ALEMBIC LABS

the lab

real-time view of the autonomous research apparatus. five AI agents work continuously, validated by Boltz-2 and Chai-1.

3d floorcoming soon
Preview render of the planned 3D laboratory floor — five lab-coated AI agents at workstationspreview render — not live

Soon the lab gets a fully interactive 3D floor — a real-time visualization of the research apparatus you see described below. Each AI agent will be embodied as a workstation: click any of them to watch their current task unfold, inspect the inputs they're reading, and read the reasoning they're emitting as it happens.

  • Researcher's desk: see the peptide registry pull, the modification hypothesis being drafted, the literature query being built.
  • Structural bay: watch Boltz-2 / Chai-1 returning the predicted complex into the 3D viewer in real time, with confidence metrics streaming onto an adjacent screen.
  • Communicator's terminal: follow the report being typed token-by-token, with each citation lighting up as it lands.

The image above is a pre-production mockup. Architecture, lighting and agent placement are final intent — geometry and animations will ship as the standalone scene at lab.alembic.bio.

01/

architecture

                ┌──────────────────────────────┐                │   DISTILLATION CYCLE         │                │   trigger every 45min        │                └─────────────┬────────────────┘    ┌───────────┬─────────────┼─────────────┬───────────┐    ▼           ▼             ▼             ▼           ▼┌─────────┐ ┌─────────┐  ┌─────────┐  ┌─────────┐ ┌─────────┐│RESEARCHR│ │LITERATUR│  │STRUCTRAL│  │CLINICAL │ │COMMUNICR││ ● ACTIVE│ │ ◯ idle  │  │ ◯ idle  │  │ ◯ idle  │ │ ◯ idle  │└────┬────┘ └────┬────┘  └────┬────┘  └────┬────┘ └────┬────┘     │           │            │            │           │     ▼           ▼            ▼            ▼           ▼┌─────────┐ ┌─────────┐  ┌─────────┐  ┌─────────┐ ┌─────────┐│PEPTIDE  │ │ PUBMED  │  │ BOLTZ-2 │  │ CHEMBL  │ │ POSTGRES││   DB    │ │ BIORXIV │  │ CHAI-1  │  │UNIPROT  │ │   API   │└─────────┘ └─────────┘  └─────────┘  └─────────┘ └─────────┘
02/

agents

researcher

Formulates hypotheses, designs peptide modifications, decides research direction.

modelclaude-opus-4-7
data sourcesPubMed, internal peptide DB
inputspeptide name, recent literature
outputshypothesis, modified sequence
statusactive
last actionselecting peptide
total runs247

literature

Reads scientific literature from PubMed and bioRxiv. Synthesizes relevant findings. Builds research context.

modelclaude-sonnet-4-6
data sourcesPubMed, bioRxiv
inputspeptide name, target protein
outputsabstract synthesis, key citations
statusidle
last action12 abstracts retrieved
total runs247

structural

Runs structure predictions, evaluates fold quality, cross-validates results.

modelclaude-opus-4-7 + boltz-2 + chai-1
data sourcesBioLM API
inputspeptide + target sequences
outputsPDB, pLDDT, pTM, ipTM, agreement
statusidle
last actionfold completed in 5m 12s
total runs246

clinical

Fetches biohacker/clinical context, ChEMBL bioactivity data, known binders, mechanism class.

modelclaude-sonnet-4-6
data sourcesChEMBL, UniProt, biohacker forums
inputspeptide name, target ID
outputsbinders, mechanism, dosage signal
statusidle
last action47 ChEMBL entries pulled
total runs246

communicator

Synthesizes all agent outputs into the 14-section detailed report: AI analysis, research brief, peptide profile, structural caption, executive summary.

modelclaude-sonnet-4-6
data sourcesinternal pipeline state
inputsfull distillation context
outputsmarkdown report, executive summary
statusidle
last actionreport drafted in 27s
total runs246
03/

the stack

reasoning layer
  • ├─claude-opus-4-7hypothesis generation, validation reasoning
  • └─claude-sonnet-4-6summarization, communication
structure prediction
  • ├─boltz-2primary structure prediction
  • └─chai-1cross-validation
knowledge
  • ├─pubmedprimary literature
  • ├─biorxivpreprints
  • ├─uniprottarget proteins
  • ├─chemblbioactivity data
  • └─internal peptide dbcurated performance peptides
infrastructure
  • ├─python + fastapiorchestration
  • ├─postgresqlexperiment storage
  • ├─replicate apiml inference
  • └─solana (planned)on-chain logging
04/

live metrics

cycles completed
83
total tokens used
2,102,892
folds generated
83
avg cycle time
11m 16s
avg tokens / cycle
25,336
lab uptime
3d
lab spend
$130.95
chai-1 runs
25 / 65 (38%)
adaptive gating active — runs only on borderline pLDDT
on-chain folds
69 / 73 (95%)
SHA-256 of fold core data · solana mainnet