AI Prompt & Data Engineer

AI Prompt & Data Engineer at YesAstrology

The AI Prompt & Data Engineer role at YesAstrology is a full-time, fully remote position open to candidates worldwide, sitting within the AI & Engineering department. You will design, optimize, and orchestrate production LLM prompt chains, retrieval-augmented generation (RAG) pipelines, and rigorous evaluation harnesses that deliver highly personalized, structured insights at scale. YesAstrology operates at the intersection of high-precision astronomical calculation and modern artificial intelligence, and this role is the engineering bridge between the two.

Day to day, you will design structured prompt architectures, implement multi-agent workflows, build robust RAG retrieval pipelines, and establish automated evaluation benchmarks that eliminate model hallucinations, tone inconsistencies, and latency bottlenecks across millions of monthly user requests. These systems power real-time interactive consultations, personalized horoscopic reports, and dynamic astrological analysis engines — which means your prompts must respect precise chart data such as house cusps, planetary degrees, and aspect orbs rather than paraphrasing them loosely. A report that misstates a Midheaven degree or confuses a trine with a square is a factual failure, and your evaluation harnesses are what catch it before a user does.

The technical surface is concrete: complex system prompts, few-shot contexts, and chain-of-thought workflows across modern LLM providers; RAG pipelines that merge domain knowledge with real-time computational data; and automated evals measuring factual accuracy, structured schema adherence (Zod / JSON Schema), and latency. You will also optimize model routing, caching strategies, and token expenditure to scale generative features cost-effectively, and collaborate with product and backend engineers to integrate intelligent agents into high-throughput Next.js and Supabase APIs. For a broader view of the platform your work will feed, explore the astrology tools that surface these calculations to users.

What We Look For

• 2+ years of professional experience building production-grade LLM applications, prompt engineering workflows, or AI agent systems
• Demonstrated proficiency in TypeScript, Python, or Node.js with modern AI SDKs and vector databases (pgvector, Pinecone, or Qdrant)
• Strong understanding of prompt optimization techniques, structured output enforcement, and automated eval methodologies
• Experience with context management, embeddings, semantic retrieval, and multi-model routing (OpenAI, Anthropic, Gemini, open-source models)
• Fluent written and verbal English communication with high attention to detail

Common Questions

What models and infrastructure will I work with? An ensemble of modern LLMs (Anthropic Claude, OpenAI, Google Gemini) coordinated with Next.js, TypeScript, Supabase, and vector embeddings.

Is prior astrology knowledge required? No. Strong engineering instincts, curiosity about data structures, and proficiency in LLM mechanics matter most.

Is this role fully remote? Yes — 100% remote, open to candidates worldwide, with flexible working hours.

Role Overview and Technical Responsibilities

The AI Prompt & Data Engineer at YesAstrology is a full-time, fully remote role open to candidates worldwide, sitting within the AI & Engineering department. The position exists to design, optimize, and orchestrate production LLM prompt chains, retrieval-augmented generation (RAG) pipelines, and rigorous evaluation harnesses that deliver highly personalized, structured insights at scale. In practice, this means architecting, benchmarking, and deploying the AI systems behind real-time interactive consultations, personalized horoscopic reports, and dynamic astrological analysis engines.

Day to day, the engineer designs, benchmarks, and optimizes complex system prompts, few-shot contexts, and chain-of-thought workflows across modern LLM providers. They architect RAG pipelines that integrate domain knowledge with real-time computational data — for example, ensuring a generated reading correctly reflects a natal chart computed with Placidus house cusps rather than drifting into generic Sun-sign language. They also implement automated evaluation harnesses that measure factual accuracy, structured schema adherence (Zod / JSON Schema), and latency, so that a response describing Saturn transiting the 10th house is both technically valid and astrologically coherent.

Beyond prompt craft, the role carries clear engineering ownership. The engineer optimizes model routing, caching strategies, and token expenditure to scale generative AI features cost-effectively, and collaborates with product and backend engineers to integrate intelligent agents into high-throughput Next.js and Supabase APIs. The stated goal is to eliminate model hallucinations, tone inconsistencies, and latency bottlenecks across millions of monthly user requests — a mandate that directly protects the credibility of interpretations covering zodiac signs, planetary aspects, and house relationships.

Core Responsibilities

• Design, benchmark, and optimize complex system prompts, few-shot contexts, and chain-of-thought workflows across modern LLM providers.
• Architect retrieval-augmented generation (RAG) pipelines integrating domain knowledge with real-time computational data.
• Implement automated evaluation harnesses to measure factual accuracy, structured schema adherence (Zod / JSON Schema), and latency.
• Optimize model routing, caching strategies, and token expenditure to scale generative AI features cost-effectively.
• Collaborate with product and backend engineers to integrate intelligent agents into high-throughput Next.js and Supabase APIs.

Requirements and Working Environment

Candidates should bring 2+ years of professional experience building production-grade LLM applications, prompt engineering workflows, or AI agent systems. Demonstrated proficiency in TypeScript, Python, or Node.js with modern AI SDKs and vector databases (pgvector, Pinecone, or Qdrant) is required, alongside a strong understanding of prompt optimization techniques, structured output enforcement, and automated eval methodologies. Experience with context management, embeddings, semantic retrieval, and multi-model routing (OpenAI, Anthropic, Gemini, open-source models) rounds out the technical profile, together with fluent written and verbal English communication skills and high attention to detail.

On infrastructure, the team uses an ensemble of modern LLMs (Anthropic Claude, OpenAI, Google Gemini) coordinated with Next.js, TypeScript, Supabase, and vector embeddings. No prior astrology domain background is required — strong engineering instincts, curiosity about data structures, and proficiency in LLM mechanics are what matter most. The role is 100% remote with flexible working hours. Candidates can explore the broader product context through the YesAstrology tools before applying.

AI Prompt Engineering and LLM Optimization

AI prompt engineering and LLM optimization at YesAstrology means designing system prompts, few-shot contexts, and chain-of-thought workflows that force a language model to respect exact chart data — house cusps, planetary degrees, and aspect orbs — instead of paraphrasing them loosely. The engineer benchmarks these prompt chains across an ensemble of modern LLMs (Anthropic Claude, OpenAI, Google Gemini, and open-source models), then tunes model routing, caching, and token expenditure so generative features scale without waste. The measure of success is simple: a response describing Saturn transiting the 10th house must be both technically valid and astrologically coherent.

Prompt architecture is where most accuracy is won or lost. A well-built system prompt passes the computed chart as structured input — for instance, a natal chart calculated with Placidus house cusps, the Midheaven degree, and a list of applying and separating aspects — and instructs the model to reason only from those values. Few-shot examples anchor tone and format: one example might show a Capricorn Midheaven with Saturn in the 10th interpreted as earned mastery and slow-built authority, while another shows Jupiter in the 10th read as expansive public visibility. Chain-of-thought steps then walk the model through sign, house, and aspect in sequence, which reduces the drift into generic Sun-sign language that plagues ungrounded horoscope generation.

Optimization is continuous and measurable. Automated evaluation harnesses score every prompt revision on factual accuracy, structured schema adherence (Zod / JSON Schema), and latency, so a regression in how the model handles an opposition versus a square is caught before a user sees it. Model routing sends simpler tasks — formatting a daily horoscope summary — to faster, cheaper models, while complex multi-factor synthesis across the 10th house career axis, the 6th house daily work, and the 2nd house income stays on stronger models. Caching stores repeated computations and stable interpretive passages, cutting token expenditure across millions of monthly user requests.

How Grounding and Evaluation Work Together

Retrieval-augmented generation pipelines merge domain knowledge with real-time computational data, so the model retrieves the correct interpretive framework for a Chiron placement or a nodal axis rather than inventing one. The evaluation harness then tests that retrieved context actually reached the output: does the generated reading name the correct house system, preserve the exact degree of the Ascendant, and distinguish a conjunction from a trine? Because the platform's calculations feed directly into these prompts, the engineering work stays tied to the same chart mechanics users see in the personalized reports.

Incomplete birth data is handled transparently rather than silently. When an exact birth time is missing, a solar-noon chart default preserves planetary sign and aspect accuracy while flagging that house cusps and the Midheaven are approximate — and the prompt is instructed to communicate that limitation instead of asserting a precise 10th-house degree. Timing layers add a second dimension: transits, solar arcs, and progressions activate chart potential over time, so prompts must distinguish a natal promise from a currently transiting influence, such as Pluto crossing the Midheaven versus Pluto natally placed in the 4th house.

Training Data Preparation and Dataset Management

Training data preparation and dataset management in this role means curating, structuring, and validating the astrological knowledge and computational records that ground every LLM prompt chain and RAG pipeline at YesAstrology. The engineer assembles domain corpora covering zodiac signs, planets, house systems, and aspect dynamics, then pairs them with live chart computation output so retrieval returns the correct degree, cusp, and orb rather than a plausible-sounding approximation. Dataset quality is treated as an engineering deliverable, not an editorial afterthought.

Concretely, this involves building and maintaining embedding stores in pgvector, Pinecone, or Qdrant, chunking interpretive material so that a query about Saturn transiting the 10th house retrieves passages about career mastery and earned authority rather than generic Saturn text. The engineer must reconcile competing house systems: a chart cast in Placidus places house cusps at unequal intervals, while Whole Sign assigns each house a full zodiac sign, and Equal House divides the wheel into thirty-degree segments from the Ascendant. Because a single birth time can shift a cusp by several degrees, the dataset schema records which system produced each stored chart so retrieval never mixes Placidus cusps with Whole Sign placements in the same response.

Dataset management also covers the handling of incomplete birth data. When an exact birth time is unavailable, the pipeline defaults to a solar-noon chart, which preserves planetary sign positions and aspect accuracy while leaving house cusps and the Midheaven provisional. The engineer tags these records accordingly so downstream prompts can state which factors are reliable and which are estimated, and so evaluation harnesses do not penalize a model for correctly flagging an uncertain Ascendant. Aspect orbs are stored explicitly, letting retrieval distinguish a tight conjunction from a wide one and preventing a trine from being described as a square.

Validation, Versioning, and Refresh Cycles

Every dataset release is versioned and gated by automated checks: schema adherence through Zod or JSON Schema, factual spot-checks against computed ephemeris values, and latency budgets for retrieval calls. When a new interpretive corpus is added, the evaluation harness re-runs benchmark prompts to confirm that outputs still respect structured fields such as planetary degrees and house placements. Timing cycles add a further dimension, since transit, solar arc, and progression data must be refreshed as charts advance; the engineer schedules these updates so cached embeddings do not serve stale positions. Candidates curious about how these grounded calculations reach readers can review the platform's personalized astrology reports, which depend directly on the dataset discipline described here.

AI Content Generation Workflows and Automation

AI content generation workflows and automation at YesAstrology are the orchestration layer that turns a computed natal chart into a finished, reader-ready horoscope without human intervention on every request. The AI Prompt & Data Engineer designs these workflows as multi-step prompt chains: the system prompt receives structured chart data, a retrieval step pulls the matching interpretive framework, a generation step drafts the reading, and an evaluation step validates accuracy, schema adherence, and latency before delivery. Automation is what allows millions of monthly user requests to be served consistently, and the engineer owns the routing, caching, and validation logic that keeps that pipeline both fast and factually sound.

A concrete workflow makes the mechanics visible. For a personalized horoscopic report, the pipeline first ingests the chart computed with Placidus house cusps, the exact Midheaven degree, and a list of applying and separating aspects. A retrieval call then fetches passages keyed to those placements — for a Capricorn Midheaven with Saturn in the 10th, the retrieved framework emphasizes earned mastery and slow-built authority, whereas Jupiter in the 10th retrieves material on expansive public visibility. The generation step walks sign, house, and aspect in sequence through chain-of-thought prompting, and the evaluation harness confirms the output names the correct house system and preserves the Ascendant degree before the response is returned. If the model drifts into generic Sun-sign language or confuses a trine with a square, the harness flags the response and the prompt revision cycle begins.

Automation also governs how the workflow handles incomplete data and timing layers. When an exact birth time is missing, the pipeline defaults to a solar-noon chart, which preserves planetary sign positions and aspect accuracy while leaving house cusps and the Midheaven provisional; the prompt is instructed to state that limitation rather than assert a precise 10th-house degree. Timing cycles add a second dimension, because transits, solar arcs, and progressions activate chart potential over time, so the workflow must distinguish a natal promise from a currently transiting influence — Pluto crossing the Midheaven reads differently from Pluto natally placed in the 4th house. Model routing sends simpler tasks, such as formatting a daily horoscope summary, to faster, cheaper models, while complex multi-factor synthesis across the 10th house career axis, the 6th house daily work, and the 2nd house income stays on stronger models. Caching stores repeated computations and stable interpretive passages so token expenditure stays controlled as volume grows.

How the Automated Pipeline Is Validated

Every workflow revision is gated by automated evaluation harnesses that score factual accuracy, structured schema adherence through Zod or JSON Schema, and latency. When a new interpretive corpus is added to the embedding stores in pgvector, Pinecone, or Qdrant, the harness re-runs benchmark prompts to confirm that outputs still respect structured fields such as planetary degrees and house placements, and that retrieval never mixes Placidus cusps with Whole Sign placements in the same response. Aspect orbs are stored explicitly so the pipeline can distinguish a tight conjunction from a wide one. Because transit, solar arc, and progression data must be refreshed as charts advance, the engineer schedules updates so cached embeddings do not serve stale positions. Readers who want to see how these grounded calculations reach them can explore the platform's personalized astrology reports, which depend directly on the workflow discipline described here.

Data Quality, Validation, and Output Evaluation

Data quality, validation, and output evaluation at YesAstrology means building automated evaluation harnesses that score every generated reading on three measurable axes: factual accuracy against computed chart data, structured schema adherence enforced through Zod or JSON Schema, and latency. These benchmarks exist to eliminate model hallucinations, tone inconsistencies, and latency bottlenecks across millions of monthly user requests. In practice, a response describing Saturn transiting the 10th house must be both technically valid and astrologically coherent before it ever reaches a reader.

Validation begins with the chart record itself. A natal chart computed with Placidus house cusps stores unequal cusp degrees, while Whole Sign assigns each house a full zodiac sign and Equal House divides the wheel into thirty-degree segments from the Ascendant. Because a single birth time can shift a cusp by several degrees, the schema records which house system produced each stored chart, so retrieval never mixes Placidus cusps with Whole Sign placements in the same response. Aspect orbs are stored explicitly, letting the harness distinguish a tight conjunction from a wide one and preventing a trine from being described as a square. When an exact birth time is unavailable, the pipeline defaults to a solar-noon chart, which preserves planetary sign positions and aspect accuracy while leaving house cusps and the Midheaven provisional; the engineer tags these records so downstream prompts state which factors are reliable and which are estimated, and so the harness does not penalize a model for correctly flagging an uncertain Ascendant.

What the Evaluation Harness Measures

• Factual accuracy: does the output preserve the exact Midheaven degree, name the correct house system, and keep a Capricorn Midheaven with Saturn in the 10th distinct from Jupiter in the 10th?
• Schema adherence: Zod and JSON Schema checks confirm every structured field — planetary degrees, house placements, aspect types — is present and correctly typed.
• Latency: retrieval and generation calls are held to defined budgets so interactive consultations stay responsive.
• Tone consistency: few-shot anchored outputs are compared for drift into generic Sun-sign language.

Evaluation is continuous rather than a one-time gate. Every prompt revision re-runs benchmark prompts, so a regression in how the model handles an opposition versus a square is caught before a user sees it. Model routing sends simpler tasks — formatting a daily horoscope summary — to faster, cheaper models, while complex multi-factor synthesis across the 10th house career axis, the 6th house daily work, and the 2nd house income stays on stronger models. Caching stores repeated computations and stable interpretive passages, cutting token expenditure without serving stale positions.

Timing layers add a second dimension to validation. Transits, solar arcs, and progressions activate chart potential over time, so prompts must distinguish a natal promise from a currently transiting influence — Pluto crossing the Midheaven today versus Pluto natally placed in the 4th house. The engineer schedules refreshes of transit, solar arc, and progression data so cached embeddings never serve outdated positions, and re-runs the harness whenever a new interpretive corpus is added to confirm outputs still respect structured fields such as planetary degrees and house placements. Readers can see the results of this discipline in the platform's personalized astrology reports, which depend directly on the dataset and evaluation rigor described here.

API Integration and AI Model Performance Optimization

API integration and AI model performance optimization in this role means wiring intelligent agents into high-throughput Next.js and Supabase APIs while tuning model routing, caching, and token expenditure so generative features scale without waste. The engineer treats every chart-derived payload as structured data that must survive the round trip intact — a Midheaven degree, a house cusp, an aspect orb — and holds each integration to measurable budgets for factual accuracy, schema adherence, and latency across millions of monthly user requests.

Integration begins with the contract between the calculation layer and the language model. A natal chart computed with Placidus house cusps produces unequal cusp intervals, while Whole Sign assigns each house a full zodiac sign and Equal House divides the wheel into thirty-degree segments from the Ascendant; the API payload must declare which system produced the chart so downstream prompts never blend Placidus cusps with Whole Sign placements in a single response. Aspect orbs travel with the payload explicitly, letting the model distinguish a tight conjunction from a wide one and preventing a trine from being rendered as a square. When an exact birth time is missing, the pipeline defaults to a solar-noon chart, which preserves planetary sign positions and aspect accuracy while leaving house cusps and the Midheaven provisional — the API flags that uncertainty so the generated reading states the limitation rather than asserting a precise 10th-house degree.

Performance optimization is continuous and measured. Model routing sends simpler tasks, such as formatting a daily horoscope summary, to faster, cheaper models, while complex multi-factor synthesis across the 10th house career axis, the 6th house daily work, and the 2nd house income stays on stronger models in the ensemble (Anthropic Claude, OpenAI, Google Gemini, and open-source models). Caching stores repeated computations and stable interpretive passages so token expenditure stays controlled as volume grows, and retrieval calls are held to defined latency budgets so interactive consultations stay responsive. Every prompt or integration revision re-runs benchmark prompts, catching a regression in how the model handles an opposition versus a square before a user ever sees it.

How Timing Layers Stress the Integration

Transits, solar arcs, and progressions activate chart potential over time, so the API must distinguish a natal promise from a currently transiting influence — Pluto crossing the Midheaven today reads very differently from Pluto natally placed in the 4th house. Because transit, solar arc, and progression data must be refreshed as charts advance, the engineer schedules updates so cached embeddings never serve stale positions, and re-runs the evaluation harness whenever a new interpretive corpus is added to confirm outputs still respect structured fields such as planetary degrees and house placements. Readers can see the results of this discipline through the platform's astrology and horoscope platform, which depends directly on the integration and performance rigor described here.

Required Skills, Experience, and Qualifications

Candidates need 2+ years of professional experience building production-grade LLM applications, prompt engineering workflows, or AI agent systems, plus demonstrated proficiency in TypeScript, Python, or Node.js with modern AI SDKs and vector databases such as pgvector, Pinecone, or Qdrant. A strong understanding of prompt optimization techniques, structured output enforcement, and automated evaluation methodologies is required, alongside experience with context management, embeddings, semantic retrieval, and multi-model routing across OpenAI, Anthropic, Gemini, and open-source models. Fluent written and verbal English communication with high attention to detail completes the profile. No prior astrology knowledge is required.

These qualifications map directly onto the chart mechanics the role protects. Structured output enforcement means a Zod or JSON Schema check confirms that every planetary degree, house placement, and aspect type is present and correctly typed before a reading ships. Semantic retrieval experience means knowing how to chunk interpretive material so a query about Saturn transiting the 10th house returns passages on career mastery and earned authority rather than generic Saturn text. Multi-model routing experience means recognizing that a Capricorn Midheaven with Saturn in the 10th demands a stronger model than a routine daily summary, and that Jupiter in the 10th should retrieve material on expansive public visibility rather than slow-built authority.

Working Environment and Infrastructure

The position is full-time and 100 percent remote, open to candidates worldwide with flexible working hours, sitting within the AI & Engineering department. The team works with an ensemble of modern LLMs — Anthropic Claude, OpenAI, and Google Gemini, alongside open-source models — coordinated with Next.js, TypeScript, Supabase, and vector embeddings. Strong engineering instincts, curiosity about data structures, and proficiency in LLM mechanics matter most; domain background in astrology is not expected, and the platform's own calculation output supplies the astrological precision the prompts must respect.

• 2+ years building production-grade LLM applications, prompt engineering workflows, or AI agent systems
• Proficiency in TypeScript, Python, or Node.js with modern AI SDKs and vector databases (pgvector, Pinecone, or Qdrant)
• Strong grasp of prompt optimization, structured output enforcement, and automated evaluation methodologies
• Experience with context management, embeddings, semantic retrieval, and multi-model routing
• Fluent written and verbal English with high attention to detail

How to Apply for the AI Prompt & Data Engineer Position

To apply for the AI Prompt & Data Engineer position at YesAstrology, submit your application for the full-time, fully remote role in the AI & Engineering department, which is open to candidates worldwide with flexible working hours. Before you apply, confirm that you meet the core bar: 2+ years of professional experience building production-grade LLM applications, prompt engineering workflows, or AI agent systems, plus demonstrated proficiency in TypeScript, Python, or Node.js with modern AI SDKs and vector databases such as pgvector, Pinecone, or Qdrant. No prior astrology knowledge is required — strong engineering instincts, curiosity about data structures, and proficiency in LLM mechanics matter most.

Your application should make the overlap between your experience and the role's technical surface explicit. If you have built retrieval-augmented generation pipelines, describe how you chunked interpretive material and handled embeddings, semantic retrieval, and context management. If you have enforced structured output, name the tooling — Zod or JSON Schema — and explain how you validated fields such as planetary degrees, house placements, and aspect types before shipping. If you have tuned model routing across OpenAI, Anthropic, Gemini, or open-source models, quantify the effect on latency and token expenditure. Concrete evidence of automated evaluation harnesses measuring factual accuracy, schema adherence, and latency carries more weight than general claims about prompt craft.

Because the role protects precise chart mechanics, a short note on how you would keep a generated reading faithful to computed data helps your application stand out. For instance, explain how you would prevent a response from blending Placidus house cusps with Whole Sign placements, or how you would instruct a model to flag a solar-noon chart default when an exact birth time is missing rather than asserting a precise 10th-house degree. Demonstrating that you think in terms of structured payloads — a Midheaven degree, an aspect orb, a house system declaration — signals the engineering rigor the position demands. To understand the product context your prompts would feed, review the platform's astrology blog before you write your cover note.

What to Include in Your Application

• A clear statement of your years of experience building production-grade LLM applications, prompt engineering workflows, or AI agent systems.
• Your primary language and stack — TypeScript, Python, or Node.js — with the AI SDKs and vector databases (pgvector, Pinecone, or Qdrant) you have used in production.
• Specific examples of prompt optimization, structured output enforcement, and automated evaluation methodologies you have implemented.
• Evidence of context management, embeddings, semantic retrieval, and multi-model routing experience across OpenAI, Anthropic, Gemini, or open-source models.
• Fluent written and verbal English communication with high attention to detail, reflected in the clarity of your written materials.

Finally, treat your application itself as a demonstration of the attention to detail the role requires. The position sits within a team coordinating an ensemble of modern LLMs — Anthropic Claude, OpenAI, and Google Gemini, alongside open-source models — with Next.js, TypeScript, Supabase, and vector embeddings, so precision in how you describe your past work is a meaningful signal. State your experience honestly, name the systems you actually built, and explain the measurable outcomes. Candidates who can articulate how they would eliminate model hallucinations, tone inconsistencies, and latency bottlenecks across millions of monthly user requests will stand out most clearly.