Best Content Recommendation Platforms for eLearning (2026)

Updated October 2026: Azure Personalizer retired on October 1, 2026; Pinecone, Vertex AI and Algolia prices re-checked; new open-source engines table with KerasRS and Gorse; EU AI Act high-risk date moved to December 2027.

More on this topic: read our complete guide — AI Video Analytics for Online Learning (2026).

Why Fora Soft wrote this playbook

Fora Soft has been shipping multimedia and AI software since 2005. Of our 250+ delivered products a substantial slice runs in EdTech, where the gap between “a recommender that works in a notebook” and “a recommender that ships measurable completion-rate uplift” is wider than most teams expect.

Our case studies in this space include BrainCert (a WebRTC virtual classroom and LMS that scaled to multi-million-dollar revenue), Scholarly (a test-prep platform with 15,000+ active users and live classes of up to 2,000 students), Tutrex (a white-label virtual classroom), InstaClass (live tutoring on WebRTC), and Career Point (AI career coaching). In products like these, the recommender layer is the difference between a course catalogue and a learning experience.

This guide is the playbook we hand to founders and product leaders evaluating which recommendation platform to put behind their LMS in 2026. It is opinionated, ranked by what holds up in production, and grounded in the trade-offs we use when scoping a build. The same thinking goes into the LMS and virtual classroom products our e-learning software development team ships every quarter.

Which content recommendation platform is best for eLearning?

Short answer (October 2026): for most LMS and EdTech products, the best content recommendation platform is Amazon Personalize or Google’s Vertex AI recommendations if you have no ML team, Algolia Recommend if learners mostly search, and RecBole or KerasRS on top of a vector database such as Qdrant if you have ML engineers and need prerequisite-aware sequencing. No single platform wins every moment of the learner flow, so the eLearning stacks we ship combine one recommender with a reranker.

The eight options below are ranked by what holds up in production for course catalogues, not by marketing reach. Native-advertising widgets such as Taboola and Outbrain also call themselves content recommendation platforms, but they push third-party articles for ad revenue and do not belong inside an LMS. For the wider map of content recommendation platform types for video and media, see our companion guide.

What “content recommender” actually means in eLearning

An eLearning recommender is not a Netflix clone. The reward signal is different, the constraints are different, and the failure mode is different. Get this taxonomy right and you save yourself the pain of optimizing for the wrong KPI.

1. Course discovery. A new learner lands on the platform and needs the next course to enroll in. Cold-start dominates. Content-based filtering on course metadata and learner profile signals is the fallback.

2. In-course sequencing. A learner is mid-course; the system has to pick the next module, video, or exercise. Pedagogical prerequisites matter. Sequence models (SASRec), knowledge-graph recommenders that encode prerequisites (KGAT, RippleNet) and contextual bandits do well here.

3. Adaptive review. Spaced repetition, knowledge tracing (DKT, BKT), and remedial-content recommendation. Classical recommenders fail here; you need a knowledge-tracing model in front of the recommender. Our Learn lesson on adaptive learning paths walks through that layer.

4. Search-time personalization. A learner queries “regression analysis”; the ranker should reflect their level, language, and history. This is where vector retrieval plus a learned reranker shines.

Figure 1. The four moments where an eLearning recommender earns its keep; each maps to a different model family.

A 60-second decision matrix

Ten platforms ago everyone wrote “evaluate Amazon Personalize, then build something custom.” In 2026 that is wrong half the time. Match your constraint to the family below and stop reading sales pages.

Platform 1: Amazon Personalize — the managed default

Amazon Personalize is the AWS-managed recommendation service: you push events, it trains models behind a REST API, and it returns ranked items. Built-in recipes cover user-personalization, similar-items, popularity, and contextual recommendations.

Why pick it

If your team has 1–2 backend engineers and zero ML headcount, Personalize gets you to a working recommender in weeks. It handles cold-start with metadata fallback, supports A/B testing, and scales without you touching infrastructure. On AWS’s October 2026 price list, the current v2 recipes cost $0.15 per 1,000 real-time recommendations with a 1 TPS minimum billed per campaign even when idle, plus $0.002 per 1,000 interactions for training and $0.05 per GB of ingested data. Budget around $600–900/month at 100K MAU for a typical eLearning workload. Personalize is not on AWS’s June 2026 list of services closing to new customers; Amazon Kendra is, from July 30, 2026. The full breakdown is in the cost model below.

Limits

Black-box: you can’t inspect or steer the model. Pedagogical sequencing (prerequisites, knowledge tracing) cannot be encoded directly. Vendor lock to AWS. If your eLearning product needs explainable recommendations for instructors, Personalize is the wrong layer.

Platform 2: Google Vertex AI Recommendations — the GCP equivalent

Google now sells its recommendation models inside Vertex AI Search for commerce, the successor to the standalone Recommendations AI console; the 2026 price list names them AI Commerce Search Recommendations. Models are exposed via REST and BigQuery integration, and importing catalogue and user-event data is free.

Why pick it

Strong if your warehouse is BigQuery and you want recommendations driven by SQL-defined event tables. List price in October 2026 is $0.27 per 1,000 predictions for the first 20M a month, $0.18 up to 300M and $0.10 above that, plus $2.50 per node-hour for training and tuning (Google Cloud pricing page). Optimization objectives such as click-through or conversion are set as configuration rather than custom code.

Limits

Slower setup than Personalize the first time you do it; documentation is thinner. Same vendor-lock risk. Requires Google Cloud-native event ingestion to get the most out of it. The product is modelled on retail catalogues, so expect to map courses to “products” and bend the schema a little.

Platform 3: Algolia Recommend and Coveo — search-meets-personalization

Algolia and Coveo come from the search world. Their recommendation modules are bolted onto an existing search index, which is exactly what most LMS products need: a learner searches, the index returns ranked results personalized to that learner.

1. Algolia Recommend. On Algolia’s October 2026 pricing page, the Free plan includes 5,000 recommend requests a month; the pay-as-you-go Grow and Grow Plus plans include 10,000 and then bill $0.60 per additional 1,000. Built-in “related products,” “trending items,” and “frequently bought together” recipes that map cleanly onto “related courses” and “trending lessons.” UI components ship for React, Vue, and vanilla JS.

2. Coveo. Enterprise tier; bundled with broader DXP capabilities; pricing typically $10K–$50K/month at enterprise event volumes. Strong if you need to merge content across multiple repositories (LMS + knowledge base + community).

Platform 4: TensorFlow Recommenders and KerasRS — Google’s two-tower toolkits

TensorFlow Recommenders (TFRS) is Google’s open-source two-tower retriever framework. The two-tower design — one tower for users, one for items, dot product or cosine in between, is the standard architecture for large-scale retrieval. TFRS is still maintained (version 0.7.7 reached PyPI in January 2026), but Google’s newer work goes into Keras Recommenders (KerasRS), a Keras 3 library that runs on JAX, TensorFlow or PyTorch. Google said in May 2025 that KerasRS helps power the Google Play feed.

Why pick it

Production-proven, Keras-friendly, scales to millions of items. Plays well with Vertex AI for serving and ScaNN for fast ANN. If you already have a TensorFlow stack, the marginal cost is low.

Limits

Requires ML engineering. You write your own loss functions, negative samplers, and evaluation pipelines. Cold-start handling is on you. PyTorch teams should start with KerasRS on its PyTorch backend, or with RecBole or Microsoft Recommenders.

Platform 5: RecBole and Microsoft Recommenders — the algorithm libraries

RecBole is the de-facto academic standard: a single PyTorch framework that implements 100+ recommendation algorithms behind a unified API — collaborative filtering, sequence models (SASRec, BERT4Rec, GRU4Rec), graph neural networks (LightGCN, NGCF), knowledge-graph based recommenders. Microsoft Recommenders covers a similar surface area with stronger production patterns and notebook walkthroughs, and has lived in the Linux Foundation’s LF AI & Data Foundation since October 2023. Both last shipped version 1.2.1 (RecBole in February 2025, Recommenders in December 2024), so treat them as stable rather than fast-moving.

Why pick them

If your eLearning product needs a non-standard recommender — knowledge-graph aware, sequence-based, session-based — you will not find it on Personalize or Vertex. RecBole already implements it. You inherit benchmarked baselines, evaluation suites, and a path to publish a paper if that matters to your team.

Limits

Research-grade by default. You wrap it in your own serving layer, monitoring, and feature store. Plan 2–4 ML engineers ongoing.

Platform 6: NVIDIA Merlin — production-scale, GPU-accelerated

NVIDIA Merlin is an end-to-end recommender stack: NVTabular for feature engineering on GPU, Merlin Models for training, Triton Inference Server for serving. It is what you reach for when your catalogue is millions of courses or videos and your event volume crosses 100M/day.

Why pick it

Training throughput on GPUs that off-the-shelf TensorFlow Recommenders cannot match. Best-in-class for large embedding tables (1B+ parameters). Open source under Apache 2.0, free.

Limits

GPU-required end to end. Steep learning curve. Overkill for any LMS under 1M MAU. Release cadence has slowed: the latest Merlin Models release on PyPI is 23.12.0 from January 2024, so check repository activity before you start a new build on it.

Open-source recommendation engines compared (2026)

Short answer: the open-source recommendation engines worth shortlisting in 2026 are KerasRS and TensorFlow Recommenders for two-tower retrieval, RecBole and Microsoft Recommenders for algorithm breadth, NVIDIA Merlin for GPU-scale pipelines, and Gorse when you want a ready-to-run recommendation server rather than a library. All six are free; the cost is the engineers who wire them into your LMS.

Versions from PyPI and GitHub release pages, checked October 16, 2026. Apache PredictionIO is left out: it moved to the Apache Attic in 2021.

Platform 7: Pinecone, Qdrant, Weaviate — the retrieval layer

In 2026, almost every modern recommender includes a vector database. You compute dense embeddings for users and items (with a two-tower model, an LLM, or a multimodal encoder), store them in a vector index, and retrieve the top-k candidates by approximate nearest neighbor in milliseconds.

1. Pinecone. Fully managed. Serverless Standard starts at a $50/month minimum, then usage-based: storage around $0.33/GB/month, reads $16 per 1M read units and writes $4 per 1M write units (Pinecone pricing page, October 2026; rates vary slightly by cloud and region). Pod-based clusters are now legacy. Lowest engineering load.

2. Qdrant. Open source, Rust core, strong filtering primitives. Self-hosted Docker for free; Qdrant Cloud from ~$50/month. Popular in production EdTech.

3. Weaviate. Open source with a hybrid search story (BM25 + vector) and built-in modules for OpenAI, Cohere, Hugging Face embeddings. Cloud and self-hosted options.

Platform 8: Cohere Rerank, OpenAI embeddings, GPT-as-ranker

The 2024–2026 wave is foundation models inside the recommendation pipeline. Three patterns matter:

1. Embedding generation. OpenAI text-embedding-3-large is $0.13 per 1M tokens in 2026; the lighter text-embedding-3-small is $0.02 per 1M. Cohere Embed 4 lists at $0.12 per 1M text tokens. Either is a good replacement for a custom item-tower when you have rich text content (course descriptions, transcripts).

2. Reranking. Cohere Rerank 4 Fast, the successor to Rerank 3.5, costs $2.00 per 1,000 search units in 2026, where one unit covers a query plus up to 100 document chunks of about 500 tokens; Rerank 4 Pro is $2.50. Take 100 candidates from your retrieval layer, ask Cohere or a small LLM to reorder them, return the top 10. Measure the gain with an online A/B test on course completion; offline ranking metrics alone tend to overstate it.

3. Generative recommendation. The frontier: ask the LLM to generate a recommended path (“here is the next module, here is why”) instead of returning a list. Useful for adaptive learning interfaces; expensive at scale.

Reference architecture: from learner event to recommendation

A production-grade eLearning recommender has six stages. The pattern is the same whether you are on a managed platform or a self-built stack; only the components change.

Figure 2. The six-stage retrieve-then-rerank pipeline; candidates narrow from a full catalogue to the top ten shown.

1. Event ingestion. Kafka, Kinesis, or Pub/Sub captures every learner interaction — video play/pause, quiz answer, time on page, course enrollment. Volume drives the rest of the pipeline.

2. Feature store. Tecton, Feast, or a BigQuery / Snowflake feature view. Maintains learner profiles, course embeddings, interaction histories. The single source of truth for online and offline.

3. Retrieval. Vector ANN (Pinecone / Qdrant / Weaviate) plus sparse lexical (Algolia / OpenSearch) returns 100–500 candidates from a 10K–1M course catalogue.

4. Ranking. A ranking model (gradient-boosted trees such as LambdaMART, DCN, or a sequence-aware ranker fed with SASRec features) scores the candidates. This is where most of your offline NDCG / recall@k gains live.

5. Reranking. Cohere Rerank 4 or a small LLM reorders the top 50–100 with cross-encoder attention. Optional but worth it on text-heavy content.

6. Exploration and serving. Contextual bandit or epsilon-greedy injects diversity. Redis caches results. A retrain job runs nightly or on a schedule keyed to fresh-content arrival.

For a deeper architectural read on the recommendation layer, see our companion guide on AI content recommendation systems for video, and our Learn primer on what eLearning video is.

The eight-platform comparison matrix

Side-by-side view across the dimensions that actually matter when scoping a build. Pricing is 2026 list pricing; treat it as a starting point for negotiation.

How much does a recommendation engine cost? An eLearning cost model

Short answer: at 100K monthly active learners, a recommendation engine costs roughly $600–900 a month on Amazon Personalize (about $750–1,100 on Vertex AI), $1,100–1,300 a month on a hosted retrieve-and-rerank stack, and about $900 a month in compute for a self-hosted build, plus the ML engineers who run it (2026 list prices).

A 100K MAU eLearning product with 5,000 courses is the line where the math flips between managed and self-hosted. Below is a back-of-envelope cost model we use in early scoping calls. Numbers are 2026 list prices; we typically come in lower in real estimates because Agent Engineering, our AI-agent-assisted delivery process, compresses the pipeline-build loop.

Figure 3. Infrastructure run-rate by path at 100K MAU (2026 list prices). Engineer headcount, not infra, drives the self-hosted total.

1. Managed path (Amazon Personalize). 100K MAU × 4 sessions / month × 5 recommendation calls per session = 2M requests / month. At $0.15 / 1K that is ~$300/month, but the 1 TPS minimum bills ~2.6M requests a month anyway, so plan on ~$390/month plus a few tens of dollars of training. Add $200–500/month in event pipeline and storage. Total: ~$600–900/month, no ML team needed. On Vertex AI the same 2M predictions cost ~$540 at $0.27 / 1K plus $2.50 per node-hour of training, so budget roughly $750–1,100/month there.

2. Hybrid path (Pinecone + Cohere Rerank). Pinecone serverless ~$50–200/month. Cohere Rerank is the swing factor: at $2.00 per 1,000 queries, reranking once per session across ~400K sessions/month runs ~$800/month. Embedding generation via OpenAI ~$30–50/month. Add a small inference VM (~$200/month). Total: ~$1,100–1,300/month, plus 1 backend engineer. Hosted reranking makes this path pricier than managed, not cheaper; self-host a cross-encoder reranker and the bill drops back toward the managed range.

3. Self-hosted path (TFRS / RecBole + Qdrant). Two GPU training runs / week on an L4 spot instance ~$400/month. Qdrant self-hosted on a $200/month VM. Inference on CPU ~$300/month. Total compute: ~$900/month. Plus 2 ML engineers ongoing — this is where the real cost lives.

4. Build cost. A working recommender MVP — ingest, retrieval, ranker, dashboard, A/B test rig, takes our team roughly 10–16 weeks with Agent Engineering. We usually estimate before kickoff in a 30-minute scoping call rather than online; ranges thrown without your data would be off by ~40% in either direction. If the recommender ships as part of a new learning platform, our custom LMS development team builds the event pipeline, course catalogue and recommendation surfaces in one scope.

Mini case: the learner data behind Scholarly’s 15,000 users

Scholarly is a Sydney test-preparation platform we built: 15,000+ active users and live classes of up to 2,000 students at once. It shows what an eLearning recommender feeds on, because one product already produces every signal in this guide: live-class attendance, quiz answers, homework submissions and reviews, video-on-demand views, and course enrolments in a built-in CRM.

The stack is Go and Node.js microservices on Kubernetes, with PostgreSQL for records and ClickHouse for analytics. That shape decides the recommender path. Event data already sits in a columnar store, so the cheapest first step is a content-based and popularity fallback computed straight from ClickHouse, then a managed or open-source ranker once interaction volume passes the ~100,000-interaction mark from the FAQ below. Teams that start with the model instead of the event pipeline usually end up rebuilding the pipeline first. Want the same review of your platform’s data?

A decision framework — pick your platform in five questions

Figure 4. Five questions that route you to a platform family, from managed services to a custom introspectable model.

Q1. Do you have an ML team? No — Personalize / Vertex / Algolia. Yes — consider TFRS, RecBole, or Merlin.

Q2. Do you need pedagogical sequencing or knowledge tracing? No — managed platforms are fine. Yes — you need RecBole-class graph or sequence models, or a custom DKT/BKT layer in front.

Q3. Is your dominant interaction search or browse? Search — Algolia / Coveo. Browse — Personalize / Vertex / custom retrieve-then-rerank.

Q4. How text-heavy is your content? Heavy (transcripts, articles, captions) — LLM embeddings and rerankers pay off fast. Light (mostly metadata) — classical recommenders are enough.

Q5. Do instructors need to explain recommendations? Yes — build on RecBole or knowledge graphs where you can introspect the path. No — black-box managed services work.

Five pitfalls that kill eLearning recommenders

1. Optimizing for clicks, not completion. The recommender that maximizes session length often surfaces shallow content. Optimize for course completion rate, exam pass rate, and time-to-mastery — the actual learner outcomes.

2. Cold-start by default. If you do not ship a content-based or popularity-based fallback for new learners, the first three weeks of every term will produce empty recommendations. Build the fallback first, then the personalized layer on top.

3. Offline-online mismatch. A recommender that wins on offline NDCG@10 often loses on online completion-rate uplift. Build an online A/B test rig from day one and trust it more than offline numbers.

4. Filter bubbles and popularity bias. Popular courses get more popular; niche specialist content disappears. Mitigate with diversity constraints, inverse-popularity weighting, and a small exploration budget (5–10% of impressions).

5. Ignoring compliance until launch. COPPA if your audience is K-12, GDPR for any EU learner, FERPA for US schools that receive Department of Education funding. Decide your data-handling posture before you decide on a platform, because you may not be allowed to send learner events to a third-party API at all.

KPIs to track from day one

Quality KPIs. Course completion rate uplift versus a static-list control (target ≥ 8% absolute on a 30-day window), recall@10 (target ≥ 0.35 on a held-out test set), NDCG@10 (target ≥ 0.18).

Business KPIs. Time-to-mastery reduction (target 15–25%), exam pass rate uplift, learner retention at week 4 and week 12, paid course conversion uplift.

Reliability KPIs. p95 recommendation latency (target < 200 ms end-to-end including reranker), retraining cadence honored (target nightly or weekly), input-distribution drift score (alarm if PSI > 0.2 between two weekly snapshots), exploration budget honored (5–10% of impressions).

Compliance: COPPA, GDPR, FERPA, EU AI Act

1. COPPA. Any platform aimed at learners under 13 in the US must obtain verifiable parental consent before collecting personal information. Behavior-based personalization is allowed, but third-party recommendation services may not be permitted to receive identifiable learner events. Audit your data flow before signing with Personalize / Vertex. The FTC’s amended COPPA Rule has been enforceable since April 22, 2026 and requires a written data-retention policy with a defined time period, so set retention for learner event logs before they feed a recommender.

2. GDPR. Article 6 requires a lawful basis (legitimate interest is the typical one; consent is safer). Profiling for educational outcomes is generally permitted but must be transparent and offer opt-out. Data subject access requests apply — you must be able to export and delete a learner’s recommendation history.

3. FERPA. Schools that receive U.S. Department of Education funding cannot share student education records with third parties without consent, except under narrow exceptions such as the school-official exception for contracted vendors. If your LMS is sold to schools, your recommender architecture must keep events on-premises or in a clearly contracted processor relationship.

4. EU AI Act. Annex III lists AI systems that evaluate learning outcomes, including when those evaluations steer the learning process, as high-risk, which can catch adaptive sequencing built on knowledge tracing. The Digital Omnibus, Regulation (EU) 2026/1744 (published July 24, 2026), moved those Annex III obligations from August 2, 2026 to December 2, 2027. Plan for a documented risk assessment, audit logging, and human oversight now.

When NOT to build a recommender

There are three situations where you should ship a curated list, not a recommender.

1. You have fewer than 200 courses or fewer than 1,000 active learners. A handcrafted curriculum and a curated “next up” list outperforms any algorithm at this scale. Recommenders earn their keep on long catalogues.

2. Your content is highly sequenced and prerequisite-heavy. A regulated certification course with a fixed module order does not benefit from a recommender. Use a knowledge-tracing model for adaptive review only.

3. You cannot commit to monthly retraining and an A/B test rig. A recommender without ongoing care drifts and starts producing worse results than a static list within a quarter. If you cannot dedicate the engineering, partner with someone who can — or ship a curated list and revisit later.

FAQ

Should I still consider Apache PredictionIO?

No. Apache PredictionIO was retired in 2020 and moved to the Apache Attic in 2021. Migrate any active deployment to RecBole, Microsoft Recommenders, or a managed service. Articles that still recommend PredictionIO are out of date.

Is Azure Personalizer still available?

No. Microsoft retired Azure AI Personalizer on October 1, 2026, and it accepted no new resources after 2023. Move contextual-bandit workloads to Vowpal Wabbit, the open-source library Personalizer was built on, or to a managed recommender such as Amazon Personalize or Vertex AI.

What are the best open-source recommendation engines in 2026?

KerasRS or TensorFlow Recommenders for two-tower retrieval, RecBole or Microsoft Recommenders for algorithm breadth, NVIDIA Merlin for GPU-scale pipelines, and Gorse if you want a self-hosted recommendation server with a REST API. All are free; budget for the engineers instead.

What is the cheapest way to ship a recommender for an early-stage LMS?

Algolia Recommend (the Free plan covers 5,000 requests a month; Grow includes 10,000, then $0.60 per 1,000) if your dominant interaction is search, or Amazon Personalize at a few hundred dollars/month if you are already on AWS. Both let you ship in a single sprint without an ML team. Move to a custom stack only when you have real data and a measurable ceiling on the managed approach.

Do I need a vector database to build a modern recommender?

If you generate dense embeddings (with a two-tower model, an LLM, or a multimodal encoder), yes. Pinecone is the lowest-effort managed option; Qdrant is the popular self-hosted choice. If you stay on classical collaborative filtering with low cardinality, you can skip the vector layer.

How much labeled data do I need before a deep recommender pays off?

Roughly 100,000 interactions and 1,000 active learners. Below that, classical baselines (Alternating Least Squares, popularity, content-based) usually win. Above that, two-tower retrievers and sequence models start to differentiate.

Can ChatGPT or Claude be your recommender?

As a final reranker on top of a retrieval layer, yes — that is what most modern stacks do. As the only ranker for thousands of items, no — latency and cost both blow up. Use the LLM where it earns its keep: cross-encoding the top 50–100 candidates, or generating a natural-language explanation of the recommendation.

How often should I retrain the model in production?

Daily for the retrieval layer, weekly for the ranker, monthly for any deeper feature engineering. Trigger out-of-cycle retrains if your input-distribution drift score (PSI on the user / item feature vector) crosses 0.2 between snapshots.

What KPI should I report to my CEO?

Course completion rate uplift versus a static-list control. It is a learner-outcome metric the business cares about, it’s simple to A/B test, and it correlates with paid retention. Offline NDCG and recall@10 are engineering metrics — share them in the team Slack, not in the board deck.

Where do graph neural networks fit in eLearning?

Anywhere prerequisites or shared topics matter. Knowledge-graph recommenders such as KGAT and RippleNet encode “course A is a prerequisite for course B” or “both courses share a topic” as graph edges, which classical collaborative filtering ignores; LightGCN and PinSage learn from the learner–course interaction graph instead. The trade-off is engineering effort — you need a maintained knowledge graph in your feature store.

What to read next

Ready to ship an eLearning recommender that lifts completion?

Pick by team shape, not by hype. No ML headcount — Amazon Personalize, Vertex AI Recommendations, or Algolia Recommend. ML team and a custom taxonomy — TFRS, RecBole, or Microsoft Recommenders. Massive catalogue — NVIDIA Merlin. Text-heavy content — retrieve-then-rerank with Pinecone and Cohere Rerank or an LLM. Every production eLearning recommender we have scoped has been a stack — not a single platform.

Plan from the start for cold-start fallback, online A/B testing, and a compliance posture that survives COPPA, GDPR, and the EU AI Act. Optimize for course completion uplift, not for clicks. If you want a partner that has built EdTech platforms such as BrainCert, Scholarly, Tutrex, and InstaClass — and ships with Agent Engineering for faster, cheaper estimates — that is exactly what we built Fora Soft to do. We also handle the broader AI integration work that surrounds the recommender layer.