Three systems that had to work.

The challenge

Bond market making is not equity market making. The instrument universe is large, quoting is request-driven, liquidity is fragmented across venues, and pricing depends on curve construction and model calibration that cannot possibly run inside the order path.

The platform in place coupled pricing, execution and risk into components that had to be released together. Onboarding a venue meant a full-platform release. Changing a pricing model meant the same. The desk was effectively choosing between shipping improvements and keeping the platform stable.

The approach

We redrew the boundaries around failure and change rate rather than around teams, then made latency an explicitly measured requirement instead of a shared hope.

• Split by rate of change — pricing, quoting, execution, position and risk, market data and venue adapters became independently deployable.
• ION bus as the backbone — the middleware the desk already trusted, with explicit message contracts and versioning between services.
• Heavy maths out of the hot path — curve construction and calibration pushed into C++ and Julia behind JNI bridges.
• Latency as a gated requirement — percentile tracking on every hop, allocation-free critical sections, regression gates in CI.

What we built

• A 50+ service platform communicating over ION bus middleware, each service owning its own release cycle.
• FIX connectivity through QuickFIX/J: session lifecycle, sequence recovery, drop-copy reconciliation and message-level auditability.
• JNI bridges from JVM services into native C++ and Julia libraries for pricing and calibration.
• Real-time position and risk, continuously recomputed, with pre-trade limit checks inside the order path and kill-switch semantics that genuinely stop flow.
• A latency measurement harness wired into CI, so a change that costs microseconds shows up before production does.

The outcome

• Sub-millisecond execution latency sustained across the hot path, measured at the percentiles that matter rather than the average.
• Venue onboarding became an adapter change, not a platform release — new connectivity ships without touching pricing or risk.
• Model iteration decoupled from platform cadence: quants change pricing behaviour without waiting for a release train.
• Regulatory review supported by the record — every order carries a message-level trail from decision to venue acknowledgement.

The challenge

An acceptance test involves at least four parties, each with partial visibility: the manufacturer, the supplier, the client and an independent testing agency. Coordination ran on email and shared drives — scheduling by thread, evidence as attachments, status as an opinion.

Non-conformities lived in spreadsheets that drifted away from the ERP within days. When an auditor asked for the evidence behind an acceptance, assembling it took people off their actual work for the better part of a week.

The approach

Digitising the form would have saved paper. We digitised the process itself: the legal transitions, the evidence each one requires, and who is allowed to make it.

• Inspection as an explicit state machine — not a status column. Transitions are validated, attributed and irreversible by design.
• ERP as system of record — orders, items and suppliers integrated rather than duplicated, with reconciliation and replay instead of one-shot imports.
• Documents as first-class evidence — versioned, immutable, attached to the transition that produced them.
• Role-scoped portals — each party sees its own view of the same underlying truth.

What we built

• A Java/Spring backend with a React front end over PostgreSQL, deployed on AWS.
• Bidirectional ERP interfaces with reconciliation, replay and conflict reporting.
• S3-backed document management with versioning, retention and full traceability.
• State-machine workflows covering scheduling, execution, non-conformity, rework and final acceptance.
• Real-time compliance dashboards, with non-conformities attached to the ERP object they belong to.

The outcome

• One source of truth across four organisations — status is derived from the record, not negotiated over email.
• Audit evidence assembled, not reconstructed: the trail already exists because the process produced it.
• Non-conformities stopped drifting from the ERP, because they are attached to it rather than copied out of it.
• Coordination cost moved into the software — the portal replaced the mailbox as the coordination layer.

The challenge

Planning ran on spreadsheets: one forecasting method applied uniformly across a heterogeneous catalogue, safety stock set by rule of thumb, and no way to evaluate a policy change other than adopting it and waiting a quarter.

Demand variability and supplier lead-time variability were being treated as if they were the same problem — so buffers were sized for the wrong risk and placed at the wrong echelon.

The approach

Better forecasts do not fix wrongly placed buffers. We treated the two as separate problems and measured each against a baseline that could not be argued with.

• Forecasting separated from policy — each evaluated on its own terms, against an explicit naive baseline.
• Stochastic demand and lead time — modelled independently, so buffers absorb the variability that actually causes the stock-out.
• Optimisation under constraints — capacity, service level and cost expressed as a trade-off curve rather than a single answer.
• Delivered into the planners' tools — results land where the work already happens, so the model survives contact with the working week.

What we built

• A Python/Pandas modelling stack with ML demand forecasting and reproducible backtesting.
• Multi-echelon inventory optimisation over stochastic demand and lead-time distributions.
• Constraint-based optimisation exposing the cost-versus-service-level frontier.
• Data pipelines on Apache NiFi and RabbitMQ over Oracle sources, with Jenkins driving scheduled runs and retraining.
• Scenario simulation so a planner can compare policies before committing to one.

The outcome

• Measurable reduction in stock-outs, evaluated against the pre-existing policy rather than against the model's own optimism.
• Scenario comparison in minutes instead of a spreadsheet rebuild — planning conversations moved from "what does the file say" to "which trade-off do we want".
• Forecast quality became a measured number, with a baseline anyone could reproduce.
• Inventory policy became reviewable — assumptions written down, versioned and testable.