Selected work

The platforms we build,
and the work behind them

Every capability we claim is backed by something we shipped. These are our own products first — architecture, models, interface and deployment — then the client engagements that sharpened the practice.

01 / Own platform

ATLAS‑LRO Nexus

A Living Research Operating System

Flagship platform

The problem

A research paper is a static description of work nobody else can re-run. Its claims cannot be queried, its methods cannot be executed, and its provenance is a log file — which is why reproducing published science is so expensive, and why regulators cannot simply trust it.

What we built

A platform that compiles a paper into a Living Research Object: queryable claims with mandatory citations, a causal graph for counterfactual reasoning, methods that execute in sandboxed containers, and a cryptographic ledger that makes provenance verifiable rather than asserted.

Why it matters

A paper becomes a callable function. You can ask what happens at a different dose, compose one paper's method with another's, and hand a regulator an audit trail that cannot be quietly edited.

Engineering highlights

  • Executable method runtime — pinned images, data and seeds, replayable byte-for-byte, with author-bounded parameter sweeps
  • Causal engine — do-calculus over a causal graph, not similarity search dressed up as reasoning
  • Cryptographic provenance ledger — tamper-evident and regulator-ready
  • Federated knowledge mesh — cross-publisher truth layer, controversy detection and an LRO exchange
  • Deployment-profile architecture — runs from a single-node pilot up to a full Kubernetes install; add services as you add capability
  • Compliance-aware — SOC 2 Type I control mapping, HIPAA technical safeguards, FDA/EMA/ICH narrative engine

Stack

  • 20+ microservices
  • FastAPI
  • Next.js
  • TypeScript
  • Temporal
  • PostgreSQL
  • pgvector
  • Neo4j
  • Kafka
  • MinIO
  • Docker
  • Helm
  • Kubernetes
  • Python / TS / CLI SDKs

Database-agnostic storage, provider-agnostic inference and pluggable extension points throughout — no component assumes a single vendor.

02 / Own platform

AgentOS

Enterprise agentic AI platform

Enterprise platform

The problem

Enterprises wanted agents that could actually complete work, but every option demanded a vector database, cloud embeddings and a rebuild for each new use case — none of which suits regulated or air-gapped environments.

What we built

A platform where agents plan, call tools and execute multi-step tasks, retrieval runs without any vector store, and every agent, prompt, model and branding decision is configured through the admin UI rather than in code.

Why it matters

It deploys where the data already lives. No embedding service to call, no per-tenant fork to maintain, and a white-label rebrand that takes one config block instead of a release.

Engineering highlights

  • Retrieval without lock-in — vector, hybrid or fully vectorless (BM25, FTS5, TF-IDF and a knowledge graph fused together), so it runs on whatever datastore you already have
  • Model- and provider-agnostic — one registry in front of every LLM, swappable without touching application code
  • Multi-agent orchestration — 10+ pre-built agents, plus custom agents authored in the UI
  • Voice — Whisper speech-to-text with Edge TTS or OpenAI output
  • SOLID module boundaries — loosely coupled, independently testable

Stack

  • Python 3.12
  • FastAPI
  • SQLAlchemy 2.0
  • Pydantic v2
  • Next.js 14
  • TypeScript
  • shadcn/ui
  • Tailwind
  • PostgreSQL 16
  • SQLite FTS5
  • Redis
  • LiteLLM
  • Whisper
  • NetworkX
  • Docker
  • GitHub Actions
03 / Own platform

Agrozent

Agricultural efficiency platform

Live in production

The problem

Agricultural supply chains lose value at every hop, and the people carrying the most risk — farmers — are paid last and have the least access to working capital.

What we built

One system joining prediction, traceability and settlement: yield and pest models, farm-to-fork provenance, tokenised real-world assets for financing, and instant payment.

Outcome

A 10–40% efficiency improvement, because a prediction is wired to a financial consequence instead of ending its life in a dashboard.

Engineering highlights

  • Layered architecture — Flask UI over a FastAPI service layer over the AI engine
  • Prediction suite — yield forecasting, pest detection, price prediction
  • Traceability — provenance recorded from farm to fork
  • RWA tokenisation — working capital unlocked against real assets

Stack

  • Flask
  • FastAPI
  • PostgreSQL
  • Alembic
  • Custom ML engine
  • Docker
  • Nginx
Visit agrozent.com
04 / Own platform

Haramayn Guide

Offline-first AI pilgrimage companion

Web + Android & iOS

The problem

Millions of pilgrims need step-by-step guidance in the one place where connectivity collapses — inside Masjid al-Haram, surrounded by hundreds of thousands of phones competing for the same network.

What we built

An app whose core works with the network off: ritual flows, context-aware duas, a grounded assistant, and sensor-based detection that counts Tawaf and Sa'i circuits without GPS.

Why it matters

Offline-first is an architectural commitment, not a cache. Knowledge, inference and state all had to work on-device, in six languages, on mid-range Android hardware.

Engineering highlights

  • On-device ML — accelerometer-based Tawaf/Sa'i circuit detection
  • Grounded AI — Quran and Hadith sourced answers; offline engine, API when online
  • Six-language voice — Hindi, Urdu, English, Arabic, Bangla, Tamil
  • One UI, three targets — Turborepo monorepo shipping web PWA plus native Android and iOS

Stack

  • Next.js 14 PWA
  • Capacitor 7
  • Turborepo
  • TypeScript
  • FastAPI
  • Dexie / IndexedDB
  • Whisper
  • On-device ML
05 / Also in the workshop

Other products we run, and the work we do alongside clients

Not everything is a flagship. These are smaller products of our own, plus platforms we build jointly with a client's team — where the code and the data belong to them, and we bring the AI engineering.

Our product

AI Lab

Choosing a model by reading marketing pages does not work. AI Lab runs one prompt across many models in parallel and scores them on the things that decide the choice — judge scoring, embedding similarity, latency, token cost and human annotation — so a model decision is evidence-backed.

  • Parallel execution across OpenAI, Anthropic, Cohere and custom endpoints
  • Composite scoring: LLM-as-judge, similarity, latency, cost, feedback
  • API keys encrypted at rest with Fernet; role-based access
  • Run history and analytics for reproducible comparisons
  • Flask
  • SQLAlchemy
  • asyncio
  • httpx
  • PostgreSQL
Our product

ZeroCourse

Most courses teach everyone the same material and test at the end. ZeroCourse inverts it: diagnose what a learner already knows through adaptive assessment, model that state on a knowledge graph, then generate only the micro-curriculum that closes the actual gap.

  • Adaptive diagnostic assessment as the starting point
  • Knowledge-graph skill modelling in Neo4j
  • Evidence-based certification and continuous reassessment
  • Dynamically generated micro-curriculum per learner
  • FastAPI
  • Neo4j
  • PostgreSQL
  • Redis
  • Docker
Client & collaboration

Procena.ai

Agentic procurement intelligence, built together with the client's team and running at procena.ai. Twenty-five configured agents read the customer's own spend, supplier, contract and project data and apply named procurement frameworks — Kraljic, Porter, TCO, should-cost, category strategy — to produce advice a CPO can act on.

  • Every monetary figure computed by a deterministic engine, never by the model
  • Uncited figures stripped before output — enforced in code, not prompts
  • Hexagonal FastAPI backend: 24 bounded contexts, domain logic free of ORM
  • Adding an agent is configuration — a YAML file and a prompt, no Python
  • Provider-agnostic LLM registry
  • FastAPI
  • TanStack Start
  • PostgreSQL
  • Supabase
  • Alembic
  • Docker
procena.ai
06 / Client engagements

Work delivered for clients across publishing, pharma and finance

Shorter engagements, each solving one concrete problem. Client names are withheld under the terms we work to.

Alt-text & long-description generation

An accessibility tool that writes alt-text and long descriptions for images at scale, exposed as both a REST API and a web interface so editorial teams and pipelines could use the same engine.

  • Long descriptions for complex figures and charts
  • Batch processing for existing archives
  • Vision models
  • REST API
  • Python

Drug safety intelligence

A pharmacovigilance platform that reads Individual Case Safety Reports with NLP and LLM/SLM models, classifies and summarises them, and scores risk — producing output a regulator will accept.

  • Automated safety classification and summarisation
  • Risk scoring and anomaly detection
  • Compliance-ready regulatory reporting
  • NLP
  • LLM / SLM
  • Python
  • Pharma IT integration

Embedded search for academic research

Semantic search over a large research corpus, embedded directly into the publisher's existing reading experience so researchers never leave the page they are working in.

  • Meaning-based retrieval, not keyword matching
  • Embedded widget, no separate destination
  • Embeddings
  • Vector search
  • Python

Crypto trading bot

An automated trading system with signal generation, position sizing and hard risk limits — built so that a bad model cannot become an unbounded loss.

  • Strategy backtesting before live capital
  • Enforced risk and exposure ceilings
  • Python
  • Exchange APIs
  • Time-series ML

Multi-domain web scraping

Resilient data extraction across many sites with differing structures, including change detection so a silent layout shift surfaces as an alert rather than as corrupted data.

  • Schema normalisation across sources
  • Scheduling, retries and failure alerting
  • Python
  • Scrapy
  • Playwright
  • PostgreSQL

E-commerce platform

A full commerce build — catalogue, checkout, payments and an admin back office — delivered end to end and deployed to production.

  • Payment gateway integration
  • Inventory and order management
  • Python
  • REST APIs
  • PostgreSQL

PDF document automation

Extraction and transformation of structured content from large PDF sets — the unglamorous work that turns an archive into something queryable.

  • Layout-aware extraction, including tables
  • Automated generation and templating
  • Python
  • OCR
  • Document parsing

Custom model development

Bespoke models trained and fine-tuned on client data where an off-the-shelf model could not meet the accuracy, latency or cost requirement.

  • Dataset curation and labelling strategy
  • Fine-tuning, evaluation and serving
  • PyTorch
  • Fine-tuning
  • Model serving
06 / Next

Have something harder than this?

The work above started as problems nobody had a clean answer for. Tell us yours and we will say honestly whether we are the right team for it.

Start a project What we do