Production
AI Systems.

Shipped to production, not proof-of-concept — the architecture and engineering decisions behind each one.

Coming· Sep '26
Agricultural Computer-Vision logo

DEFRA funded

Agricultural Computer-Vision

On-device livestock vision — a quantised YOLO model on iPhone estimates cattle weight and body condition from the camera, fusing LiDAR depth for weighbridge-grade accuracy.

  • On-device, offline-first — a quantised YOLO model runs on iPhone at 30+ FPS (INT8) — inference is fully on-device, with no internet dependency.
  • Camera + LiDAR depth fusion — camera vision (2–10m) augmented by LiDAR depth (0.3–2m optimal) for ±5% weight accuracy against a calibrated weighbridge.
  • Vet-validated multi-output — weight, body-condition score, and visual health indicators, with BCS validated to 85% agreement against a veterinary surgeon.
  • Data flywheel — COCO transfer learning fine-tuned on 3,000+ CVAT-annotated cattle images; anonymised data feeds back for continuous refinement, trained on a GPU cluster (AWS).
Computer VisionYOLOv11-segCoreMLLiDARARKitMLOpsAWS
Coming· Aug '26
Agentic Property Intelligence logo

Antec Global

Agentic Property Intelligence

A multi-agent property assistant — a router delegates to specialist Property, Neighbourhood and Risk agents, with structured output and automatic model failover.

  • Capability-based routing — automatic failover across Claude Sonnet 4 → GPT-4o → GPT-4o-mini.
  • Structured output everywhere — Zod-governed; a GPT-4o planner composes multi-step plans for specialist agents to execute.
  • Response caching — LLM caching middleware (Upstash Redis) — 60–80% cost cut on repeat queries.
  • Price prediction + on-device vision — LightGBM (FastAPI) on 30M+ UK transactions; YOLOv11n-seg → CoreML at 30+ FPS in an ARKit iOS app.
Multi-agentClaude Sonnet 4GPT-4oVercel AI SDK v5PythonFastAPILightGBMYOLOv11n-segCoreML
Live· Feb '26
PullBot logo

Google + Dentsu

PullBot

An agentic PR reviewer in CI — Claude Opus 4.6 reads each diff via the Anthropic SDK and returns schema-constrained findings, with review guidelines chosen per file type.

  • Schema-constrained findings — JSON with severity, type, and category — enforced by the API, never prose.
  • Token-aware chunking — exact pre-flight token counts pick single-pass or chunked review; a bin-packing chunker splits big PRs, worst-verdict wins on merge.
  • Prompt caching as cost architecture — cache hits logged on every review.
  • Runtime review skills — fetched from a GitHub-hosted library, selected per file type with per-repo overrides.
Claude Opus 4.6Anthropic SDKStructured outputPrompt cachingGitHub ActionsGCPTypeScript
Live· Jan '26
ATS Check logo

Antec Global

ATS Check

A multi-model CV-screening and research pipeline — routes Claude, GPT-4.1, Groq and Grok by task inside a durable workflow that pauses at a human checkpoint.

  • One gateway, model per task — Claude Sonnet 4.5 and Haiku, GPT-4.1, Groq Llama 3.3, Grok 4.1 — each assigned where it fits.
  • Eval-driven model swaps — a comparison eval harness documented 40–67% cost cuts per task with quality held.
  • Durable workflows — long-running jobs as durable workflows (Vercel Workflow DevKit) with human-in-the-loop re-entry.
  • Layered JSON repair — retry, fallback model, then jsonrepair — so structured output never breaks the pipeline.
Vercel AI SDK v6Claude Sonnet 4.5GPT-4.1Groq Llama 3.3Grok 4.1Vercel WorkflowsVercel AI GatewayRAG
Live· July '25
Ticket Badger logo

Antec Global

Ticket Badger

Semantic event discovery over 50K+ listings, by text or photo — pgvector search fused with date, venue and price filters in a single query.

  • pgvector inside Postgres — semantic search (text-embedding-3-small, 1,536-dim) fused with ordinary SQL filters — date, venue, price — in one query.
  • Zero-LLM query classifier — extracts constraints before any model is called — routing queries as constraint, semantic, or hybrid.
  • Multimodal photo input — photo → personality cues → event recommendations via GPT-4o-mini vision, with a three-stage fallback chain.
  • Two-tier caching — in-process LRU plus Upstash Redis over embeddings and vector search.
pgvectorSQLtext-embedding-3-smallGPT-4o-miniRAGVercel AI SDKStrands SDKEval Harness

Want the full story — teams led, platforms scaled, startups founded?

Full experienceGet in touch