Location: New Haven, CT, on-site
Full-time
Reports to: Chief Technology Officer
Compensation: $200,000 - $260,000/yearly plus bonus
The Role
We are seeking a Principal Software Engineer to own the software through which Bexorg's scientists design experiments on living human brain tissue and interrogate the data those experiments produce. You will lead two flagship efforts: our custom LIMS, which supports two applications, Cortex for BrainEx and NeuroscaleOps for NeuroScale; and Neurolens, our AI research harness, which you will re-architect as a multi-tenant platform for large pharma partners.
This is a hands-on principal role reporting to the CTO. You will set the technical direction, write a good share of the code, and ship to users who sit a few meters away. The first users of everything you build are the surgeons, perfusionists, neuroscientists, and drug discovery scientists who run the platforms. Their experiment designs become the data model, and the data model becomes the evidence behind our therapeutics.
You will spend real time in the lab learning how an experiment on a human brain is actually designed: donors, protocols, dosing arms, sampling timepoints, regions, slices, plates, assays, readouts. You will turn that understanding into systems that make the right thing easy and mistakes hard, and that still make sense of the data years later.
You will build on a strong foundation rather than a blank page: a multi-account AWS estate, an S3-based datalake, a LIMS in daily production use, and a working AI harness with 200+ skills. The job is to take systems that already work for one company and make them precise enough for the FDA and general enough for our partners.
What You'll Build
-1. The LIMS: experiment design as the data model
Our custom LIMS supports two applications, Cortex for BrainEx and NeuroscaleOps for NeuroScale, on the backend data models beneath them. It is in daily production use. Its most pressing gap is also the most interesting software problem in the company: a first-rate experience, and a first-rate data model, for designing experiments.
An experiment on a donated human brain is designed once and run once. Its design determines what every downstream dataset means: which regions are sampled, which compounds are dosed at which concentrations and timepoints, which slices go to which plates, wells, and assays. Captured as free text, the design leaves the datalake full of results nobody can fully interpret. Captured as validated structure, every readout arrives already knowing what it measures.
You will build the loop that makes the second case the default:
• Scientists describe our systems and protocols in natural language.
• LLMs translate those descriptions into ontologies: the entities, relationships, and constraints of a BrainEx run or a NeuroScale slice campaign.
• The ontologies compile into schemas that validate the data structures representing each experiment, in the UI while the scientist designs it and in the backend as data lands.
• Ontologies, schemas, UI, and experiment records are versioned together, so the production system gains new capabilities continuously without orphaning a single historical run.
The UX is half the problem. The people designing these experiments are world experts with no patience for clumsy tools. The design surface must feel like it was built by someone who understands the biology, because it will be.
-2. Neurolens: a multi-tenant AI research harness for pharma
Neurolens is our custom AI meta-harness, deployed on AWS. Scientists ask research questions in natural language, and the harness answers by composing a bounded action space of 200+ skills and bioinformatics tools over our datalake. Every step is recorded so that any result can be reproduced exactly: the standard of provenance that FDA compliance demands.
Today Neurolens serves one organization over one datalake. You will re-architect it as a multi-tenant platform on AWS that becomes the basis for working with large pharma partners. At the center is a sophisticated permissions model that composes a custom datalake for each partner: every partner's view of our data is assembled from exactly the datasets, compounds, and tissue their individual licensing terms grant, and nothing else. In a second phase, partners will upload their own data and tools into that composed datalake at peer level with ours, so their scientists run Neurolens across both, with hard guarantees that no partner's data, skills, or provenance ever mixes with anyone else's.
That is a serious cloud engineering problem:
• A licensing-aware permissions model: each partner's datalake composed from the datasets, compounds, and tissue their agreement grants, enforced at the storage, query, and skill level, and re-composed as agreements change.
• Peer-level partner contributions, in the second phase: partner-uploaded datasets and tools that sit beside ours in the composed datalake with the same validation, provenance, and isolation guarantees.
• Tenant isolation that is designed rather than assumed: account and VPC boundaries, per-tenant IAM and KMS, data residency and egress controls, and a provenance ledger that is itself tenant-scoped and tamper-evident.
• Identity federation with partner identity providers, so a partner's scientists sign in with their own credentials and inherit exactly the entitlements their agreement defines.
• A governed registry for skills and tools: versioned, tested, and auditable, so that an analysis run this year can be re-executed identically in a regulatory review years from now.
• Scalable, cost-attributed compute for bioinformatics workloads over a large multimodal omics datalake: transcriptomics, proteomics, imaging, and physiological telemetry from the rigs.
• Observability and evaluation for agentic runs: what the harness did, why, at what cost, and whether it was right.
What You'll Do
LIMS and Experiment Design (Cortex & NeuroscaleOps)
• Own the end-to-end architecture of the LIMS and its two applications, Cortex and NeuroscaleOps: web front ends, APIs, backend data models, and the pipelines that connect them to the datalake.
• Sit with scientists, perfusionists, and robotics engineers to understand how BrainEx runs and NeuroScale campaigns are designed, then encode that understanding as ontologies and schemas rather than tribal knowledge.
• Design and ship the experiment-design UX: protocol templates, dosing and sampling plans, plate and slice layouts, validation at the point of entry, and clear diffs between versions of a design.
• Build and operate the LLM-assisted ontology pipeline and its versioning model: natural-language descriptions in, reviewed and versioned ontologies out, with human approval gates and regression tests that catch semantic drift.
AI Research Infrastructure (Neurolens)
• Lead the re-architecture of Neurolens into a multi-tenant AWS platform, and own the licensing-aware permissions model at its center: each partner's datalake composed from what their agreement grants, kept correct as agreements and data evolve.
• Extend the platform, in its second phase, so partners can contribute their own data and tools at peer level with ours, under the same validation, provenance, and isolation guarantees.
• Make provenance a product property: a governed skills and tools registry, immutable run records, deterministic re-execution, and exports a regulatory reviewer can follow without us in the room.
• Scale the execution layer with per-tenant, per-run cost attribution, and partner with our Head of AI on evaluation, so the harness is measured on correctness and reproducibility, not only on whether it produced an answer.
Platform, Cloud & Data Engineering
• Own production: infrastructure as code, CI/CD, observability, and operational practices appropriate for systems whose users are mid-experiment on irreplaceable tissue.
• Evolve the datalake's data models alongside the LIMS so that physiological telemetry, imaging, and omics data join cleanly on experiment design rather than on file names.
• Hold the security and compliance line across tenants: least privilege, encryption, audit logging, and the controls that pharma partners' security teams will scrutinize.
Leadership & Collaboration
• Serve as the senior technical authority on application and cloud architecture at Bexorg, setting standards for design, code review, testing, and documentation.
• Mentor engineers, and the scientists who increasingly build their own tools with LLMs, raising the floor of what the whole company can ship.
• Represent Bexorg's software platform to the technical and security teams of biopharma partners, foundation partners, and government grantors.
• Communicate clearly with technical and non-technical stakeholders, and write things down: design documents, decision records, and runbooks that outlast any one person.
Requirements
• Extensive hands-on experience (typically 10+ years) designing, building, and operating production software, including several years as the technical lead or architect of systems that demanding users depend on daily.
• Demonstrated track record of owning products end-to-end: discovery with users, data modeling, UX, implementation, deployment, operation, and iteration.
• Strong expertise in most of the following areas:
o Full-stack web application development: a modern front-end stack (e.g., React and TypeScript), API design, and backend services in Python or a comparable language.
o Data modeling for complex scientific or operational domains: relational design (PostgreSQL), schema evolution and migrations, and versioned or bitemporal records.
o Formal schemas and ontologies: JSON Schema, Pydantic, LinkML, OWL/SHACL, or equivalent, and the discipline of validating data at the boundary.
o LLM-powered systems in production: tool use and agents, structured outputs, bounded action spaces, evaluation harnesses, and a clear-eyed view of where models fail.
o AWS at depth: multi-account Organizations, IAM, VPC networking, KMS, S3-based datalakes, container and serverless compute, and infrastructure as code (CDK, Terraform, or CloudFormation).
o Authorization and entitlements: fine-grained, policy-based access control at the row, column, and object level (e.g., AWS Lake Formation, Cedar, or OPA) that must mirror contractual terms exactly.
o Provenance, auditability, and data integrity: immutable logs, deterministic re-execution, and controls in the spirit of 21 CFR Part 11, GxP, HIPAA, or SOC 2.
• Real interest in drug discovery, and the patience to learn experiment design in depth from the scientists who do it.
• Product sense: you notice when a tool is clumsy and fix it, and you can tell the difference between the feature scientists asked for and the one they need.
• Fluency with LLM-assisted engineering as a daily practice, paired with the judgment to know when the model is wrong.
• Excellent written and verbal communication, and demonstrated experience mentoring engineers and setting technical standards.
• Proven ability to operate as an aggressive self-starter: you take ownership, move quickly, and drive projects forward with minimal supervision.
• Bachelor's degree in Computer Science, Software Engineering, Applied Mathematics, or a related field, or equivalent industry experience with a strong track record of impact.
• Based in, or willing to relocate to, the New Haven, CT area for an on-site role.
Nice to Have
• Experience building or operating large multimodal omics datalakes: single-cell or single-nucleus transcriptomics, proteomics, metabolomics, or high-content imaging at scale, and the bioinformatics tooling around them (workflow engines such as Nextflow or Snakemake; scanpy, Seurat, or similar).
• Experience with LIMS, ELN, or sample-tracking systems, commercial (e.g., Benchling) or custom-built, and strong opinions about why scientists dislike most of them.
&bull
