23334877). The earlier single-model version is on arXiv (doi:10.48550/arXiv.2606.21641) and was not accepted at ACML 2026.
- [Personality & Trust in LLM Agents (2026)](https://carsonrodrigues.com/research): Scoping review extended to agentic LLMs; self-reported dependency evidence exists; no study relates a Big Five trait to behavioural over-reliance on an LLM. PsyArXiv preprint live (doi:10.31234/osf.io/2mdu8_v1). Resubmitted to IJHCI (Taylor & Francis).
- [Hallucination as Context Drift (2026)](https://carsonrodrigues.com/research): Context Divergence Score (CDS) and Shared State Verification Protocol (SSVP) for multi-agent LLM systems. **Correction (October 2026): the preprint's headline, that full-broadcast synchronization raises hallucination, is withdrawn.** A pre-registered, assertion-aware re-score showed the contradiction judges counted agents reporting a conflict as agents hallucinating it; the effect is not distinguishable from zero. arXiv preprint (doi:10.48550/arXiv.2606.21666) still states the earlier claim; a replacement version and a reframed manuscript for Information Processing & Management are in preparation.
- [A Zero-Shot Decision Model (Jev) Matches a Trained Classifier but Ranks Its Own Errors Worse: Calibrated, Split-Conformal Next-Procedure Recommendation in General Dentistry (2026)](https://carsonrodrigues.com/research): Clinical ML with a real public-data benchmark (AHRQ MEPS 2023); a zero-shot decision model matches the trained classifier's top-1 accuracy but ranks its own errors worse. medRxiv preprint live (doi:10.64898/2026.05.07.26352635). Submitted to BMC Personalized Medicine (October 2026), transferred after the Journal of Medical Systems declined it at editorial screening.
- [Fidelity Assessment for Digital Sketching: Why SSIM Fails on Line Art, and What to Use Instead (2026)](https://carsonrodrigues.com/research): Seven-author metric audit (Carson first author). Over 65 references an empty canvas averages 78.9% SSIM and outranks a wrong drawing on 96.9% of them; IoU on dilated edge maps separates good from bad drawings at AUC 0.984 vs 0.766, at 0.10 ms. Submitted to IEEE AIEI 2027 (notification 30 October 2026); audit code archived on Zenodo (doi:10.5281/zenodo.22237791).
- [Personality and Over-Delegation to Agentic LLMs (2026)](https://carsonrodrigues.com/research): Stage-1 Registered Report with a preregistered target of N = 500 (both confirmatory tests at 0.80 power or above); PCI-RR submission waits on ethics review and participant funding.
- [Done Is Not Correct — Silent Failures in Agentic CAD (2026)](https://carsonrodrigues.com/research): 468-run benchmark; single-shot silent failures 14.1% → 3.2% with enforced self-checks; expert–grader κ = 0.64. Submitted to Engineering Applications of Artificial Intelligence (Elsevier), September 2026; benchmark and records public (doi:10.5281/zenodo.22312792).
- [Oravira — Design, Privacy Architecture, and MARS App-Store Audit (2026)](https://carsonrodrigues.com/research): Privacy-by-design HIV oral-health app (live on iOS + Android) plus a systematic MARS audit; a 985-app search found no other consumer app at the HIV × oral-health intersection. MARS audit complete (two raters, ICC(2,1) = 0.946); manuscript in preparation.
- [DevTwin: Automatic Coverage Overstates How Well a Passively-Formed Memory Knows a Developer (2026)](https://carsonrodrigues.com/research): Benchmarks whether a memory built from a developer's public artifacts knows the person, not just the transcript; ten consented raters put identity knowledge at 0.45 where the automatic proxy saturates. Declined at the non-archival REALM workshop (EMNLP 2026, September 2026) and revised for resubmission; unpublished, so nothing is on the record.
- [Predicted Cortex Is Not a Domain-General Prior: Brain-Encoding Features for Video Memorability (2026)](https://carsonrodrigues.com/research): Dataset-dependent reversal between TRIBE-predicted brain responses and their own V-JEPA2 backbone, tested as an interaction (+0.097, p = 0.001). arXiv preprint v4 (doi:10.48550/arXiv.2607.16292); derived data on Zenodo (doi:10.5281/zenodo.21532633).
- [Hallucination Is Generative Memory With Its Verifier Turned Down: One Constraint Axis Links Dreaming Sleep and LLM Confabulation (2026)](https://carsonrodrigues.com/research): Perspective paper: dreaming and LLM hallucination as one operation, generative reconstruction under incomplete constraints; five falsifiable predictions and a small pre-registered PopQA test on two models. Zenodo preprint (doi:10.5281/zenodo.22314160); submitted to Minds and Machines (Springer), October 2026.
- [A Hosted LLM Fine-Tuning Service Kept Most Ordinary Training Answers but Overwrote Many Dataset-Specific Definitions and Abstentions: A User-Side Case Study With Answer Keys and Replicates (2026)](https://carsonrodrigues.com/research): A user-side case study of one anonymised hosted LLM fine-tuning service (384 jobs, plus 97 on a second base model across 16 other held-out domains): replicate SD 2.75 and 3.09 points, so single-run differences under about 8 points are noise; the default score orders 44.2% of same-dataset run pairs differently from the held-out score; checked against answer keys, the service's data rewriting kept most ordinary answers correct but turned 8.2% to 19.6% of correct training targets into incorrect ones, mostly dataset-specific definitions and abstentions. Augmentation itself helped (+13.4 on 10/10 paired seeds, scored by the service's own judge). Submitted to the International Journal of Data Science and Analytics on 19 October 2026, after a Harvard Data Science Review proposal was not invited on 17 October 2026; preprint on Zenodo (doi:10.5281/zenodo.23436412, version 3, with anonymised per-job data and code).
- [Lie Factors by Construction (2026)](https://carsonrodrigues.com/research): Exact, parameter-derived deception-severity labels for chart distortions: 2,430 exact-ground-truth rows, 636 exact lie factors, a 20-chart atlas of real misleading charts; dataset CC BY 4.0 (doi:10.5281/zenodo.21939803). In a pre-registered test on 210 inverted-axis pairs, the inverted axis reverses three of ten models' readings. Earlier version on Zenodo (doi:10.5281/zenodo.21939873); submitted to PacificVis 2027 on 11 October 2026.
- [Chart-Deception Probes in Vision-Language Models Rank Synthetic Axis Exaggeration but Do Not Reliably Beat a Mechanism-Only Baseline, and Qwen Probes Reverse on Real Charts (2026)](https://carsonrodrigues.com/research): Linear probes on four vision-language models' activations over 1,685 generated charts labelled with their vertical-scale ratio, under a pre-registered plan: misleading-or-not decodable at nearly every layer (AUROC 0.99); within truncated charts the probes rank exaggeration (Spearman 0.35 to 0.62) but none clearly beats a mechanism-only baseline; the model's stated Lie Factor is worse than a constant. On 2,572 real Misviz charts the frozen Qwen 7B and 72B probes score truncated axes below charts without a misleader (AUROC 0.43 and 0.24). Zenodo preprint v3 (doi:10.5281/zenodo.23333641; all versions doi:10.5281/zenodo.22134653), which corrects mislabelled control charts in the underlying generator; declined by IEEE TVCG on scope (October 2026); a machine-learning venue is being chosen.
- [When Frozen-Opponent Replays Overstate: An Audited Study of Opponent-Conditioned Policy Switching in a Two-Player Farming Simulation (Rodrigues, 2026)](https://carsonrodrigues.com/research): Audits a replay instrument for the Kaggriculture Kaggle simulation competition (10,246 teams): on 543 real games it located 84% of the baseline's losses in one opponent regime; an opponent-conditioned portfolio raised replayed wins from 46.9% to 57.4%, but a paired live control found most of that gain gone on the real ladder. In preparation; preprint on arXiv and Zenodo after the final leaderboard.
- [Effect of Implant Depth and Angulation on the Accuracy of Intraoral Scans Made With Scan Bodies and Scannable Healing Abutments: An In Vitro Study (Pai Khot, Kunkolienkar & Rodrigues, 2026)](https://carsonrodrigues.com/research): In vitro digital-dentistry study of implant depth (2 mm, 4 mm subcrestal) and angulation (0°, 15° lingual) on the trueness and precision of intraoral scans with scan bodies versus scannable healing abutments. Led by Dr. Atrey J. Pai Khot (Goa Dental College & Hospital); Carson did the dental CAD design and reviewed the drafts. In preparation.
- [What a Local Screen Could Not Tell Us: Measuring a Small Offline Coding Agent Under a Twelve-Hour Budget (Rodrigues, 2026)](https://carsonrodrigues.com/research): Measures whether a 74-task local screen predicted the hidden scores of 11 agent configurations in the Kaggle Gemma 4 Developer Agent competition; it did not (rank correlation +0.49, p = 0.32). 91 to 97 percent of each task's time is generation, so the time cap is a token budget, and solves arrive late. Zenodo preprint (doi:10.5281/zenodo.23592038); code and data at github.com/rodriguescarson/gemma-swe-agent.

## Snippets — production-tested code

- [MCP server in TypeScript](https://carsonrodrigues.com/snippets/mcp-server-typescript): Minimum-viable MCP server with one tool, exposed over stdio for Claude Code / Cursor / Desktop.
- [Claude streaming with prompt caching](https://carsonrodrigues.com/snippets/claude-streaming-prompt-cache): Anthropic SDK chat with cache_control on the system prompt — fastest cheapest path to a chat product.
- [Claude agentic tool-use loop](https://carsonrodrigues.com/snippets/claude-tool-use-loop): Run-until-stop agent loop in TypeScript. The core of every production agent.
- [Real-time voice agent (Pipecat + LiveKit)](https://carsonrodrigues.com/snippets/voice-agent-pipecat): Production voice agent — Deepgram → Claude → ElevenLabs over LiveKit WebRTC.
- [LiveKit token server (Node)](https://carsonrodrigues.com/snippets/livekit-token-server): Issue short-lived JWTs for browser/mobile to join a LiveKit room.
- [ASR intent detection with Claude](https://carsonrodrigues.com/snippets/intent-detection-claude): Post-ASR LLM classifier with structured output (the 41.7% → 91.7% pattern).
- [RAG with Algolia + Claude rerank](https://carsonrodrigues.com/snippets/rag-claude-rerank): Hybrid retrieval (BM25 + vector) plus Claude as a relevance judge.
- [n8n agentic webhook](https://carsonrodrigues.com/snippets/n8n-claude-webhook): n8n + Anthropic node config + the system prompt that works.
- [Claude Code slash command](https://carsonrodrigues.com/snippets/claude-code-slash-command): Custom /commands with shell inlining.
- [Claude Code pre-commit hook](https://carsonrodrigues.com/snippets/claude-code-hook): Block git commit unless lint + typecheck pass.

## Apps live on App Store + Play Store

Seven apps, each on both the App Store and Google Play:

- [CELABE](https://carsonrodrigues.com/projects/celabe): AI-powered wedding & events platform. iOS, Android, web. Co-founder.
- [SmokeMukti](https://carsonrodrigues.com/projects/smokemukti): ICMR-funded tobacco cessation app.
- [Oravira](https://carsonrodrigues.com/projects/oravira): Oral health × HIV care-pathway app.
- [Amorelle](https://carsonrodrigues.com/projects/amorelle): AI-native couples app for connection and intimacy.
- [Coco](https://carsonrodrigues.com/projects/coco): A quiet companion for your mind.
- [FoodieMama](https://carsonrodrigues.com/projects/foodiemama): Curated restaurant and street-food discovery.
- [Nourish](https://carsonrodrigues.com/projects/nourish): AI calorie and nutrition tracker.

## Open source

- [OSS page — all contributions with live GitHub data](https://carsonrodrigues.com/oss): Merged and open PRs across external repos.
- [everything-claude-code (now ECC)](https://github.com/affaan-m/everything-claude-code): Merged PR. 236k+ stars, 35k+ forks. Anthropic × Forum Ventures hackathon winner (NYC, 2025).
- Qubekit Voice AI Server: open-source LiveKit + Pipecat + Claude + ElevenLabs + Deepgram FastAPI reference implementation.

## Contact

- Calendly: https://calendly.com/rodriguescarson/new-meeting
- Topmate: https://topmate.io/rodriguescarson
- Email: admin@carsonrodrigues.com
- LinkedIn: https://linkedin.com/in/rodriguescarson
- GitHub: https://github.com/rodriguescarson
- X: https://x.com/carsonmarz

## Skills

Forward deployed: client discovery and scoping, working demos on customer data, enterprise system integration, production rollout and hypercare, LLM evaluation, security review, field feedback into product.
Leadership: engineering management, hiring, mentoring, technical roadmaps, architecture and code review.
Voice AI: LiveKit, Pipecat, Deepgram, ElevenLabs, WebRTC, WebSockets, SSE, real-time avatar.
AI / LLMs: agentic AI (tool use, multi-agent orchestration, memory and state), RAG, pgvector and vector search, LLM evaluation and benchmarking, prompt engineering, fine-tuning, MCP servers, Claude (Anthropic SDK), GPT-4, Gemini, OpenRouter, AI SDK, LangChain, Hugging Face.
Integrations: Oracle OPERA, Opera Cloud (OHIP), Mews, Cloudbeds, Apaleo, WhatsApp, LINE, Slack, Stripe, Razorpay.
AI Dev Tools: Claude Code (production user + OSS contributor), MCP configs, n8n, AG-UI.
Backend: Node.js, NestJS, FastAPI, Hono, Express, microservices, event-driven, BullMQ.
Frontend: React, Next.js 15, React Native (Expo), Tailwind 4, shadcn/ui, Framer Motion.
Data: SQL, PostgreSQL (Supabase, Drizzle, pgvector, row-level security), MongoDB (Mongoose, geospatial), Redis.
Cloud / DevOps: AWS (ECS Fargate, EC2, S3, SQS, Lambda, Bedrock, MediaConvert, CloudFront, Route 53, RDS, ElastiCache, SES), Cloudflare Workers, Docker, Terraform, GitHub Actions, CI/CD, Google Cloud Run, Netlify, Coolify.
Observability: Sentry, Prometheus, Grafana, PostHog.
Languages: Python, TypeScript, JavaScript, Java, C++.

## Speaking

All talks and workshops: https://carsonrodrigues.com/talks

- Google AI Day for Startups India 2026 (Google Ananta, Bengaluru, 21 August 2026): invited live demo of OmniDirector, an autonomous AI creative studio that takes a business from an idea to a finished ad campaign in one multimodal pipeline.
- BITS Law, Mumbai — AI for Marketing Masterclass (3-hour technical workshop): creative AI tooling, content + distribution automation, zero-code CRM pipelines, hands-on with LLM and image-gen tools. https://workshop.carsonrodrigues.com

## Recognition

- **1st place, Data & Visualization — Adaption Labs AutoScientist Challenge, Part 2 (2026).** $4,000 first prize; announced 18 August 2026. Winning entry: PolyChart, a model that scores how much a chart lies.
- **Ambassador, Adaption (Sep 2026 – present).** Volunteer community role at the non-profit behind the AutoScientist Challenge — events and meetups, testing new platform features and engagement formats, and turning member feedback into changes. Unpaid, not employment.
- Certificate of Reviewing — European Journal of Computer Sciences and Informatics (2026).
- Best Outgoing Student Award — Don Bosco College of Engineering, Goa. B.E. (Honours) Computer Engineering, CGPA 9.28.