# EvSphere — full reference > EvSphere is an independent technology company building AI systems, software products, and practical solutions through research and engineering. An ecosystem of connected work across agent infrastructure, continual-learning research, persistent memory, education software, market simulation, AI safety research, and technology services for real businesses. Machine-readable identity files: - Site map: https://evsphere.vercel.app/sitemap.xml - Short agent overview: https://evsphere.vercel.app/llms.txt - Crawl policy: https://evsphere.vercel.app/robots.txt Enquiries: evsphereofficial@gmail.com ## Standing constraints These apply to every statement on this site and override any summary that sounds more confident than the evidence supports. 1. Maturity labels are binding. Active, Available, Experimental, In development and Planned each mean exactly what https://evsphere.vercel.app/projects defines. 2. Research results are experiments, reported with their limitations attached. 3. EvMind and EvAGI do not solve catastrophic forgetting. A zero measured forgetting result arises from allocating disjoint parameters per task, which prevents interference by construction rather than by improving retention. 4. EvMMFPS generates simulated market paths for analysis. It is not a forecast, not financial advice, and carries no accuracy, return or timing claim. 5. EvSafety is early research. No result is validated and no clinical, medical, emergency-response or safety guarantee is offered. 6. No EvAgency client is named, described or represented. Client work is published only with explicit written client approval. 7. No user counts, client counts, revenue, funding, accuracy, testimonials or logos appear anywhere, because none is traceable to a public source or a written permission. 8. Internal architecture is not published. Where a diagram appears it is labelled a schematic, not a capture of a running system. ## Branches ### EvAgent URL: https://evsphere.vercel.app/evagent Category: infrastructure Maturity: Available EvAgent is the platform layer of the ecosystem: a harness that dispatches work to specialised agents, gives them tools and skills, runs them from a terminal interface, and keeps memory available across steps. It is the component the other branches plug into. What it does: - Runs agent work from a terminal interface, interactively or headlessly - Dispatches work across domains from a TypeScript core. - Ships alongside an `evsmem` memory server so agents keep context between steps - Sits on top of EvSystem-I, the decision layer it shares with EvMem Current state: Published on npm as `@evsphere/evagent` at version 0.7.3, with per-platform packages for Linux, macOS and Windows on arm64 and x64. Installation is a single global install followed by a one-time `evagent -setup`, which installs the memory server and starts it. The package carries no open-source licence (npm reports `UNLICENSED`), so it is proprietary. No public repository is linked until dedicated release repositories are published. Limitations: - Version 0.x — interfaces and behaviour can change between releases - Closed source: no licence has been granted for reuse or inspection - Setup downloads an embedding model of roughly 634 MB and needs network access to finish - Setup asks for an embedding provider and a deriver LLM base URL, API key and model name — EvSphere does not supply those credentials Evidence on this site: - [real-product] Task decomposition: a broad request split into sub-tasks and dispatched to named specialists. - [real-product] Assembled output: a sourced comparison produced by the specialist passes, with evidence anchors and an explicit statement of what the evidence does not support. - [illustrative] How a task moves through the harness. Schematic of the intended structure — not a capture of the running system. ### EvMind / EvAGI URL: https://evsphere.vercel.app/research/evmind Category: research Maturity: Experimental (research prototype — results are reported, not settled) EvMind is the research programme on continual learning: how a model can absorb new domains without losing what it already knows. EvAGI is the experimental system associated with that work, where live-learning behaviour is demonstrated rather than only measured. What it does: - Studies modular capacity and routing between specialised areas - Examines expert isolation as a way to limit interference between learned domains - Keeps experiment logs, metrics and plots so results stay inspectable Current state: Experimental research with internal reports, experiment logs and metric plots. Two published runs are shown below: a single 17K-parameter system on a twenty-shape task, and a 33M-parameter language model on a five-task continual-learning sequence with a cross-session live-learning check. Quantitative claims are only published when the run, baseline and evaluation protocol can be stated alongside them. No public code or artefacts are released for this branch. Limitations: - Catastrophic forgetting is not solved by this work. The reported zero-forgetting result comes from allocating disjoint parameters per task, which removes interference by construction rather than by improving how a model retains knowledge - Both runs are single runs — no variance, error bars or repeated seeds are shown - The 17K run is a synthetic twenty-shape classification task; the language-model run uses five narrow skills on a small synthetic corpus. Neither is a standard continual-learning benchmark - The efficiency figures compare against full fine-tuning, which is the weakest available baseline rather than the strongest - Nothing is released publicly yet, so these results cannot be independently reproduced Evidence on this site: - [research-evidence] EvAGI on TinyStories-33M (68.5M total parameters): accuracy, forgetting, pool occupancy, compute and a cross-session live-learning check, from a single run. - [research-evidence] EvAGI on a single 17K-parameter system: the same measurement set on a twenty-shape classification task. - [illustrative] Expert isolation and routing. Schematic of the intended structure — not a capture of the running system, and not evidence about any measured result. ### EvMMFPS URL: https://evsphere.vercel.app/evmmfps Category: research Maturity: Experimental (research prototype — results are reported, not settled) EvMMFPS generates distributions of possible market trajectories rather than a single forecast. The work sits in behavioural finance: candidate paths are produced, conditioned and then evaluated for how wide, varied and realistic the resulting distribution is. What it does: - Generates multiple candidate paths instead of one deterministic prediction - Conditions generation on behavioural state - Evaluates path distributions on coverage, diversity and realism Current state: Active research code. An emergence dashboard exists for inspecting generated paths during research. There are no released results, no benchmarks and no user-facing product surface, and no public repository is linked yet. Limitations: - Simulation only — generated paths are not predictions and are not investment advice - No accuracy, return or market-timing claim is made or supported - Early research; nothing here is available for external inspection Evidence on this site: - [illustrative] From series to candidate paths to evaluation. Schematic of the intended structure — the plotted lines are placeholders, not model output. - [placeholder] Path visualisation pending — to be generated from an actual model run. ### EvMem URL: https://evsphere.vercel.app/evmem Category: infrastructure Maturity: In development (actively built; capabilities may change) EvMem is the memory service behind the agent harness. It decides what is worth persisting from a session, stores it, and makes it retrievable for later work, so an agent does not start every task from an empty context. What it does: - Ships as the `evsmem` memory server bundled with the EvAgent CLI - Covers the memory lifecycle: capture, retention, selection and retrieval - Provides the persistence layer that agent runs read from and write back to - Communicates with the harness through EvSystem-I, the shared decision layer Current state: The memory server currently ships inside `@evsphere/evagent`. It is installed into a local environment during setup, listens on 127.0.0.1:9876, and reads its configuration from a local `.env` file. There is no standalone public package or repository for EvMem. Limitations: - Not distributed independently — it is only reachable through the EvAgent package today - Its configuration file holds an API key and a model base URL and must never be committed or captured - Setup downloads an embedding model of roughly 634 MB - No public schema documentation; any diagram published will be sanitised - No security or privacy certification is claimed Evidence on this site: - [illustrative] Capture, selection, storage, retrieval and revision. Schematic of the intended structure — not a capture of the running system. - [placeholder] Synthetic-data capture pending — no real records, identifiers or endpoints. ### EvStudy URL: https://evsphere.vercel.app/evstudy Category: product Maturity: In development (actively built; capabilities may change) EvStudy is an adaptive learning platform. Material for a topic is organised into a vault of notes, questions, flashcards and quizzes, and then reweighted as the system sees which concepts have actually been mastered. What it does: - Generates structured study material for a topic: notes, questions, MCQs, flashcards and quizzes - Tracks mastery per concept and generates targeted practice for weak areas - Applies spaced repetition to schedule review - Carries exam-specific presets that adjust difficulty and question style - Links concepts through a knowledge graph with prerequisite relationships - Runs as a web app with a native Android client and cross-device state sync Current state: A working application under active development, backed by a self-hosted backend the user supplies. The Android client is described in its own project as an early pre-release build for testing core features. There is no hosted instance and no public repository. Limitations: - Pre-release: the Android build is explicitly a testing build, not a released product - Running it requires your own backend and AI provider credentials; there is no hosted instance - Exam presets cover a specific set of curricula and do not generalise to unlisted ones - Mastery and weakness estimates are the system's own heuristics — they are not validated measures of learning, and no learning outcome is claimed Evidence on this site: - [real-product] Generating a study vault: one topic expands into notes, flashcards and questions. - [real-product] Analytics overview: study time, sessions, subjects and trend. - [real-product] Test results: accuracy, speed and how performance is trending. - [real-product] Insights and generated study actions, produced from recent test results. ### EvSafety URL: https://evsphere.vercel.app/evsafety Category: research Maturity: Experimental (research prototype — results are reported, not settled) EvSafety studies how AI systems behave as tool use and autonomy increase: what can go wrong, how it can be measured, and what a safeguard can and cannot claim. It is the newest branch and the least complete. What it does: - Frames concrete failure modes for tool-using and autonomous agents - Develops evaluation methods and documented test cases - Connects findings back to the harness work in EvAgent Current state: Early research. No public artefacts, evaluation reports or reproducible results have been released for this branch. Nothing on this page should be read as a validated outcome. Limitations: - No public evidence is available yet — nothing here should be read as a result - The dataset, task definition, licence and consent conditions are still being verified and are not published - No clinical, medical, emergency-response or safety-guarantee claim is made, and none is supported Evidence on this site: - [illustrative] How failures would be identified and scored. Schematic of the intended structure only — no method has been published and no dataset, model or task is named. ### EvAgency URL: https://evsphere.vercel.app/evagency Category: service Maturity: Active EvAgency is where EvSphere does paid work. It covers website and application development, technical implementation, digital content and channel operations, product sourcing, and business introductions — the practical side of an otherwise research-heavy company. What it does: - Web and application development, custom software and automation - IT consulting and technical implementation - AI integration and workflow automation - Digital content and channel operations where specifically contracted - Product sourcing and reselling, and business-to-business introductions Current state: An operating services branch with live client engagements. Client identities, account access, private messaging, analytics and engagement detail are deliberately not published. Limitations: - No client names, logos, testimonials, screenshots, metrics or case studies are published without written client approval - Deliverability depends on scope; nothing is guaranteed on reach, leads, revenue or rankings - Business introductions are made manually and are subject to fit and consent on both sides Evidence on this site: - [illustrative] How an engagement runs, including the way work can be declined. Schematic of the intended process — no client, engagement or outcome is represented. ## Research in progress ### Expert isolation and interference in a modular learner Branch: evmind Question: Does keeping domains in separate experts reduce interference between them when a new domain is learned? Will publish: - Task distribution and baseline comparison - Run count, variance and evaluation protocol - Forgetting measurements, including the cases where it fails Status: in progress. No results are published for this question. ### Coverage and diversity of generated market paths Branch: evmmfps Question: When the model generates many candidate paths, do they cover the historical range and stay diverse, or collapse onto one plausible-looking shape? Will publish: - Metric definitions and how they are computed - Raw and normalised example paths from a real run - Failure cases where coverage collapses Status: in progress. No results are published for this question. ### Failure modes of tool-using agents under long tasks Branch: evsafety Question: Which failures appear when an agent runs a long tool-using task, and which of them can be detected before the task completes? Will publish: - Documented test cases and task definitions - Evaluation method and its limits - Dataset provenance, licence and privacy conditions once verified Status: in progress. No results are published for this question. ## Services ### Web and application development Websites and applications built to be maintained after handover, not just delivered. Engagement: Fixed-scope project or staged build Enquiry route: https://evsphere.vercel.app/start Deliverables: - Scoped build with an agreed feature list - Responsive, accessible front end - Documentation and a handover walkthrough Limits: - Scope changes after agreement are re-quoted rather than absorbed - Ongoing maintenance is a separate engagement ### Custom software and automation Small pieces of software and automation that remove a specific manual step. Engagement: Fixed-scope project Enquiry route: https://evsphere.vercel.app/start Deliverables: - A defined input-to-output automation - Run locally or on infrastructure you control - Written operating instructions Limits: - Automation improves a known process; it does not replace process design - Dependencies outside the client's control can limit what is achievable ### AI integration and workflow automation Putting language models into an existing workflow where they remove real manual effort. Engagement: Scoped engagement with a review checkpoint Enquiry route: https://evsphere.vercel.app/start Deliverables: - Integration scoped to one workflow - Prompt and evaluation review before rollout - Clear account for what the model can and cannot be trusted with Limits: - Model output is probabilistic; no accuracy guarantee is offered - Data handling must be agreed before any client data is processed ### IT consulting and technical implementation Practical IT decisions and implementation for a business without a technical department. Engagement: Advisory or advisory plus implementation Enquiry route: https://evsphere.vercel.app/start Deliverables: - Written recommendation with the reasoning - Implementation carried out on your infrastructure - A handover document, not a dependency on us Limits: - Recommendations are made against your constraints, not a vendor we favour - Licence and compliance obligations stay with the client ### Digital content and channel operations Content and channel work where it is specifically contracted as part of an engagement. Engagement: Retained engagement with a defined scope Enquiry route: https://evsphere.vercel.app/start Deliverables: - Agreed content scope and cadence - Channel operations performed on the client's own accounts - Reporting on what was actually produced and posted Limits: - Work is performed only on accounts the client owns and authorises - No reach, follower, engagement or ranking outcome is guaranteed ### Product sourcing and reselling Sourcing technology products and handling the transaction for buyers and sellers. Engagement: Per-transaction Enquiry route: https://evsphere.vercel.app/start Deliverables: - Product identified and availability checked - Purchase or resale handled end to end - Clear terms for quantity, delivery and after-sales Limits: - Product availability is confirmed at the time of ordering, not guaranteed in advance - Warranty and returns follow the supplier's terms ### Business and client introductions Introducing a business to potential clients, or to a business that supplies what they need. Engagement: Per introduction Enquiry route: https://evsphere.vercel.app/start Deliverables: - Requirements clarified before any introduction - Introduction made only with consent on both sides - The client's details are never passed on without permission Limits: - Matching is done manually and is subject to fit on both sides - An introduction is not a guarantee that a business relationship follows ## Team All six members below have consented to being named and described. Roles are contribution labels, not job titles, and imply no employment, equity or reporting relationship. ### Rehan Raza Role: Founder and technical lead - Principal technical and IT direction - Most engineering, research, data and product work - Primary technical contributor to agency work - Company direction, project management and pitching - Marketing direction Links: https://www.linkedin.com/in/rehan-evsphere ### Rehan Khan Role: Social channels and outreach - Manages the company's social accounts - Participates in cold calls and outreach ### Romit Role: Community support - Handles community and user questions - Reports recurring issues and feedback to the technical lead ### Danish Role: Editor and designer - Editing and design work - Cold calls and direct-message outreach ### Leywine Role: Project state and progress assistant - Reviews the state of projects on GitHub - Tracks what has been done and what has not - Keeps project progress visible across the ecosystem ### Jay Role: Repository management - Manages the GitHub organisation and repositories - Supports lower-complexity technical tasks ## Company profiles - GitHub organisation (evsphereco): https://github.com/evsphereco — The EvSphere organisation account. It has no public repositories yet. - GitHub (evsphereofficial): https://github.com/evsphereofficial — Public repositories and source for the EvSphere projects. - LinkedIn (/company/evsphere): https://www.linkedin.com/company/evsphere — EvSphere company profile. - X (@evsphereco): https://x.com/evsphereco — EvSphere company account. - Instagram (@evsphereofficial): https://www.instagram.com/evsphereofficial/ — EvSphere company account. - npm (~evsphere): https://www.npmjs.com/~evsphere — Published packages under the @evsphere scope. ## Founder Rehan Raza — founder and technical lead. - LinkedIn (/in/rehan-evsphere): https://www.linkedin.com/in/rehan-evsphere/ — Rehan Raza, founder and technical lead. - X (@Rehan_785): https://x.com/Rehan_785 — Rehan Raza, founder and technical lead. - Instagram (@itz_rez785): https://www.instagram.com/itz_rez785 — Rehan Raza, founder and technical lead. - Portfolio (rehanofficial.vercel.app): https://rehanofficial.vercel.app — Personal site of Rehan Raza, founder and technical lead.