Principal Engineer: Retrieval Geometry & Vector Quantization in Cambridge

FactTrace is building the world's most precise content provenance infrastructure. From our base in Cambridge UK, we trace how language travels, mutates, and resurfaces across different systems — mapping the lineage of every claim from its origin through every paraphrase, reframing, and reinterpretation that follows. We are an early-stage, research-intensive company operating at the intersection of representation learning, information retrieval, and applied cryptography. Working towards a reference corpus exceeding the 100 million documents and continues to scale. The systems we are designing today will underpin a forthcoming Zero‑Knowledge cryptographic layer — and they demand mathematical structure of the highest order.

This is not a role for engineers who want to wrap foundation models in thin orchestration. We are designing proprietary representational spaces from first principles, and we are hiring the engineer who will lead that effort.

The Mission

You will architect the next generation of retrieval — systems that move beyond the implicit compromises of the current AI stack and set a new standard for what retrieval can be. Standard dense retrieval groups text by topical proximity; the geometry conflates what a passage is about with what it actually asserts. That gap is the entire problem space. We are constructing intent‑aware vector spaces in which the model can natively distinguish between a faithful paraphrase, a direct refutation, and a subtle distortion — preserving factual intent as a first‑class geometric quantity. The retrieval pipeline you build must answer a single question with sub‑second latency against a 100M+ document corpus: where did this claim originate, and how has it evolved? Answering it well means retrieving with precision that conventional architectures cannot reach, and doing so without the computational cost that has, until now, been the price of that precision.

What You’ll Lead

• Single-stage, precision-first retrieval architecture. You will design retrieval pipelines that achieve cross‑encoder‑grade semantic precision in a single forward pass, eliminating the two-stage retrieve‑and‑rerank paradigm as a computational bottleneck. This is the central research and engineering challenge of the role: producing the discriminative geometry of a cross‑encoder at the latency and throughput of a bi‑encoder, at corpus scale. Your work here defines our retrieval ceiling.
• Vector quantization and binarization at scale. You will pioneer compression regimes for our embeddings that preserve full semantic resolution at massive scale — moving beyond naïve dimensionality reduction into learned quantization, product and residual quantization, and binarized representations that retain the geometric structure our downstream systems depend upon. Compression without distortion is the design constraint, not an aspiration: the representational fidelity you preserve is the input to our cryptographic layer, where any geometric degradation would propagate irrecoverably.
• Intent‑aware representation learning. You will own the training objectives, contrastive structures, and hard‑negative regimes that produce our proprietary vector spaces — drawing on metric learning and Natural Language Inference to encode entailment, contradiction, and neutral stance as native geometric relationships rather than downstream classification tasks.
• The mathematical foundation. Every retrieval index, quantization scheme, and embedding you ship is also cryptographic substrate. You will work in close partnership with our cryptography track to ensure the geometry you produce is pristine, well‑conditioned, and amenable to the Zero‑Knowledge proofs that will operate over it. This is representation learning with a mathematical mandate that few teams in the world are positioned to pursue.

What We’re Looking For

We expect deep, demonstrable expertise across representation learning, metric learning, and Natural Language Inference, plus the engineering judgment to ship these systems at corpus scale. Specifically:

• Representation & metric learning fluency. Substantial experience designing contrastive, triplet, and supervised metric learning objectives; familiarity with hard‑negative mining, debiased contrastive losses, and representation collapse — and how to design around them.
• NLI as a training signal. Strong grasp of entailment, contradiction, and neutral relation modelling, and a track record of using NLI signals to shape representational geometry rather than treating them as a classification head.
• Retrieval systems at scale. Experience with ANN indexing (HNSW, IVF, ScaNN, or comparable), and a clear thesis on single‑stage retrieval — why and how it can surpass two‑stage retrieve‑and‑rerank pipelines.
• Vector quantization & binarization. Working knowledge of PQ, OPQ, residual quantization, and learned binary codes; understanding of the precision–compression frontier and how to reason about it rigorously rather than empirically.
• Production engineering. Comfortable owning systems from research prototype to high‑throughput production, with strong Python, PyTorch (or JAX), and a deep understanding of the GPU memory and latency profiles of retrieval workloads. A research profile is welcome but not required; we weight demonstrated, shipped systems above publication counts. What is required is the conviction that the current retrieval stack is not the ceiling — and the ability to build the one above it.

Why This Role

• A research‑grade problem at production scale. Few teams are simultaneously pushing the frontier of intent‑aware representation and operating 100M+ document retrieval under hard latency budgets.
• Direct cryptographic impact. The geometry you produce will be reasoned over by Zero‑Knowledge proofs — a rare and demanding constraint that elevates every design decision you make.
• Cambridge, at the centre of it. We are rooted in one of the world's deepest concentrations of ML, cryptography, and information‑retrieval talent.