harness ready.
Nare Sakhoyan
AI Engineer · Full-Stack — ships LLM systems into production
accepting connections — remote · contract · immediate
Anyone can call a model.
I build the harness around it —
the loop · the limits · the retrieval · the graders
— the part that decides whether it can be trusted.
this page is one. keep scrolling.
anatomy
This site is the schematic
the model never sees a request the limiter didn’t approve
every answer traces to a retrieved chunk — cite or decline
latency and cost are printed under every reply, not hidden in logs
8 cases · 3 traps defined — run pending, failures will be included
[ you are here ]
accepting connections — remote · contract or EOR · start immediately
Looking for Applied AI Engineer · Forward Deployed Engineer · AI Infrastructure / Agent Platform Engineer · AI Systems Engineer.
modules
module 01/05
Job-search agent
agent harness · 4 graders · human approval gate
Autonomous pipeline that finds postings, scores fit, tailors a résumé per role with Claude tool use, and fills application forms via browser automation.
its failures would ship under my name, so it earns autonomy through evals
- TypeScript
- pnpm/turbo
- Nest.js
- Next.js
- Postgres
- Docker
module 02/05
Ask Nare
rag harness · live · you're using it
The chat widget on this site: RAG + tool use + guardrails. Live now — the eval scorecard is in progress.
retrieval, tools, guardrails, telemetry — the site answers for itself
- Next.js
- @anthropic-ai/sdk
- BM25 / embeddings
- Vitest
module 03/05
llm-evalkitin progress
pure eval harness · 3 models · cost/latency report
TypeScript eval harness comparing Haiku / Sonnet / Opus on résumé tailoring: schema, hallucination, keyword-coverage and LLM-judge graders with a cost/latency report. In progress — repo and report not yet published.
“which model?” is an empirical question
- TypeScript
- Anthropic SDK
- Zod
- Vitest
module 04/05
Armenian NLP toolkit
eval when no benchmark exists
NLP for a low-resource language, with a README that states results and limitations honestly.
for a low-resource language you build the measuring stick first
- Python
- Armenian
- Low-resource NLP
module 05/05
Foody
classic full-stack · the foundation
Recipe + pantry app: pantry-based recommendations, favourites, drafts, search history.
the substrate every harness stands on
- Nest.js
- Drizzle ORM
- Postgres
- Clerk
- Cloudinary
- Next.js
runtime history
jan 2020 → dec 2021 · yerevan, armenia
Matemat
Full-Stack Engineer
foundations: full-stack, search, data
- Full-stack development with Node, Vue, MongoDB, GraphQL, and Elasticsearch.
- Shipped search and data features end to end for a small product team.
mar 2022 → dec 2023 · us · remote
Hearst Autos
Full-Stack Engineer
scale: the platform behind Car and Driver & every Hearst site
- Worked on the platform powering Car and Driver, Autoweek, and the shared system behind all Hearst sites.
- Built Next.js front-ends and Node.js / Python microservices at publisher scale.
may 2024 → mar 2025 · us · remote
Turing
Software Engineer — AI training data
inside the reward model — I wrote and graded what teaches models “correct”
- Authored and graded reference solutions (HTML/CSS/JS/TS/Node/React/SQL) used to train reward models that improve LLM coding accuracy.
- Learned how models are evaluated from the data side — rubric design, grading consistency, and failure taxonomies.
nov 2025 → mar 2026 · netherlands · remote
Prostrive
Senior Full-Stack Engineer
orchestration: multiple LLM providers, automated validation of AI answers
- Core engineer on PromptTracker (now Answer One), an Answer Engine Optimization platform — Next.js + Nest.js/Node.
- Connected multiple LLM providers to automate prompt generation, scraping, and validation of how brands appear in AI answers. Role ended when funding ended.
cs coursework · université toulouse iii · 2019–2023
self-test
A harness that can’t fail its own tests isn’t testing anything.
deterministic checks + llm judge · fails the build under threshold · full scorecard →
console
This is a harness with a model inside.
interrogate it.
Ask Nare
Answers only from my published profile · cites sources · 20/20 questions left
Powered by Claude. Won’t discuss salary, won’t invent facts; if it doesn’t know, it says so. See the eval scorecard →
status
accepting connections
Remote worldwide · contract or EOR · start immediately. Looking for Applied AI Engineer · Forward Deployed Engineer · AI Infrastructure / Agent Platform Engineer · AI Systems Engineer — shipping LLM systems into production: harnesses, evals, RAG, tool use.