Lab notes

Finance

Autonomous financial investigation platform

Explains why profitability moved, with an evidence trail an auditor can follow.

Reference platform: CI, docs, synthetic ERPAutonomy: advise

Read the full note

The challenge

Finance questions draw on ERPs, external filings and internal documents. An answer without provenance cannot be audited or signed off.

The solution

An investigation platform in which code, not the prompt, controls the process: planner → investigation loop → adversarial critic → remediation loop → deterministic verifier.

Controls and evidence

• Every fact is stored with its source. Every claim is classified (fact, calculation, assumption, scenario or inference) and lists the evidence for and against it.
• Database access is read-only and guarded. State, memory, recovery and permissions are explicit policy.
• A configuration-driven evaluation suite scores the agent before any demonstration.

Data sources

Guarded database queries, public regulatory filings, public macroeconomic series, internal documents, anomaly detection and a scenario engine.

Status

Full platform with continuous integration, documentation and a demonstration script. It runs on a labelled synthetic ERP, so it can be evaluated without client data.

Autonomy level

Advise. The platform reads, analyses and reports, and a person acts on the findings.

Relevant to

Corporate finance, internal audit and risk functions.

Technology

Python 3.11, FastAPI, SQLAlchemy (Postgres/SQLite), Pydantic, OpenAI Responses API, Next.js, Tailwind, Docker Compose, GitHub Actions (ruff, mypy, pytest)

Supply chain, maintenance, field service

Operations automation agents

Six agents that investigate a business event, decide, act through controlled tools and escalate to a person when required.

Reference build: six agents, synthetic dataAutonomy: act with approval

• 6agents on one shared pipeline
• 7steps from event to verified action

Read the full note

The challenge

An agent that can commit spend or change records needs hard limits and a full audit trail before it can be connected to core systems.

The solution

Agents for supplier disruption, maintenance triage, field-service rescheduling, procurement exceptions (three-way match), inventory stockouts, and a cross-domain operations control tower.

Controls and evidence

• One shared pipeline: gather → assess → validate → decide → approval gate → execute → verify.
• Spend limits, approved vendors and tolerances are enforced in code. Writes are idempotent and verified by read-back. Every tool call is audited.
• Run state is persisted, so a human approval can resume the work later.
• The stockout agent survives a forced shutdown and replays safely.
• Each agent ships with deterministic evaluations and unit tests. Without a model API key it falls back to a deterministic heuristic.

Status

All data is synthetic. The tool layer is designed to map onto a production ERP or enterprise asset-management system.

Autonomy level

Act with approval. Consequential actions wait for a person.

Relevant to

Global manufacturing, utilities, procurement and field-service operations.

Technology

Python 3.12, LangGraph, MCP, FastAPI, Pydantic, SQLite (simulated ERP/EAM), TypeScript, Vite

Retail and e-commerce

Conversational commerce agent

A single conversation takes a customer from intent to a working payment link, in three languages.

Working applicationAutonomy: act with approval

Read the full note

The challenge

Gift purchasing involves search, comparison, delivery checks, payment and tracking. Each extra step between intent and payment is an opportunity to lose the customer.

The solution

A chat-first agent integrated with a retailer's catalogue and ordering systems through the Model Context Protocol. Built for an industry agent-building competition in 2026.

Capabilities

• Search, delivery check, cart, guest checkout with a live payment link, and order tracking, all within the conversation.
• Photo-based search, a bundle builder that works to a stated budget, side-by-side comparison, and personalised message previews in a non-Latin script.
• Three languages, including a non-Latin script and a mixed local-English style.
• Persistent memory of recipients, occasions, location, language and budget, with a local holiday calendar.
• Streaming responses with parallel tool calls, rendered as rich product cards and delivery quotes.

Resilience

A dependency-free integration client with synonym retry, browse fallback when search returns nothing, stale-on-error caching, jittered retries and session reconnect. Installable progressive web app that meets accessibility requirements (live regions, focus management, reduced-motion support).

Autonomy level

Act with approval. The customer reviews the cart and completes payment.

Relevant to

Retail, gifting and marketplace operators.

Technology

Next.js 14 (App Router), TypeScript, Tailwind, OpenAI function calling, Model Context Protocol, custom SSE protocol, Vercel

Advisory services

Knowledge-graph advisory assistant

An assistant that answers from a knowledge graph first, and states clearly when the graph holds no answer.

Client engagementAutonomy: advise

Read the full note

The challenge

Interviews and case notes hold an organisation's expertise, but the material cannot be searched. Plain vector search misses how people, companies and events relate.

The solution

Two connected components for an advisory product.

• The assistant: Answers questions from a Neo4j graph through a Cypher query tool, falls back to web search when the graph has no answer, and summarises long conversations to manage memory. Runs with Qdrant indexing and a Redis/Postgres-backed LangGraph server.
• The extractor: Converts raw interview transcripts into a graph of people, companies, regions and case stories, with typed output ready to load.

Related build

An incident-knowledge assistant in which a router selects vector search or generated graph queries for each question.

Status

Client engagement, delivered with colleagues. Client confidential.

Autonomy level

Advise. The assistant answers questions and does not change records.

Relevant to

Advisory, consulting and any firm with years of unstructured case material.

Technology

LangGraph, Neo4j, Qdrant, GPT-4o, Groq, Pydantic, Docker Compose

Life sciences

Ontology-grounded drug-interaction analysis

Drug-interaction answers that trace to a formal medical ontology rather than to a model's recollection.

Public demonstrationAutonomy: advise

Read the full note

The challenge

In a clinical setting, a fluent but incorrect answer is worse than no answer.

The solution

Four specialist agents (interaction, conflict, similarity and alternative medication), orchestrated with LangGraph. Each queries an OWL/RDF ontology. A language model then converts the structured results into a readable answer.

Controls and evidence

Facts come from the ontology, and the language model only phrases them. Every query is logged.

Status

Public demonstration with inference tracking, a web interface and a command-line tool.

Autonomy level

Advise. The system answers questions and does not act on them.

Relevant to

Pharma, clinical decision support and any domain governed by a formal rulebook.

Technology

Python, LangGraph, LangChain, RDFlib/OWL, Groq (open-source models), Streamlit, Hugging Face Spaces

Research-intensive teams

Technical document intelligence

Dense technical papers made easier to read, with the source text left intact.

Three iterations, end-to-end demonstration

Read the full note

The challenge

Specialists spend hours on terminology, equations and figures before reaching the substance.

The solution

Three iterations, from a research prototype to a full-stack application to a containerised retrieval stack.

• Parallel analysis agents cover terminology, figures (through vision models), mathematics and document structure, and report progress in real time.
• The latest version routes each request (question, glossary or figure) across a vector store, cache and relational database.

Evidence

Run end to end on a landmark machine-learning paper: sections parsed, figures described, glossary built and a quality report produced.

Relevant to

R&D, due-diligence and compliance teams with heavy technical reading.

Technology

FastAPI, LangGraph, Azure OpenAI (GPT-4o and embeddings), Next.js 14, KaTeX, Qdrant, Redis, Postgres, Docker Compose

Retail data

Retail price intelligence

A natural-language shopping list matched to the lowest-priced in-stock items across several supermarket chains.

Team engagement

Read the full note

The challenge

Retail data is inconsistent, changes daily and arrives in a different format from every source.

The solution

Our scope was the data collection and the retrieval and agent layers.

• Collectors for three regional supermarket chains write dated snapshots with error manifests and normalise unit prices and promotions into a single product model.
• A LangGraph agent parses the free-text list, runs vector search, reranks and summarises.
• Hybrid filtering on store, branch, maximum price and stock, using Firestore vector search with Gemini embeddings. An earlier Qdrant implementation is retained.

Controls and evidence

The model provider can be swapped (Groq, OpenAI, DeepSeek or Gemini), and the architecture is documented with UML and Mermaid diagrams.

Status

Team engagement.

Relevant to

Retail intelligence, pricing and marketplace teams.

Technology

Python, Playwright, Crawlee, LangGraph, Firestore, Qdrant, Streamlit, Expo/React Native

Emergency response, logistics

Autonomous response-planning simulators

Teams of autonomous drones learn to coordinate, and a live map interface lets stakeholders compare strategies before committing resources.

Research paper, open code

• 30scenes across 3 regimes
• 0learned policies beat the simple rule

Read the full note

The challenge

Emergency and logistics plans are costly to test in the field. A simulator lets decision-makers test them first.

The solution

Two connected platforms.

• Wildfire response: Drones search bushland for fires and suppress them. Policies trained with DQN, MAPPO and MADDPG are compared against a hand-written greedy expert across 30 scenes in three regimes, including a stress regime with eight fires, lossy radio and a binding battery constraint. Documented in a research paper in conference format.
• Flood relief: Drones deliver medicine and water to cut-off communities, scout roads and monitor slopes at risk of failure. Built on real terrain, road, population and river data from a recent major flood event. Runs on a standard CPU laptop.

Controls and evidence

• The flood model is calibrated against satellite-derived flood extents. Every calibration score and every failed acceptance gate is recorded.
• Fixed seed pools separate training, checkpoint selection and held-out reporting.
• Reported finding: A simple nearest-fire rule proved very difficult to beat, and no learned policy outperformed it. The paper lists its own limitations.

Applicable to

Logistics, emergency response, infrastructure monitoring and any problem of the form "compare strategies before committing capital". A deterministic simulation core, a streaming API and a map front end.

Technology

Python, NumPy, PyTorch, PettingZoo, FastAPI, WebSockets, React, MapLibre, Cesium, pytest

Agriculture

Field-operations and compliance platform

A collaborative farm-management application, taken from market research to a working product.

Full-stack MVP

Read the full note

The challenge

Farm records feed audits, safety rules and cost control, and fields often have no connectivity.

The solution

Market study → product requirements → technical architecture report → working application.

• Offline-first progressive web app with an idempotent sync queue, so a retry never creates a duplicate record.
• A plan → recommendation → work order → actual pipeline that validates application rates and warns automatically about pre-harvest and re-entry intervals.
• Audit evidence-pack PDF export, a cost-of-production ledger, five user roles and map views on PostGIS.

Relevant to

Agriculture and any field-operations business with compliance obligations.

Technology

Next.js 14, TypeScript, Prisma, PostgreSQL/PostGIS, Leaflet, turf, pdf-lib, jose

Subscription businesses

Churn prediction for data-sparse businesses

Identifies customers likely to leave, explains why, and keeps a record of every decision.

Demonstration on synthetic data

Read the full note

The challenge

Businesses with limited history and strict privacy obligations need predictions they can defend.

The solution

A clean-architecture churn system. Business rules run first and machine learning second. SHAP explains each prediction. A GDPR-style audit log records every decision. Served through an API and a command-line tool.

Status

Demonstration trained on synthetic data.

Relevant to

Subscription and membership businesses.

Technology

scikit-learn, SMOTE, SHAP, FastAPI, pytest

Agriculture, quality control

Explainable visual inspection

Finds, measures and explains defects on produce, and shows the reasoning behind each diagnosis.

Award-winning research

• 85.0%F1 on a public fruit-disease benchmark
• 4segmentation methods per defect

Read the full note

Recognition

Originated as an academic research project and became the basis of a Best Paper Award at an international AI conference in 2025.

The solution

Automatic background removal, then defect segmentation by four methods (colour ranges, K-means, adaptive thresholding and edge detection). For each defect: size, colour, centre, bounding box and share of total coverage.

Controls and evidence

A plant-disease ontology with a reasoner encodes symptoms, colour profiles and morphology, so every diagnosis arrives with its logic.

Result

85.0% F1 on a public fruit-disease benchmark, competitive with established deep-learning image models while remaining fully interpretable.

Also delivered

A web application with live parameter tuning and a real-time camera mode.

Relevant to

Agri-tech, food quality control and any inspection task where the reason matters as much as the result.

Technology

Python, OpenCV, rembg, scikit-learn, OWL, Pellet, Streamlit

Security and fraud

Multi-model phishing detection

Three independent models vote on whether a link is malicious, and the service continues if one drops out.

Benchmarked ensemble

• 97.1%LSTM accuracy on 62.5k held-out URLs
• 99.6%LSTM ROC AUC on the same set
• 92.0%ensemble accuracy, 1.71 s per URL

Read the full note

The challenge

Any single model has a blind spot. Three different model types cover for one another.

The solution

XGBoost on 20+ URL features plus TF-IDF, a bidirectional LSTM on character sequences, and Llama 3.1 (via Groq) for semantic analysis with reasoning. An orchestrator fuses their confidence-weighted scores. If the language model is unavailable, the remaining models are re-weighted automatically.

Results

• The LSTM, trained on 250k URLs, reached 97.1% accuracy, 98.3% precision, 95.8% recall and 99.6% ROC AUC on a 62.5k held-out set.
• The full ensemble reached 92.0% accuracy at 1.71 seconds per URL on a 600-sample benchmark.
• Benchmark scripts for each model and for the ensemble are included, so the figures can be reproduced.

Also delivered

A web interface with confidence indicators and per-model breakdowns, a command-line tool and a Python API.

Relevant to

Security, fraud and trust-and-safety teams.

Technology

Python, XGBoost, PyTorch, Groq/Llama 3.1, scikit-learn, Streamlit

Drug discovery, gene therapy

Molecular docking and binding-affinity research

Multi-engine docking pipelines and foundation-model research for drug discovery and gene therapy.

Proprietary engagement

Read the full note

Note

Proprietary engagement. Described at summary level only.

The challenge

Different docking engines produce different poses. Teams need a way to compare, rescore and validate them.

Scope

• Docking pipelines that combine several engines (AutoDock Vina, DiffDock and Boltz-2), with pose rescoring and validation.
• Foundation-model research on protein-ligand binding affinity and pose-confidence estimation.
• Molecular dynamics simulation and engineering of viral capsids for gene-therapy delivery, and quantum methods for molecular simulation.
• Agentic architectures with memory, planning and tool orchestration, plus research on agentic explainability.

Relevant to

Biotech, pharma and gene-therapy R&D.

Tools

AutoDock Vina, DiffDock, Boltz-2, GNINA, molecular dynamics, Qiskit, PennyLane

Cheminformatics

Drug-likeness modelling with validity review

A labelled dataset and classifier built from 17,967 compounds in a public medicinal-plant database, with an independent review of what the score means.

Research with independent validity review

• 0.999headline AUC, partly circular
• 0.945AUC from a single-feature rule

Read the full note

The solution

A pipeline that cross-references 142 approved-drug seeds across 5 public databases. Clustering and PCA outlier removal yield about 14.4k drug-like and 3.2k non-drug-like compounds. Fingerprint-based structural splits, SMOTE/ADASYN for class imbalance, and an XGBoost classifier.

The validity review

The headline AUC of about 0.999 is partly circular, because the labels derive from the same features. A single-feature rule already reaches 0.945 AUC, so the next phase requires verified negative examples.

Why it is included

It shows the validity check we apply to any model before calling it complete.

Relevant to

Cheminformatics and ML teams that want an independent review of a result.

Technology

Python, RDKit, pandas, scikit-learn, imbalanced-learn, XGBoost, ChEMBL API

Quantum computing

Hybrid quantum-classical compression

A hybrid compressor tested on real quantum hardware and under realistic noise.

Executed on real hardware

• 50image patches scored
• 6noise levels tested

Read the full note

The challenge

Claims about quantum advantage are easy to make and hard to verify. This work runs the circuits on real hardware and measures them against a classical baseline.

The solution

A Quantum Fourier Transform splits image patches from a standard benchmark into frequency bands. A variational quantum circuit, trained with PennyLane, learns which bands to keep.

Method

Evaluated on real quantum hardware (Qiskit Runtime) and under gate noise in a simulator, against a classical thresholding baseline. Scored with PSNR and SSIM over 50 patches at 6 noise levels.

Delivered as

Self-contained packages, single-file notebook versions, verification scripts and publication-ready figure generation.

Relevant to

Organisations assessing where quantum methods are, and are not, ready.

Technology

Qiskit, Qiskit-Aer, IBM Quantum Runtime, PennyLane, NumPy