We build the prototype and everything above it.

What we do

Most consultancies work one layer. These connect end to end. The areas marked with a position sit at a point on the path from physical measurement to distributed system; those marked with a range run across it.

Each area opens to show work we have done. We describe it without naming who it was for.

• photons

  Spectroscopy and sensing

  Optical instrument design and the signal chain that makes an instrument useful. Shifted-excitation Raman difference spectroscopy, SERDS, designed and built from scratch. Fluorescence rejection, baseline and drift correction, chemometric model development, calibration transfer between instruments, and the practical business of getting a laboratory-grade measurement to survive field conditions.

  12 pieces of work

  • Common-mode rejection SERDS with asymmetric least squares baselines, recovering quantitative signal under large and varying fluorescence
  • Dual-laser instrument design with closed-loop power balancing, autofocus, auto-exposure, cosmic-ray rejection and on-board wavenumber calibration
  • Multi-illumination imaging used as an orthogonal chemical channel alongside the Raman signal
  • Etalon and detector-artefact diagnosis, and confocal versus shifted-excitation instrument selection
  • Differentiable spectral processing: unrolling iterative baseline correction into trainable layers with learnable smoothness and asymmetry
  • Loss functions derived from vibrational physics for spectral regression, respecting peak asymmetry and overtone structure
  • Riemannian metric learning over spectral manifolds, with geodesic distances and Grassmann subspace structure
  • Peak deconvolution, Richardson–Lucy and instrument-response deconvolution, wavelet and Savitzky–Golay denoising
  • Synchrotron soft X-ray spectroscopy: NEXAFS, XANES, STXM and ptychography
  • Trace-analyte Raman assays in food, feed and packaging matrices, and Raman imaging of biological tissue
  • Uncertainty estimation and conformal prediction for spectroscopic models
  • Digitising published spectra into machine-readable training sets

• electrons

  Computational chemistry and materials modelling

  Electronic structure calculation and spectral simulation: time-dependent DFT, NEXAFS spectral prediction and interpretation, plane-wave and real-space DFT with GPAW and ASE, deMon2k, Gaussian. Molecular dynamics with GROMACS. Structure–property prediction, and machine learning applied to molecular and materials data where first-principles calculation is too slow.

  12 pieces of work

  • Half-core-hole and transition-potential DFT for X-ray absorption spectra
  • Path-integral and ab initio molecular dynamics, and nuclear quantum and isotope effects
  • First-principles molecular dynamics for soil vibrational spectra
  • QM/MM, normal-mode and Hessian analysis
  • GPU-native electronic structure and machine-learned interatomic potentials, with differentiable programming in JAX and PyTorch
  • Molecular discovery pipelines using active learning and synergistic screening, closed against wet-lab validation
  • Bayesian optimisation over candidate selection in design–make–test loops
  • Generative small-molecule design, molecular docking, and ADMET, hERG and Tox21 safety screening
  • Graph neural networks and cheminformatics for structure–property prediction
  • Enzyme-inhibitor discovery for agricultural nitrogen-use efficiency
  • Covalent inhibitor design against a kinase target
  • Foundation models for spectral and chemical data

• parameters

  Applied AI and multi-agent systems

  Architecture and delivery of LLM-based systems that do real work: multi-agent orchestration with LangGraph, retrieval pipelines, tool-use and function-calling design, evaluation harnesses, and the unglamorous reliability work that separates a demo from production. Distributed fine-tuning of open-weight models across multi-GPU clusters. Inference and serving infrastructure, including mixture-of-experts models, throughput and latency tuning, and cost modelling.

  16 pieces of work

  • An autonomous discovery platform that plans its own experiments, generates and trains task-specific models, designs candidate molecules, runs simulations and triggers physical measurement
  • A distributed multi-agent network whose agents run as independently attested nodes, each with its own on-chain registration, economics and operator tooling
  • A personal chief-of-staff agent over mail, calendar and notes, with proactive behaviour driven by durable workflows: scheduled briefings, heartbeats, pattern digests and live-context resynthesis
  • Agent memory as a first-class subsystem: semantic, episodic and procedural tiers over vector search, a knowledge graph for entities and relationships, and an in-memory working set
  • Multi-agent orchestration with supervisor topologies and stateful checkpointed graphs, and bidirectional sync to external systems through an outbox with durable-workflow fallback
  • Voice agents using client-side tool execution, so no publicly reachable server is required
  • Distributed fine-tuning of open-weight models up to 480B parameters on 8×H200 clusters, using Megatron-LM, DeepSpeed ZeRO-3, FSDP and LoRA/QLoRA
  • Reinforcement-learning post-training with GRPO, colocated rollouts and reward-function design
  • Inference serving with vLLM, including dynamic LoRA adapter loading and mixture-of-experts throughput tuning
  • A merged contribution to vLLM adding LoRA support across the DeepSeek V2, V3 and R1 model family vllm-project/vllm #23971 (opens in a new tab)
  • Diagnosis of distributed-training failures: NCCL faults, mixture-of-experts deadlocks, cross-framework incompatibility
  • Framework bake-offs across MS-Swift, Megatron-LM, Axolotl and Unsloth, with each failure mode documented first-hand
  • Lattice-geometry research for model quantization, extending codebook constructions and applying basis reduction
  • Anti-hallucination validation: physical-limit gates, historical-baseline checks, statistical outlier detection, confidence scoring and human escalation
  • Enterprise semantic search over private corpora, with hybrid retrieval that mirrors source-system access control rather than flattening it
  • MCP servers for tool access and instrument control

• blocks

  Smart contracts and decentralized systems

  Smart contract design, development and review. Tokenization of real-world and physical assets. On-chain provenance and data attestation, particularly where the data originates from physical measurement, which is where the spectroscopy and blockchain sides of this company meet.

  16 pieces of work

  • A verifiable AI oracle running inside a trusted execution environment, with retrieval over a vector database and external data feeds
  • Attestation-gated infrastructure: disks unlocked only against a validated enclave quote, node admission gated on measurement registers and GPU trust data, signature-enforcing admission control
  • Zero-human-access confidential nodes with no shell, read-only root filesystems and secrets sealed to attested state
  • Confidential computing with AMD SEV-SNP and Intel TDX
  • Continuous attestation monitoring, with an edge verification service publishing measurements and quotes for third-party checking
  • Multi-cloud Kubernetes for confidential workloads
  • Solidity and Anchor contracts for permissionless node registration, staking, rewards and slashing, with reputation-weighted routing
  • Per-agent token issuance on a bonding curve, with on-chain revenue distribution to holders
  • On-chain prediction markets, and aggregation of responses across independent nodes
  • An operator dashboard for node economics and network health
  • Consensus over non-deterministic model output, via constrained schemas and on-chain response proofs
  • Permissioned-ledger provenance with Go chaincode: hashes on chain, encrypted payloads off chain, role-based isolation between competing parties
  • Zero-knowledge proofs, Bulletproofs and homomorphic encryption for verifying properties without revealing the underlying data
  • Token design across fractionalizable asset NFTs, soulbound records and fungible credits
  • Layer-1 design work, including hybrid consensus and cross-chain agent-to-agent protocols
  • Post-quantum attestation and risk assessment for signature-based chains

• photons → parameters

  Agricultural technology

  Soil sensing and the data infrastructure around it: sensor-to-model pipelines, agronomic data engineering, spatial data handling, and the design of measurement programmes that produce data a model can actually learn from.

  7 pieces of work

  • Sensor-to-model pipelines carrying a soil measurement through to an agronomic prediction
  • Conditioned Latin hypercube sampling designs over LiDAR-derived terrain covariates
  • Field-boundary delineation by ML inference over tiled satellite imagery
  • Integration across national crop, soil, elevation, hydrography, weather and satellite datasets, and farm-machinery OEM APIs
  • Soil metagenomics: assembly, gene prediction, functional and taxonomic annotation, and differential abundance
  • On-device plant disease detection by computer vision
  • Voice agents for agronomic workflows

• all layers

  Research and software engineering

  The connective work: Python systems, data pipelines, scientific computing, cloud and GPU infrastructure, API design, and translating research code into something maintainable.

  8 pieces of work

  • A spectroscopy data and MLOps platform: orchestrated import, quality-control, signal-processing and modelling flows, git-style data-lake versioning, experiment tracking and a model registry
  • A PostgreSQL warehouse over an instrument-native binary format — star schema, SCD-2 dimensions, range-partitioned facts, materialised views — and the ETL to unlock it
  • Administering a several-hundred-core, multi-terabyte-memory Slurm cluster and a managed Kubernetes estate, including tracing a GPU fault to a single accelerator instance
  • Model-promotion governance mapping analyte to model to sensor per environment, with version history, so deployment is auditable
  • Self-hostable stacks shipped as one compose file: API, Postgres with vector search, cache, graph store and workflow engine
  • One codebase to iOS, Android and web with React Native and Expo, including app-store release engineering
  • Self-hosted multi-GPU infrastructure on Rocky Linux, with private mesh networking and tunnelled ingress
  • Privacy-preserving inference over sensitive data using homomorphic encryption

Depth axis, from physical measurement to distributed abstraction.

How we work

The people who scope your work are the people who do it.

TensorChem sells deep technical work, not staff augmentation. Our practitioners deliver client work personally and carry their own client relationships, so the person who understands your problem is the person you keep talking to.

Who we are

The experience behind each layer of the stack, and where it comes from.

Doctoral-level computational chemistry

Our electronic-structure and spectral-simulation work rests on doctoral research in quantum chemistry, carried out at a national synchrotron light source. Beamline experience is the reason a spectrum gets read as physics rather than as a curve.

Instruments we have built, not specified

Our spectroscopy practice comes from inventing and building the SERDS soil-sensing instrument at an agricultural sensing company, and leading the software stack above it. That work is named on patents in agricultural technology. We have owned an optical signal chain end to end, from the excitation source to the calibration model.

LLM systems that do real work

Our multi-agent work includes a LangGraph-based system that won a national AI competition. We run distributed fine-tuning and inference infrastructure on multi-GPU clusters, which is a different discipline from calling a hosted API.

Research code that survives production

Python systems, data pipelines, scientific computing, cloud and GPU infrastructure, and API design. Most of the value in a research codebase is lost on the way to something maintainable, and closing that gap is a large part of what we are hired for.

Fields we work in

• Chemistry and materials science
• Agriculture and agricultural technology
• Distributed ledger and smart contract projects
• Agentic deep research
• Automated molecular discovery
• Agentic memory construction

Start a conversation

Tell us what you are working on. Enquiries are read and answered by the people who would do the work.