Doriel Chiche.
Open to full-time / contract-to-hire · Jerusalem · open to Tel Aviv / hybrid

Doriel Chiche

AI / Full-Stack Engineer

I build LLM-powered products end to end — the model layer, the backend behind it, and the native clients on top. Seven years in production, the last three and a half spent making AI features reliable enough to ship.

Download CVLinkedInGitHubEmail

Routing layer, live

Animated diagram of the multi-provider fallback chain: the primary model times out after 1.4 seconds, the fallback answers in 720 milliseconds, and the response streams back with no user-visible error.

A replay of the real fallback chain. Fixed timings, no API call — the production version is in the Leirod backend.

7
years shipping production software
4
production codebases, built solo
1,500+
automated tests
~2 s
from photo to full nutritional breakdown

Selected work

Leirod — AI nutrition & coaching platform

Live on the App Store and Google Play. Designed, built and released solo across four production codebases: a Hono/Node.js API, native iOS and Android clients, and a Next.js site — carried by 1,500+ automated tests. The food-recognition pipeline pairs a vision model with the USDA FNDDS and French CIQUAL databases in PostgreSQL, returning a full nutritional breakdown in about two seconds per scan.

StackSwift / SwiftUIKotlin / ComposeNode.jsHonoPostgreSQLDrizzle ORM

Multi-provider LLM routing with real failover

Every AI feature declares a primary → fallback → tertiary chain across Gemini, Claude, GPT and Mistral. Model selection is config-driven, so swapping a model is an environment variable rather than a deploy, and a provider outage degrades that one feature instead of breaking the product. Google models are pinned to Vertex AI so EU data residency holds.

StackGeminiClaudeGPTMistralOpenRouterVertex AI

Streaming chat agent with tool use

A provider-agnostic multi-turn tool loop acting on the user's own data — log a meal, correct a portion, query history. Per-turn credit billing is protected by HMAC-signed, user-bound, short-lived tokens, so it cannot be bypassed from the client. Scores shown alongside it are computed in code, never by a model: the LLM only writes the narrative.

StackTool useSSE streamingHMACNode.jsPostgreSQL

LLM agents for small businesses

Telegram-based assistants automating back-office and operations workflows, built on the Hermes framework. Delivered to three clients, saving each around ten hours a week.

StackHermesLLM agentsTelegram APINode.jsn8n

Skills

AI / LLM

LLM application designAgents & tool useRAGVision modelsStreamingPrompt engineeringFallback & cost control

Languages

TypeScriptJavaScriptPythonSwiftKotlinSQLPHP

Backend

Node.jsHonoREST APIsPostgreSQLDrizzle ORMMongoDBMySQL/MariaDBOAuth 2.0

Frontend

ReactNext.jsVueTailwind CSSHTML/CSS

Mobile

Swift / SwiftUIKotlin / Jetpack ComposeApp Store & Play Store releases

Infrastructure

DockerLinuxGitHub ActionsNginxCloudflarepm2n8n

Quality

VitestJUnitXCTestPlaywrightSentryPostHog

About

Doriel Chiche

Full-stack engineer with seven years of production experience, now focused on LLM-powered products. Trained at École 42. Before going independent I spent four years at Retraite Plus in Jerusalem, building CRM systems used daily across the company and rolled out to enterprise clients, then coordinating delivery for a cross-functional team. I take features from database schema to store release, and I care about AI that holds up in production rather than in a demo. French (native), English (fluent), Hebrew (professional working proficiency).

Contact

Based in Jerusalem · open to Tel Aviv / hybrid.

doriel.chiche@gmail.com
LinkedInGitHub