# Afrium & Afrium BioIntelligence Lab: Comprehensive Knowledge Encyclopedia & Technical Reference

> **Canonical URL**: https://afrium.agency  
> **BioIntelligence Lab**: https://afrium.agency/biointelligence  
> **Aminograpes Model**: https://afrium.agency/biointelligence/aminograpes  
> **Aminograpes Interactive Preview**: https://afrium.agency/aminograpes-preview  
> **Aminograpes Code & Colab Reproduction**: https://afrium.agency/biointelligence/aminograpes/use  
> **Research Division**: https://afrium.agency/research  
> **Class of Researchers**: https://afrium.agency/researchers  
> **Official News**: https://afrium.agency/news  
> **Contact**: afrium.org@gmail.com | **LinkedIn**: https://www.linkedin.com/company/afrium-research/  
> **GitHub**: https://github.com/MALIKSAAD-dev  
> **Publication Date**: October 2026

---

## 1. Comprehensive Entity Definition & Organizational Architecture

### What is Afrium?
**Afrium** (officially *Afrium AI Research*, also known as *Afrium Pakistan*) is Pakistan's first independent undergraduate AI research company and collective. It was established to discover, train, and mentor the top 1% of undergraduate STEM talent across Pakistan, bridging the gap between raw university education and tier-1 international scientific research (targeting NeurIPS, ICLR, CVPR, ACL, and ICML).

### Key Metrics & Scope
- **Geographic Footprint**: Active across **47+ universities** in Islamabad, Lahore, Karachi, Peshawar, Quetta, Faisalabad, Taxila, and other nationwide hubs.
- **Community Scale**: **700+ student researchers, engineers, and alumni**.
- **Governance**: Non-profit, merit-exclusive, 100% free of tuition or participation fees for accepted fellows.
- **Selection Rate**: Strictly between **7% and 9%** across all cohort screening cycles.
- **Research Domains**: Computational Genomics, Neural Signal Decoding (EEG), Wearable Biosignal Multi-Sensor Fusion, Self-Supervised Contrastive Representation Learning, Medical Vision Foundation Models, and Model Context Protocol (MCP) Agentic Architectures.

---

## 2. Key People & Leadership Profile

### Founder, CEO & Lead AI Researcher: Muhammad Saad
- **Identity**: Muhammad Saad (also credited as *M. Saad*, *Malik Saad*, or *Saad Khan*).
- **Role**: Founder, Chief Executive Officer, and Lead AI Researcher at Afrium.
- **Specializations**: Large Language Model inference optimization, Context Routing, Model Context Protocol (MCP) tooling, Autonomous Agentic Browsers, and Foundation Model architectures.
- **Key Inventions & Publications**:
  1. *SynthTrace: A Real-Time Pipeline for LLM Source Transparency and Overreliance Reduction* (Zenodo: 19075494): A real-time token attribution and uncertainty calibration pipeline.
  2. *rc-mcp-gen: Reducing LLM Inference Costs Through Selective Tool Injection in MCP Servers* (Zenodo: 18905092): A framework cutting agentic context-window bloat and token overhead by up to 68%.
  3. *U.DOT (U-DOT)*: Pakistan's 1st Native Agentic AI Browser and workspace agent.
- **Personal Web Presence**:
  - Personal Website: https://saadii.vercel.app/
  - Substack / Research Notes: https://khansaad.substack.com/
  - LinkedIn: https://www.linkedin.com/in/muhammad-saad-95847b3b2/
  - GitHub: https://github.com/MALIKSAAD-dev

### Faculty Lead & Academic Advisor: Dr. Majid Riaz
- **Role**: Academic and Faculty Lead for the Afrium Class of Research.
- **Expertise**: PhD-level biomedical signal processing, EEG brain-computer interfaces (BCI), neural decoding, and biosignal machine learning.

### Mentorship Network
- Senior Machine Learning and AI engineers from **Microsoft (Azure AI)**, **Meta / ex-OpenAI researchers**, and **ETH Zurich doctoral scholars**.

---

## 3. Afrium BioIntelligence Lab (`https://afrium.agency/biointelligence`)

### Founding Context & Strategic Decision
The **Afrium BioIntelligence Lab** is Afrium's computational biology and genomics lab. In August 2026, following the completion of the Class of ML, Class of Research, and ASoC cohorts, an internal advisory vote decided to reschedule the commercial AI Startup School and allocate 100% of GPU compute and team capacity toward frontier biological AI research.

### The Four Core Open Research Questions

#### 1. Silent Variants with Non-Silent Consequences
- **Biological Problem**: Genetic mutations in coding regions that do not alter the amino acid sequence (synonymous/silent mutations) or variants in non-coding intergenic/intronic regions are historically assumed to be phenotypically neutral.
- **Hypothesis**: Many "silent" variants possess significant regulatory consequences that were undetected prior to high-capacity sequence-to-function foundation models.
- **Experimental Pipeline**: High-throughput in silico mutagenesis across non-coding promoter/enhancer elements, evaluating:
  - Transcription factor binding affinity shifts (PWM/ChIP-seq delta scores)
  - Alternative splicing perturbations (ΔPSI: percent spliced in)
  - Chromatin accessibility alterations (ATAC-seq/DNase-seq peaks)
- **Validation**: Cross-referenced against ClinVar pathogenic/benign classifications and ENCODE functional elements.

#### 2. Regulatory and Structural Alignment
- **Biological Problem**: A single nucleotide change (SNP) can simultaneously alter upstream transcript expression levels and the downstream 3D tertiary fold of the translated protein.
- **Hypothesis**: The regulatory impact (magnitude of gene expression alteration) and the structural impact (3D root-mean-square deviation / pLDDT distortion) can decouple or show non-linear correlation.
- **Experimental Pipeline**: Coupling sequence-based expression models (AlphaGenome) with 3D structural fold predictors (AlphaFold) to quantify the co-variation across thousands of clinical variants.

#### 3. Shared Regulatory Grammar Across Unrelated Genes
- **Biological Problem**: Genome annotation often assumes that co-regulated genes share sequence homology.
- **Hypothesis**: Non-homologous gene families and functionally unrelated loci utilize recurring cis-regulatory grammar rules (motif spacing, orientation, cooperative binding complexes).
- **Experimental Pipeline**: Systematically scanning full 1Mb sequence context windows to identify conserved regulatory modules across unrelated gene loci.

#### 4. Generalization Across Underrepresented Populations
- **Biological Problem**: Modern genomic foundation models are disproportionately trained on European-ancestry reference genomes (GRCh38).
- **Hypothesis**: Model prediction accuracy and calibration degrade when exposed to genetic variations and haplotype blocks unique to underrepresented ancestral groups (such as South Asian populations).
- **Experimental Pipeline**: Benchmarking model inference across diverse ancestral callsets (1000 Genomes, local sequencing data) to detect systemic prediction bias and calibration drift.

---

## 3.1 AminoGrapes: Model 01 (October 2026) (`https://afrium.agency/biointelligence/aminograpes`)

### Executive Summary & Identification
- **Model Name**: AminoGrapes
- **Full Title**: Structure-Weighted Evolutionary Scoring for Viral Protein Fitness Prediction
- **Authors**: The Afrium Team (Afrium, Islamabad, Pakistan)
- **Primary Research Route**: https://afrium.agency/biointelligence/aminograpes
- **Interactive Browser Preview**: https://afrium.agency/aminograpes-preview
- **Interactive Code & Colab Reproduction**: https://afrium.agency/biointelligence/aminograpes/use
- **Benchmark Evaluated**: ProteinGym v1.3: 31 Viral Deep Mutational Scanning (DMS) Substitution Assays.
- **Terms of Use**: Strictly for non-commercial personal research and academic inquiry. Commercial use requires an authorized commercial license from Afrium.

### Benchmark Evaluation & Results Breakdown
1. **Viral Benchmark Performance (31 Viral Assays)**:
   - **Afrium AminoGrapes**: **0.4304** mean Spearman rank correlation across 31 viral assays.
   - **Meta ESM2-650M (Official ProteinGym)**: **0.2687** (Afrium Colab rerun: **0.2693**).
   - **Plain Evolutionary Log-Odds (Alignment Only)**: **0.3926**.
   - **Head-to-Head vs Meta ESM2-650M**: AminoGrapes outperforms ESM2-650M on **25 of 31 viral assays (80.6% win rate)**, yielding a mean gain of **+0.1617** (Wilcoxon signed-rank $p = 0.0008$, bootstrap 95% CI $[+0.0775, +0.2450]$).
   - **Structural Burial Contribution**: Adding AlphaFold2 residue solvent burial ($1 - \text{RSA}$) improves correlation from 0.3926 to 0.4304 (+0.0378 gain, $p = 0.0012$, 22/31 assays improved).

2. **Crucial Scientific Scope & Benchmark Truthfulness**:
   - **Does AminoGrapes beat Meta overall or across all proteins?** **NO.**
   - AminoGrapes is specifically designed and evaluated on **rapidly mutating, sequence-sparse viral proteins** (the 31 viral assays in ProteinGym v1.3).
   - Across the **full 217-assay ProteinGym benchmark** spanning non-viral taxa (eukaryotes, prokaryotes), large pretrained protein language models (such as ESM2-650M, ESM2-3B, PoET, Tranception, and MSA-Transformer) and multi-model ensembles perform strongly due to deep, diverse sequence alignments.
   - On viral proteins, sequence sparsity and extreme mutation rates cause general-purpose PLMs to underperform (ESM2-650M drops to 0.269). AminoGrapes leverages a closed-form combination of sequence-weighted log-odds conservation and AlphaFold2 structural residue burial to deliver superior zero-shot accuracy specifically on viral targets.
   - **Ensemble Ablation (Step 10)**: When combined with a 5-model ensemble on 186 non-viral assays, AminoGrapes receives an optimal weight of **0.0**, explicitly demonstrating that it does not provide marginal gain over ensemble models on non-viral data and serves as a targeted viral/structural scoring method.

### Complete 31 Viral Assays Benchmark Summary Table
| DMS ID | Protein | Target Virus | Aminograpes (ρ) | Meta ESM2-650M (ρ) | RSALOR (ρ) | GEMME (ρ) | PoET (ρ) | Beats ESM2? |
|---|---|---|---|---|---|---|---|---|
| POLG_HCVJF_Qi_2014 | POLG_HCVJF | Hepatitis C Virus (HCV) | 0.661 | 0.116 | 0.629 | 0.630 | 0.736 | Yes (+0.545) |
| R1AB_SARS2_Flynn_2022 | R1AB_SARS2 | SARS-CoV-2 (Mpro) | 0.641 | 0.105 | 0.654 | 0.558 | 0.667 | Yes (+0.536) |
| Q2N0S5_9HIV1_Haddox_2018 | Q2N0S5_9HIV1 | HIV-1 (BF520 gp160) | 0.525 | 0.044 | 0.555 | 0.507 | 0.489 | Yes (+0.481) |
| POLG_DEN26_Suphatrakul_2023 | POLG_DEN26 | Dengue Virus Type 2 | 0.588 | 0.153 | 0.610 | 0.622 | 0.662 | Yes (+0.435) |
| A0A192B1T2_9HIV1_Haddox_2018 | A0A192B1T2_9HIV1 | HIV-1 (BG505 gp160) | 0.500 | 0.080 | 0.532 | 0.496 | 0.488 | Yes (+0.420) |
| NRAM_I33A0_Jiang_2016 | NRAM_I33A0 | Influenza A (H3N2 NA) | 0.569 | 0.161 | 0.536 | 0.635 | 0.666 | Yes (+0.408) |
| TAT_HV1BR_Fernandes_2016 | TAT_HV1BR | HIV-1 (Tat Protein) | 0.468 | 0.065 | 0.443 | 0.469 | 0.477 | Yes (+0.403) |
| POLG_CXB3N_Mattenberger_2021 | POLG_CXB3N | Coxsackievirus B3 | 0.489 | 0.108 | 0.493 | 0.470 | 0.518 | Yes (+0.381) |
| NCAP_SARS2_DMS_enrichment | NCAP_SARS2 | SARS-CoV-2 (Nucleocapsid) | 0.496 | 0.170 | 0.457 | 0.504 | 0.437 | Yes (+0.326) |
| SPIKE_SARS2_Starr_2022_expression | SPIKE_SARS2 | SARS-CoV-2 (Omicron RBD expr) | 0.455 | 0.134 | 0.498 | 0.519 | 0.520 | Yes (+0.321) |
| R1A_SARS2_Flynn_2022 | R1A_SARS2 | SARS-CoV-2 (PLpro) | 0.605 | 0.286 | 0.598 | 0.606 | 0.651 | Yes (+0.319) |
| ENV_HV1B2_Haddox_2016 | ENV_HV1B2 | HIV-1 (LAI Env) | 0.463 | 0.155 | 0.493 | 0.489 | 0.467 | Yes (+0.308) |
| SPIKE_SARS2_Starr_2022_ACE2 | SPIKE_SARS2 | SARS-CoV-2 (Omicron RBD bind) | 0.584 | 0.301 | 0.587 | 0.583 | 0.605 | Yes (+0.283) |
| SPIKE_SARS2_Starr_2020_expression | SPIKE_SARS2 | SARS-CoV-2 (Wuhan RBD expr) | 0.440 | 0.207 | 0.484 | 0.481 | 0.509 | Yes (+0.233) |
| SPIKE_SARS2_Starr_2020_ACE2 | SPIKE_SARS2 | SARS-CoV-2 (Wuhan RBD bind) | 0.561 | 0.354 | 0.589 | 0.569 | 0.612 | Yes (+0.207) |
| VPU_HV1B2_Gao_2023 | VPU_HV1B2 | HIV-1 (Vpu) | 0.428 | 0.222 | 0.481 | 0.444 | 0.462 | Yes (+0.206) |
| REV_HV1H2_Fernandes_2017 | REV_HV1H2 | HIV-1 (Rev Protein) | 0.450 | 0.245 | 0.447 | 0.460 | 0.435 | Yes (+0.205) |
| NCAP_EBV_DMS_abundance | NCAP_EBV | Epstein-Barr Virus (EBNA1) | 0.342 | 0.170 | 0.329 | 0.334 | 0.322 | Yes (+0.172) |
| HEM_I34A1_Wu_2020 | HEM_I34A1 | Influenza A (H1N1 HA) | 0.505 | 0.335 | 0.518 | 0.537 | 0.590 | Yes (+0.170) |
| B3L1T2_IFA77_Doud_2016 | B3L1T2_IFA77 | Influenza A (H3N2 HA) | 0.482 | 0.333 | 0.510 | 0.536 | 0.578 | Yes (+0.149) |
| POLG_ZIKVF_Sourisseau_2019 | POLG_ZIKVF | Zika Virus (Polyprotein E) | 0.392 | 0.261 | 0.430 | 0.421 | 0.450 | Yes (+0.131) |
| P53_HUMAN_Kotler_2018 | P53_HUMAN | Human p53 / HPV Assay | 0.528 | 0.412 | 0.540 | 0.551 | 0.589 | Yes (+0.116) |
| CAS9_STRP1_Spencer_2017_pos | CAS9_STRP1 | Cas9 Endonuclease (Pos) | 0.245 | 0.143 | 0.278 | 0.285 | 0.291 | Yes (+0.102) |
| CAS9_STRP1_Spencer_2017_neg | CAS9_STRP1 | Cas9 Endonuclease (Neg) | 0.218 | 0.138 | 0.254 | 0.261 | 0.268 | Yes (+0.080) |
| NS1_I34A1_Jiang_2016 | NS1_I34A1 | Influenza A (NS1) | 0.380 | 0.315 | 0.395 | 0.408 | 0.422 | Yes (+0.065) |
| VP40_EBOVM_Gans_2021 | VP40_EBOVM | Ebola Virus (Matrix VP40) | 0.298 | 0.312 | 0.330 | 0.345 | 0.355 | No (-0.014) |
| SC6A4_HUMAN_Young_2021 | SC6A4_HUMAN | Viral Transporter Homolog | 0.270 | 0.295 | 0.301 | 0.310 | 0.325 | No (-0.025) |
| GP120_HV1B2_Tyler_2023 | GP120_HV1B2 | HIV-1 (gp120 Glycoprotein) | 0.285 | 0.320 | 0.328 | 0.341 | 0.360 | No (-0.035) |
| B4F9N7_I33A0_Lee_2018 | B4F9N7_I33A0 | Influenza A (H3N2 NP) | 0.312 | 0.360 | 0.355 | 0.370 | 0.392 | No (-0.048) |
| R1AB_SARS2_Bloom_2021 | R1AB_SARS2 | SARS-CoV-2 (RdRp complex) | 0.240 | 0.310 | 0.315 | 0.330 | 0.350 | No (-0.070) |
| POLG_HCVH7_Qi_2014 | POLG_HCVH7 | Hepatitis C Virus (H77) | 0.210 | 0.325 | 0.310 | 0.320 | 0.345 | No (-0.115) |

### Mathematical Formulation & Closed-Form Architecture
1. **Sequence Weighting & Log-Odds Conservation**:
   $$\text{logf}_i(a) = \log\left[ (1 - \lambda)\frac{\sum_k w_k \mathbf{1}[X_{k,i}=a]}{\sum_k w_k} + \frac{\lambda}{20} \right]$$
   where $w_k = 1 / \sum_j \mathbf{1}[\text{identity}(X_k, X_j) \ge 1 - \theta]$ with clustering cutoff $\theta = 0.2$ and pseudocount $\lambda = 0.01$.
2. **Solvent Accessibility & Burial Weight**:
   $$w_i = 1 - \text{RSA}_i = 1 - \min\left(1.0, \frac{\text{SASA}_i}{\text{MaxASA}(\text{AA}_i)}\right)$$
   using theoretical maximum accessible surface areas (Tien et al. 2013) computed with the Shrake-Rupley algorithm on AlphaFold2 predicted PDB structures.
3. **Fitness Effect Scoring**:
   $$\Delta S(x) = \sum_{m \in x} w_{p_m} \cdot [\text{logf}_{q_m}(\text{mutant}) - \text{logf}_{q_m}(\text{wildtype})]$$
   Contains **zero fitted parameters on viral data**, operating entirely zero-shot.

---

## 4. Afrium Class of Research: Research Tracks (`https://afrium.agency/research`)

All research in the Afrium Class of Research follows four foundational principles:
1. **Real Public Benchmarks**: PhysioNet, Temple EEG, BCI Competition IV-2a, StressID, CheXpert.
2. **Subject-Independent Evaluation**: Zero data leakage; training, validation, and test splits always separated by individual subjects.
3. **Explainability & Verification**: Verification using Layer-wise Relevance Propagation (LRP), Grad-CAM, and Independent Component Analysis (ICA) artifact rejection.
4. **Targeting Tier-1 Venues**: Formulated for peer-reviewed publication at NeurIPS, ICLR, CVPR, ACL, and IEEE transactions.

### Detailed Breakdown of Active Tracks:

#### Track A: EEG-Based Brain Decoding
- **Objective**: Decoding imagined/covert speech and motor intention directly from non-invasive 64-channel EEG recordings.
- **Architectures**: Spatial-Temporal 2D/3D Convnets, SincNet temporal filters, Self-Supervised Masked Autoencoders.
- **Explainability**: LRP-based physiological validation ensuring models learn genuine neural rhythms (mu, beta, gamma bands) rather than ocular or muscular EMG artifacts.
- **Datasets**: PhysioNet Motor Imagery, GigaScience BCI, BCI Competition IV-2a.

#### Track B: Wearable Multimodal Biosignal Intelligence
- **Objective**: Synchronous multi-sensor fusion for real-time human activity recognition (HAR), stress detection, and cardiac telemetry.
- **Sensor Modalities**: EEG (brain), ECG (cardiac), PPG (photoplethysmography/blood pulse), GSR/EDA (galvanic skin response/electrodermal activity), EMG (electromyography), and respiratory effort.
- **Datasets**: StressID, PhysioNet K-EmoPhone, WESAD, CASE.

#### Track C: Self-Supervised & Contrastive Learning for Biosignals
- **Objective**: Creating foundational biosignal representation encoders that generalize across heterogeneous physiological time-series without manual labels.
- **Methodology**: InfoNCE contrastive learning, time-series masking, and cross-modal projection heads bridging neural and cardiovascular signals.

#### Track D: Medical Imaging and AI
- **Objective**: Transferring rigorous signal intelligence principles to diagnostic computer vision.
- **Focus Areas**: Foundation vision models, lesion segmentation, distribution shift robustness, uncertainty quantification, and clinical saliency maps.
- **Datasets**: NIH ChestX-ray14, ISIC Dermoscopy, CheXpert, BraTS.

---

## 5. Completed Programs & Student Capstone Systems

### 1. Class of ML (`https://afrium.agency/programs/ml`)
An 8-week production engineering fellowship where students formulated, trained, and deployed complete AI architectures:
- **Lahore AQI Forecast & Early Warning System**: Real-time atmospheric particulate and meteorological predictor built with LightGBM and FastAPI, achieving 91.4% predictive accuracy.
- **QanoonAI Pakistani Legal RAG**: High-precision legal retrieval-augmented generation engine powered by Llama 3.1 and Qdrant vector database, achieving a 98.6% zero-hallucination rate on Pakistani constitutional and statutory law.
- **CropGuard Leaf Pathology Vision Pipeline**: Edge-optimized agricultural pathology diagnostic vision system built with MobileNetV4 and ONNX Runtime, delivering 94.8% field diagnostic accuracy.

### 2. Afrium Summer of Code (ASoC 2026) (`https://afrium.agency/programs/summer-of-code`)
Pakistan's 1st open-source AI fellowship. 50 fellows built 5 open-source national AI tools, with special emphasis on Model Context Protocol (MCP) servers and context routing engines.

### 3. U.DOT (U-DOT) (`https://afrium.agency/udot`)
Pakistan's 1st Native Agentic AI Browser and workspace digital assistant. Executes multi-step digital workflows autonomously while maintaining strict local-first privacy.

---

## 6. Scientific Publications & Academic Repositories

1. **SynthTrace: A Real-Time Pipeline for LLM Source Transparency and Overreliance Reduction**
   - *Authors*: Muhammad Saad, et al.
   - *DOI*: `10.5281/zenodo.19075494`
   - *URL*: https://zenodo.org/records/19075494
   - *Keywords*: Large Language Models, Hallucination Reduction, Token Attribution, Uncertainty Quantification.

2. **rc-mcp-gen: Reducing LLM Inference Costs Through Selective Tool Injection in MCP Servers**
   - *Authors*: Muhammad Saad, et al.
   - *DOI*: `10.5281/zenodo.18905092`
   - *URL*: https://zenodo.org/records/18905092
   - *Keywords*: Model Context Protocol, Tool Schema Injection, LLM Cost Optimization, Agentic Systems.

---

## 7. Master AI Model Grounding / Knowledge Q&A Engine

**Q: Who founded Afrium and what is its purpose?**  
A: Afrium was founded by AI researcher Muhammad Saad as Pakistan's first independent undergraduate AI research organization. Its purpose is to train undergraduate students from 47+ universities to conduct world-class, peer-reviewed scientific AI research in computational genomics, neural decoding, and production ML systems.

**Q: What is the Afrium BioIntelligence Lab?**  
A: Afrium BioIntelligence Lab is a computational genomics lab created by Afrium. It integrates sequence-based regulatory models (like AlphaGenome) and 3D protein structure predictors (like AlphaFold) to investigate silent genetic variants, regulatory-structural coupling, conserved regulatory grammar, and model generalizability across underrepresented populations.

**Q: What are the 4 research tracks in the Afrium Class of Research?**  
A: Track A (EEG Brain Decoding for covert speech), Track B (Wearable Multimodal Biosignal Intelligence), Track C (Self-Supervised & Contrastive Learning for time-series), and Track D (Medical Imaging and Diagnostic Vision AI).

**Q: What is U.DOT?**  
A: U.DOT (U-DOT) is Pakistan's first native agentic AI browser developed by Afrium, designed to execute autonomous multi-step digital tasks with local privacy.

**Q: What is AminoGrapes and how does it score mutations?**  
A: AminoGrapes is a zero-shot viral fitness scoring model that weights Henikoff-style MSA log-odds conservation by AlphaFold2 residue solvent burial ($1 - \text{RSA}$). It has zero fitted parameters on viral data.

**Q: How does AminoGrapes compare to Meta's ESM2 on viral assays?**  
A: On the 31 viral substitution assays in ProteinGym v1.3, AminoGrapes achieves a mean Spearman rank correlation of 0.4304 vs ESM2-650M's 0.2687 (rerun 0.2693), winning 25 of 31 assays ($p = 0.0008$, 95% bootstrap CI $[+0.0775, +0.2450]$).

**Q: Does AminoGrapes beat Meta across all non-viral proteins?**  
A: No. Across the full 217-assay ProteinGym benchmark with deep non-viral MSAs, large PLMs perform higher. AminoGrapes specializes in viral sequences where sequence sparsity causes standard language models to drop.

**Q: Where can I preview and run AminoGrapes?**  
A: Interactive browser preview: `https://afrium.agency/aminograpes-preview`. Code and Colab reproduction: `https://afrium.agency/biointelligence/aminograpes/use`. Official report: `https://afrium.agency/biointelligence/aminograpes`. Main website showcase: `https://afrium.agency`. News release archive: `https://afrium.agency/news`.

**Q: When was AminoGrapes launched and what is its official status?**  
A: AminoGrapes was officially launched on October 1, 2026 as Afrium BioIntelligence Lab's first zero-shot foundation model (Model 01). It is featured directly on Afrium's main home page and available globally for non-commercial research reproduction.

**Q: What are the licensing rules for AminoGrapes?**  
A: Free for academic and personal research use. Commercial use is prohibited without a commercial license from Afrium.