Postscript's $230K Engineer Jobs Aren't About Building Models. They're About Building Memory Systems That Decay on Schedule.

The $150M Context

Postscript hit $150 million in annual recurring revenue last September, up from $100 million six months earlier. The company projects another 50 percent year-over-year growth, profitably — a trajectory that would put it near $184 million ARR within a year. For any SaaS business that's a milestone. For a company that built its identity on a single channel, it creates a different problem: the ceiling is no longer technical. It's strategic.

Postscript is no longer an SMS tool with integrations. It is a retention platform where AI decides which channel carries which message to which shopper at which moment. That decision — cross-channel, real-time, brand-consistent — is the engineering problem the next section examines. The $150 million ARR bought the runway to solve it. The market growth made it unavoidable. The competitive response, particularly from Attentive, is already moving.

Founded in 2018 by Adam Turner and Alex Beller, Postscript rode the Shopify boom to become the default SMS layer for direct-to-consumer brands: Brooklinen, Ruggable, True Classic, Dr. Squatch among them. By 2026 it served 20,000 clients with 309 people across a fully remote U.S. and Canadian workforce. A Series C in June 2022 brought $65 million at a $636 million valuation led by Greylock. The Fondue acquisition eight months later added discount-strategy tooling. Through it all the product remained SMS-first, Shopify-deep, and intentionally narrow.

That narrowness worked. But the market underneath it shifted. The rich communication services market grew from $10 billion to $13.6 billion in a single year, a 35 percent compound annual growth rate. Carriers are rolling out RCS as the default messaging protocol on Android; Apple adopted it in iOS 18. The channel environment is fragmenting, and the brands on Postscript's platform are asking for email, RCS, and whatever comes next without losing the attribution and segmentation they built on SMS.

Postscript's answer arrived in September 2026: an email product built on a new AI-native platform, and the Agentic Context Platform (ACP) underneath it. The ACP unifies customer context (purchase history, onsite behavior, conversation transcripts) with brand context (voice, catalog, offers, rules) and performance data from every test running across channels. Email addresses, phone numbers, and Shopify records collapse into single customer profiles. Four attribution models and migration tools let brands move existing email programs without restarting their data history.

Building the Shared Context Layer

Postscript's bet on a unified AI stack rests on a single architectural decision: the Agentic Context Platform (ACP), a shared context layer that every product (SMS, RCS, email, Shopper, Infinity Testing) reads from and writes to. The company describes it as "built to support multichannel from the ground up, not as an extension to an existing channel." That distinction matters. Most retention platforms started with one channel, then bolted on email, push, or RCS as separate services with separate data stores. Postscript's co-founders say they considered that path and rejected it: "Every leading retention platform has the same origin story: one channel done well, then years of bolting more on. More channels, AI features, integrations all stacked on software from 2012. We had a choice: do the same, or start over and build for the AI era. We started over."

The ACP holds three context types. Customer context includes purchase history, onsite behavior, and, critically, what shoppers say in conversations with Postscript's AI agents. Brand context covers voice, catalog, offers, and rules. Performance data captures every test running across the system. Unified Profiles ensure a single subscriber record across channels: "SMS, RCS, and Email all use and read from the same subscriber context. No duplicate charges or fragmented customer records. Just one unified profile for every channel." A conversation over text informs the next email; an email click reshapes the next RCS message.

This sounds straightforward until you look at the data model. Traditional customer data platforms store static attributes: favorite color, birth date, hair type. Conversational data is different. As Postscript's team puts it: "From a tech side, it's a whole different type of data structure, right? It's not like hair type, it's like memories, right? Like memories about someone. And memories about people are temporal. They like have time decay." That phrase — "memories with time decay" — captures the core engineering problem. A shopper's intent expressed in an SMS conversation last Tuesday weighs differently than the same intent expressed last November. The system must weight, decay, and reconcile these memories across channels in real time, then feed them to AI agents making millions of micro-decisions per campaign.

The channel constraints compound the difficulty. SMS operates in 160-character fragments with strict carrier regulations. Email supports rich HTML, long-form content, and different deliverability rules. RCS, which carriers enabled roughly 18 months ago and which Postscript saw its first customer use for Black Friday, adds rich media, carousels, and suggested replies, but carries higher carrier fees. Each channel has different latency budgets, different formatting requirements, different compliance regimes. Yet the ACP must serve a single brand voice and a single customer memory to all three simultaneously. "Every channel draws from the same brand and customer context, so every send and its performance informs what happens next," the company says. That means the context layer cannot be a passive store; it must be an active, write-heavy loop where every interaction updates the profile that drives the next interaction.

Infinity Testing reveals the scale. Postscript reports an average 20 percent lift over 18 months, tested across hundreds of millions of messages. The system generates thousands of message variants, serves them to segmented audiences, measures performance, and feeds results back into the ACP automatically. "New context and insights flow between ACP and its agents in a constant loop; every data point feeds back in, so performance keeps improving on its own." That loop must operate across SMS, email, and RCS without creating feedback cycles that amplify noise or violate channel-specific constraints.

The integration layer adds another dimension. Postscript syncs bidirectionally with Klaviyo and unidirectionally from Shopify, Gorgias, and Recharge. An API exists for custom integrations. Competitor Attentive bundles email and push natively with 50-plus third-party connectors. Postscript's approach, featuring deep Shopify-native integration, Klaviyo sync, and an owned context layer, means the ACP must normalize incoming data from systems with different schemas, update frequencies, and identity graphs, then project a unified view back out to channels that each expect different payloads.

RCS adoption illustrates the operational reality. Early adopters report "really, really incredible" results but hesitate on carrier fees. The architecture must support gradual rollout: a brand might run SMS and RCS in parallel, with the ACP deciding which channel reaches which subscriber based on device capability, cost, and predicted conversion. That decision logic lives in the context layer, not in channel-specific code.

The result is a system that resembles a real-time bidding engine more than a marketing automation tool. Every subscriber interaction triggers dozens of small decisions (channel, timing, content, offer, variant) multiplied across 20,000-plus brands and millions of subscribers. The ACP's job is to make those decisions consistent, auditable, and improvable. That is the architecture Postscript's engineering team is building, and it is why the roles they are hiring for look different from a typical martech stack.

Shopper AI: One Agent, Every Channel

Three and a half years ago, Postscript opened a massive office in Phoenix. They called it the E-commerce Sales Center and staffed it with dozens of humans trained to respond to customer questions, handle orders, and manage interactions via SMS. The center delivered great customer experience. It also hit a hard ceiling: linear cost structure, impossible to scale. When Postscript wanted to offer conversational commerce to all 10,000 customers simultaneously, humans were no longer an option.

That constraint forced a migration, not a swap. Postscript built Shopper, a conversational commerce agent, to replace the Phoenix center. Early Shopper wasn't perfect; it misunderstood questions, gave wrong recommendations, missed sales opportunities. So they ran hybrid operations: Shopper handled simple conversations, humans took escalations and complex cases. Over time, as Shopper improved, the ratio shifted. Eventually the human center was no longer needed.

The Phoenix center gave Postscript something most AI startups never get: proprietary data. Every conversation had a human writing messages, making real-time decisions about how to handle questions, when to escalate, when to close. That created a dataset of thousands of human customer service interactions, all tagged with outcomes: customer satisfied, sale made, escalated. When Postscript built Shopper, they used that data to train the AI. As one analysis put it, the companies that win aren't the ones with the best models; they're the ones with proprietary data from actually running the business.

Shopper today is a suite of four specialized agents, each with a different job across the customer journey, all reading from and writing to the same Agentic Context Platform. Responses handles inbound conversations, answering questions, working through objections, sending product recommendations, closing sales in real time when a subscriber texts in. Automations decides what to do next for each subscriber, replacing welcome series, abandonment flows, and lifecycle messaging with a 1:1 message every time. Campaigns decides who should get a campaign and what it should say: fewer, better messages. Ask Shopper turns every conversation into something you can search and query, so brands can ask what customers are actually telling them and receive detailed reports.

The agents operate on the shared context layer holding three kinds of context: it (purchase history, onsite behavior, what shoppers say in conversations), it (voice, catalog, offers, rules), and it from every test running in an account. For every message, Shopper agents make dozens of micro-decisions based on what they know about the subscriber and overall program performance: when to send, what strategy to use, which offer to include, how to personalize every aspect. They use the same mechanisms as Infinity Testing to continuously learn and optimize; the longer it runs, the more it learns what strategies work for different cohorts.

Making this work across SMS and RCS, the two most inherently conversational channels, required solving edge cases that don't exist in single-channel deployments. RCS supports rich cards, carousels, suggested replies, typing indicators. SMS supports none of those. The agent has to render the same intent differently per channel without losing context or breaking compliance. Shopper inherits Postscript's TCPA compliance standards platform-wide: opt-outs, quiet hours, and consent rules apply automatically. For regulated categories like health and supplements, Shopper runs on built-in guardrails for medical claims, and individual use cases get reviewed before going live.

Balancing brand voice with conversion was its own engineering battle. The models would quickly learn that more aggressive sales tactics drove better performance, but brands are only interested in that up to a point. If the agent says "only three left, buy now" when that's not true, it drives sales and the brand gets mad. Postscript's team put prompt-level constraints in place to limit aggressiveness and require subtlety, plus batch validation gates that block messages containing off-brand content before they ever go live. Hallucinations got handled by a supervisory agent that cross-references product links and information against the actual database before sending. "Do you really see a lot of hallucinations? Not anymore," a Postscript engineer said in a recent technical walkthrough.

The results show up in customer data. Beekman 1802 saw a 32 percent revenue-per-subscriber lift, 18x incremental return, and a 29 percent lower unsubscribe rate. CX teams report Shopper handles roughly 92 percent of conversations on its own, escalating only topics that truly need help: cancellations, refund requests. It also works 24/7. Brands using conversation-driven flows in tandem with Shopper typically see a 5-to-15 percent lift in revenue.

All brands onboarding to Shopper start with Brand Centre and Shopper Responses. From there they expand into Shopper Automations and Campaigns, the latter still in alpha. Support for additional file types and data sources is on the way. Shopper is expanding across the customer journey, with future channels coming. The engineering bar this reveals isn't about model selection; it's about building the evaluation loops, validation gates, and context plumbing that let a specialized agent operate reliably across channels with different constraints, compliance regimes, and rendering capabilities. That's the stack Postscript is hiring for.

Six Roles, One Build Plan

Postscript lists six open roles on its careers page as of late September 2026, all remote and all with posted salary bands — a transparency practice the company underscores with a flat "we do not do geo based salaries" policy. The median posted salary sits at $130,000 across a $76,000-to-$188,000 range, and the tech stack reads like a modern cloud-native shop: TypeScript, React, Python, Go, Java, Scala, AWS, Kubernetes, Datadog, REST. With roughly 309 employees on staff, each hire represents a meaningful percentage of headcount, and the role mix tells a clear story about where the platform is investing.

The most senior engineering slot, Staff Platform Engineer at $190,000 to $230,000, targets AWS and Python expertise in a startup context. That title and band signal Postscript is building foundational infrastructure — likely the Agentic Context Platform that now underpins SMS, email, and RCS — rather than bolting features onto the old SMS stack. A Senior Backend Engineer role at $172,000 to $203,000 (posted 21 weeks ago and still open) asks for the same polyglot stack: TypeScript, React, Python, Go, plus eight additional technologies. The breadth suggests the backend team owns services that span channels, not siloed SMS microservices. A Senior Full-Stack Engineer at the same band (listed five months ago) rounds out the core product engineering layer.

Two customer-facing technical roles reveal the go-to-market shift. A Solutions Engineer at $120,000 to $140,000, posted five weeks ago, is explicitly charged with helping customers "extend their strategy into email alongside SMS, and contribute to shaping best practices as this is a newer, evolving part of the platform." The job description adds that the hire will "use AI tools to accelerate discovery calls, demo prep, and technical content creation — and help customers understand how Postscript's own AI features can drive their results." A Sales Engineer at the same band (one month old) carries a similar brief. Both roles exist because the product now requires technical translation across channels, merchants adopting email for the first time inside Postscript need architecture guidance, not just campaign tips.

The Technical Partnership Manager at $120,000 to $146,000, posted just one week ago, completes the picture. The role scopes and builds technology partnerships that "turn into demand rather than a logo on a page." In practice, that means integrating Postscript's cross-channel AI into the broader Shopify app ecosystem (review platforms, loyalty tools, subscription engines) so the ACP's shared context layer ingests signals from every touchpoint. The hire will also "ensure that customers' requests are well documented and prioritized," feeding product direction back to the platform team.

Together, the six roles form a coherent build plan: platform infrastructure to unify context, backend and full-stack engineers to ship channel-agnostic AI services, and two solutions engineers plus a partnerships lead to make the multi-channel product sellable and integrable. The salary bands are competitive for a fully remote, 300-person company backed by Greylock and Y Combinator, but they also reflect a team that cannot afford specialization for its own sake — every engineer hired now must operate across SMS, email, and RCS from day one.

The Talent War: Two Stacks, Two Tracks

The Postscript-Attentive rivalry has always been the defining axis of ecommerce SMS. In 2026 it became an AI arms race. A year ago the choice between them came down to list size and budget: Postscript was the scrappy Shopify-first option; Attentive was the enterprise-grade platform with a bigger footprint and a higher price tag. Then both companies made major AI bets in the same cycle, and those bets created genuinely different capabilities that now favor different types of brands, and different types of engineers.

Attentive's enterprise model — revenue-based pricing, dedicated strategists, managed creative support — demands engineers who can build and maintain predictive systems that operate inside a managed-service delivery motion. That means heavier investment in ML platform tooling, feature stores, and observability for models that non-technical strategists configure daily. Postscript's self-serve, Shopify-native model demands engineers who can ship AI features that work out of the box for lean DTC teams with no data science headcount. The Shopper AI agent is the sharpest example: it has to handle product-question understanding, order-status lookup, and recommendation generation in a single conversational loop, with latency low enough to feel like texting a human, and it has to do that across all three channels without the brand writing channel-specific logic.

The market is noticing. SMS message volume grew 40 percent year-over-year through 2025, outpacing every other marketing channel, and open rates sit at 90 to 98 percent with 90 to 95 percent of texts read within three minutes. U.S. retail ecommerce sales hit $1.3 trillion in 2025 per EMARKETER, and SMS has become one of the highest-converting channels inside that volume. Those numbers are pulling talent from adjacent fields, such as ad-tech real-time bidding, fraud detection, and recommendation systems at marketplace companies, into martech for the first time at scale.

But the switching costs are brutal. Migration between platforms risks suppression windows where send logic is temporarily inactive, compliance re-capture requirements that vary by how consent was originally documented, and flow rebuilds that cannot be automated because neither platform exports automation logic in a portable format. Brands commonly underestimate migration timelines by two to three times. TCPA violations carry statutory damages of $500 to $1,500 per message. The technical risk lives at the boundary between the SMS platform and everything around it: a Shopify checkout update that changes how consent is captured, a CRM that stores a different opt-out status than the SMS platform, a customer-service tool that doesn't know a contact texted "STOP" three days ago. Engineers who understand those integration seams, not just the model layer, are the ones both companies are competing for.

The talent market is bifurcating along the same lines as the product strategies. One track leads toward enterprise ML platforms that serve managed-service delivery — heavier on platform engineering, lighter on product-facing AI. The other leads toward self-serve AI products that work for Shopify brands with five-person marketing teams — heavier on product engineering, lighter on customization tooling. Both tracks pay well. Both are hiring. The engineers who choose between them are effectively choosing which version of the martech stack they want to build for the next decade.

The Engineering Playbook

Postscript's cross-channel build is a compressed version of what every AI-native product team will face: the same model must behave consistently across SMS's 160-character constraint, email's rich HTML canvas, and RCS's interactive cards — each with different latency budgets, compliance rules, and failure modes. The engineers who ship that consistency aren't "prompt engineers." They're systems engineers who treat LLM unpredictability as a design constraint.

The defining trait of AI software is unpredictability: "When you prompt an LLM, you don't know what you'll get back." Traditional software behaves deterministically; AI systems don't. Engineers who can't design around that variance don't ship cross-channel features; they ship demos.

The evaluation discipline separates production teams from prototype teams. Running disciplined evals and error-analysis loops is the core capability. At Postscript, that means measuring whether the Shopper AI agent recommends the same product bundle whether the shopper arrives via abandoned-cart text, browse-abandonment email, or RCS carousel — and catching drift before it hits 20,000 Shopify stores. Agent benchmarks on structured tasks still fail roughly one in three attempts. That gap is where evaluation engineering lives.

Context engineering has replaced prompt engineering. The modern skill is deciding what a model actually receives (retrieved documents, tool schemas, conversation history, explicit constraints) rather than wordsmithing the ask. The Model Context Protocol (MCP) is the piece that turns "AI that chats" into "AI that does things in your stack." Once a team understands MCP, "can our AI agent look up a customer's order status" stops being a multi-week integration project and becomes a well-scoped MCP server with a clear tool schema. Most production AI work now wires agents to internal tools, databases, or APIs, which is exactly what Postscript's Agentic Context Platform does across these contexts.

RAG is what lets a model answer accurately from your team's documentation, codebase, or internal knowledge base instead of guessing from training data. Fine-tuning is the least commonly needed stage; expensive, locking you into a specific model version, and most problems people reach for it to solve are actually retrieval or prompting problems in disguise. The sequence matters: prompting → agentic coding tools → MCP → agents → RAG → fine-tuning. Skipping stages is exactly why most teams plateau at autocomplete instead of shipping AI-native features.

Conversational AI engineering now blends two camps that used to be separate: classical intent classification with rigid flows, and newer LLM-first free-form generation. Modern conversational AI engineers blend both. Voice latency has dropped below the conversational threshold; sub-second end-to-end on real workloads is becoming standard, and multimodal conversation is arriving fast. Voice agents that can see what you see, look at documents you point a camera at, or watch your screen are early production deployments, not research. Outbound conversational AI is becoming a default sales motion. Regulated industries (healthcare, finance, legal) are starting to deploy conversational AI in customer-facing contexts. Engineers who can build for those environments, with audit, escalation, and regulatory accuracy, are in extreme demand.

The salary bands reflect the scarcity. U.S.-equivalent ranges for conversational AI engineers run roughly $110,000 to $150,000 entry, $150,000 to $220,000 mid, $220,000 to $310,000 senior, and $310,000 to $420,000-plus principal. AI job postings substantially outnumber qualified candidates. The engineers who solve these problems at Postscript, or at Attentive, or at the next wave of AI-native commerce platforms, aren't learning a niche. They're learning the default stack for the next decade of software.

The Phoenix sales center is gone. The humans who staffed it left behind a dataset that now powers an agent operating across three channels, 24/7, with memories that decay on schedule. The next hire will write the evaluation loop that catches the next drift before it reaches a brand's subscribers. That loop is the product now.

Working in AI? Zero G Talent tracks the openings: see every open Databricks role, browse AI jobs, openings at Anthropic, and the people building the field.