Science · use case
An AI that lives in the lab, not at the vendor
“Every PhD that graduates takes three years of hard-won failure patterns with them. The next cohort rediscovers those failures at a cost we don’t measure but definitely pay.”
Scientific workloads are a bad match for the consumer-AI shape. Data is confidential or embargoed; the interesting archives are decades long; the instruments that generate the most useful data don’t speak HTTPS; and the papers you want to cross-reference are often behind paywalls or idiosyncratic APIs.
Eldric AI OS ships what a working research group actually needs, as one binary that runs on a workstation, a Pi, or a cluster login node. The Science Worker carries 140+ live APIs as plugins; the Matrix Memory layer gives a lab institutional recall that survives PhD-cohort turnover; the iot module speaks directly to the analytical instruments on the bench.
Value propositions
140+ scientific APIs as plugins
NASA, ESA, CERN, LIGO, USGS, PubMed, GWAS, ClinicalTrials.gov, materials, genomics, neuroscience. Each auto-surfaces as a toggle. Enable three, one query fans out.
Instrument integration
The iot module speaks OPC-UA, Modbus TCP/RTU, and MQTT Sparkplug B natively. Analytical instruments on the bench show up as data sources with no middleware in between.
Per-domain memory sizes
Matrix Memory defaults tuned per domain: particle_physics 512/1024, genomics 256/1024, seismic 256/768, robotics 128/512. No single-size-fits-nothing compromise.
Cross-cohort institutional recall
Dream cycle compresses the year’s notebooks into the lab’s memory. Next year’s team opens a chat and the archive answers. PhDs graduate; the memory stays.
Structured-paper retrieval
data.pageindex sketch outperforms vector similarity on textbook / FDA-filing / SEC-filing shaped documents. Reaches ~98.7% on FinanceBench upstream — same mechanism applies to regulatory scientific docs.
LIMS with audit trail
Science Worker ships sample tracking, experiment management, and audit trails with GLP / 21 CFR Part 11 templates (see labs & pharma).
AI-driven differentiator
A scientific corpus is bigger than any context window and structured in a way plain vector RAG mishandles. Eldric ships three retrieval primitives tuned for this: exact vector recall (papers), compressed associative recall via Matrix Memory (patterns across decades), and TOC-reasoning via data.pageindex (structured regulatory docs). The fan-out dispatches all three in parallel — the epistemics match the way a working scientist actually searches.
Scalable use cases
- Academic research groups. Per-PI tenant, per-project workgroup. Matrix Memory scoped to the group. One install on a workstation, or a share of the departmental cluster.
- Corporate R&D. Secure ingest of internal reports + lab archives. On-prem LLM means IP doesn’t leak to a vendor.
- Clinical research. IRB-approved perimeter. Patient-data
session.localposture for prompts touching PHI. Audit trail for study-conduct review. - Reproducibility initiatives. Matrix Memory captures what was tried and what didn’t work. The archive of negative results gets actually useful.
- Citizen science / amateur groups. Same stack runs on a Pi. Beekeepers, meteorologists, amateur astronomers get the 140 APIs on a €200 box.
Runs on commodity hardware
Eldric AI OS was built to land on small clusters, not on hyperscaler fleets. The whole stack is one binary; the on-prem LLM is embedded llama.cpp. The hardware plan that gets most organisations into production looks like this:
3× RTX 4090 — sweet spot
72 GB total VRAM with tensor-split. Llama 3.3 70B Q4 at 60–80 tok/s, a parallel 8B routing model, and an embedding server concurrently. One-time hardware cost ~€5–7k.
Single RTX 4090 / 4080 — team scale
24 GB. Llama 3.1 8B at 80+ tok/s, 13B comfortable, 32B Q4 possible. Enough for a small department chat with fan-out retrieval.
CPU-only — pilot scale
llama.cpp on 32+ core x86 runs 8B Q4 usefully. Matrix Memory is CPU-memory-bound. A refurbished server from the rack is enough to prove the architecture.
Scale up
Multi-node cluster with H100 / GH200 for research-grade workloads. Same binary, same role modules, topology-aware. See the HPC article.
Lab bench baseline
A single 4090 workstation runs the full stack — 8B LLM, embedding model, Matrix Memory, instrument ingest — comfortably for a 10-person research group. Scaled up, 3×4090 hits 70B at usable rates.
The arithmetic: a €6k workstation displaces a €30–60k-per-year SaaS-AI contract that still leaks IP, still can’t reach your mainframe, and still has a “we may use your data for training” clause hiding somewhere.
What the disk bill looks like
| Artefact | Size | Notes |
|---|---|---|
eldric-aios-5.0.0-3.alpha3.fc43.x86_64.rpm | ~1.4 MB | CPU baseline binary; one RPM, one systemd unit. |
eldric-aios-cuda add-on | ~512 MB | Pulled in automatically via Supplements: cuda-drivers on GPU hosts. Contains GGML_CUDA llama.cpp. |
| Llama 3.1 8B Q4_K_M GGUF | ~4.9 GB | Good default for team-scale chat on a single 4090. |
| Llama 3.3 70B Q4_K_M GGUF | ~40 GB | The sweet spot for 3×4090 tensor-split. Holds a 16k context comfortably. |
| Mixtral 8x22B Q4 GGUF | ~80 GB | Tight on 3×4090; comfortable on 4×4090 or 2×H100. |
| nomic-embed-text (embedding) | ~700 MB | CPU or GPU. One per cluster; handles vector indexing. |
Matrix Memory .emm per domain | 50–500 MB | Depends on rank × dim (see memory article). chat 64/768 ~200 kB; particle_physics 512/1024 ~500 MB. |
| Vector store per 1M chunks | ~6–10 GB | Depends on embedding dim. SQLite backend; FAISS optional. |
| Hash-chained audit log | ~200 MB / 1M calls | JSONL, append-only, rotation at 500 MB files by default. |
Three reference hardware setups
| Pilot / team | Department / BU | Production / enterprise | |
|---|---|---|---|
| CPU | 1× EPYC 7313 (16c) or i9-14900K | 2× EPYC 9354 (32c each) | 2× EPYC 9654 (96c) per node |
| GPU | 1× RTX 4090 (24 GB) | 3× RTX 4090 (72 GB) | 4× H100 (320 GB) or 8× H200 |
| RAM | 128 GB DDR5 | 256 GB DDR5 ECC | 1 TB DDR5 ECC per node |
| Storage | 2× 4 TB NVMe (RAID-1) | 6× 8 TB NVMe (RAID-10) + SSD cache | Tiered: NVMe hot + TB-scale HDD / Lustre |
| Network | 1 GbE OK | 10 GbE with link agg | 25/100 GbE or IB-HDR for multi-node |
| Power | ~1 kW typical / 1.5 kW peak | ~2 kW typical / 3 kW peak | 4–6 kW per node |
| Hardware cost | ~€4–5k | ~€12–15k | €80–250k per node |
| Serves | 8B model, 10–30 concurrent chat users | 70B Q4 at 60–80 tok/s, 200–500 users | Mixtral / Llama-405B, 2k+ users per node |
Network + ops footprint
- Ports. One outward port (443 at the edge). Internally: controller on 8880, data on 8892, inference on 8883, science on 8897, etc. — all behind the edge.
- Storage layout.
${ELDRIC_DATA_DIR}defaults to/data/eldricif writable, else/var/lib/eldric. Subdirs:models/,vectors/,memory/(matrix memory),storage/(file storage),agent/,edge/, and per-module dirs. - Backup. The audit log and
.emmfiles are the two artefacts that matter. Everything else regenerates. Snapshot the data dir nightly; off-site every week. - Updates.
dnf upgrade eldric-aios. Rollback isdnf downgrade. Zero vendor dance. - Ops team. A single systems engineer can run a pilot install. A team of two runs a department deployment. Production enterprise uses your existing Linux sysadmin rota.
SWOT — an honest read
Strengths
- 140+ scientific APIs shipped as plugins — immediate domain breadth
- Matrix Memory + vector + pageindex = three retrieval strategies in one fan-out
- OPC-UA / Modbus in-module — direct instrument integration without OPC-UA gateway software
- xLSTM distillation pipeline (Hochreiter lineage) available for custom model work
Weaknesses
- Some deep-science connectors (cryo-EM, specific mass-spec vendors) need custom extensions
- AlphaFold integration gated to Enterprise tier
- Not a replacement for domain-specific software (Schrodinger, Gaussian, VASP) — integrates, doesn’t subsume
- alpha.3 — needs ops support, not plug-and-play for non-technical teams
Opportunities
- Open-science + reproducibility push globally
- ERC + national-funder mandates on research-data management
- Reputation backlash against vendor AI training on leaked research data
- Long-tail data preservation funding (Horizon Europe, NIH, JSPS)
Threats
- Vendor LIMS (LabWare, STARLIMS, Labvantage) with bundled AI
- Cloud-scientific-computing offerings (AWS Science, GCP Science, DNAnexus)
- Consumer-AI tools leaking into researcher workflows despite IRB
- Preprint + publisher-AI integrations encroaching on the literature surface
First entry points — concrete value in 30 / 90 / 180 days
Group workstation pilot
Install on a PI’s workstation. Index the group’s NFS share. Turn on PubMed + arXiv plugins. Pilot with 3 users.
Matrix Memory seeded
Dream cycle ingests a year of notebooks. First ‘what did we learn’ query tested against group memory. Instrument ingest live for one instrument.
Departmental rollout
Multi-group deployment. Per-group tenants; shared campus KB plugin. Audit log reviewed by the RDM officer.