Industry
The margin left in logistics is in the decision, not the mile
Freight, last-mile and mobility platforms run on decisions made in seconds against data that arrives late, noisy and incomplete. We build the routing, ETA, document and exception systems that keep those decisions honest at scale.
Logistics & Mobility
Routing, ETA and exception systems that hold up when the network is noisy and the margin is thin
- Now
- 3 applications
- Next
- 3 applications
- Frontier
- 1 application
- Constraints we design around
- 5
Context
Three forces reshaping this sector.
What changed recently enough that the old architecture is now the wrong one.
The remaining margin sits in the dispatch loop
Fuel, driver pay and warehouse rent are close to floor for most operators. What is still elastic is which vehicle takes which order, in what sequence, against which promise. A 3% improvement in stops per route is a larger swing than any procurement round, and it is won in software, not in negotiation.
Regulated freight data became machine-readable
The e-way bill and e-invoice regimes, FASTag tolling reads and the Unified Logistics Interface Platform turned paper trails into APIs of record. The consequence is not fewer documents — it is that a document can now be reconciled against an authoritative source instead of being trusted because someone photographed it.
Consumer tracking expectations leaked into B2B freight
Shippers stopped accepting a status code and started asking for a promise date with a confidence attached. That is a different engineering problem: a point estimate can be wrong quietly, a distribution forces the operator to decide what risk they are willing to communicate.
Applications
What we would build here, in order.
Grouped by how close each one is to production today — not by how impressive it sounds in a board pack.
Build it this quarter
Current models, current data. The engineering is understood and the risk is schedule risk, not research risk.
Address normalisation and rooftop geocoding
Indian addresses are landmark-relative, inconsistently ordered and frequently carry a PIN code that contradicts the locality. We build a resolver that parses with libpostal-style tagging, reconciles against OpenStreetMap, Mappls and the operator’s own delivered-point history, and returns a coordinate with a confidence band — plus the reason it is uncertain. Historic successful deliveries are usually the strongest signal available, and almost nobody mines them.
ETA that ships a distribution, not a point
Gradient-boosted or sequence models over map-matched traces, trained per lane and per hour-of-week, predicting quantiles rather than a mean. The promise date is then a business choice — publish P70 and absorb the tail, or publish P90 and lose conversion. Calibration is monitored continuously, because an ETA model drifts the moment a flyover opens.
Freight document intelligence
E-way bills, lorry receipts, bills of lading, proof-of-delivery photographs and carrier invoices, extracted with a layout-aware model and gated on field-level confidence. Every extraction is reconciled against the authoritative record where one exists — the NIC e-way bill API, the GST IRP — so the model produces a candidate, not a fact. Freight-bill audit against contracted rate cards falls out of the same pipeline.
One thing has to change first
Buildable, but gated on something outside the model — labels, a consented holdout, an integration, or a process the business has to agree to change.
Exception detection and proactive customer communication
Most shipments fail quietly long before anyone raises a ticket: a vehicle idle past a threshold outside a geofence, a scan that never happened, a hub dwell time two standard deviations off its lane. Streaming detectors raise the signal, a policy layer decides what is worth telling whom, and a language model drafts the message a human approves or auto-sends by tier.
Dynamic dispatch and mid-route re-optimisation
A batch VRP solve at 6am is obsolete by 9. The working pattern is a solver kept warm — OR-Tools or a hybrid metaheuristic — re-solving a bounded neighbourhood on every material event, with hard guardrails on driver churn so the app does not reshuffle a rider’s day every ninety seconds. Stability is a constraint, not an afterthought.
Dock, yard and maintenance-bay scheduling
Doors, yard slots and workshop bays are the same problem wearing three uniforms: constrained servers, stochastic arrivals, and a queue that costs money. One scheduler over all three, fed by telematics-derived fault codes and appointment APIs, beats three spreadsheets that cannot see each other’s commitments.
Scope it as research, with a gate
Credible and worth a prototype. We would run it behind an evaluation gate and say plainly if it does not clear the bar.
Capacity matching without pooling raw data
Shippers and carriers both want a thicker market and neither wants to hand over its rate card or customer list. Matching on privacy-preserving joins and derived features, with cryptographic attestation of what was computed, makes the market possible without the trust that does not exist. Technically credible in 2026, commercially still ahead of most networks.
Constraints
What makes this sector different.
The part of the build that does not transfer from anywhere else. Every one of these has to be answered in the architecture, not in the project plan.
The location is wrong twice over
The address is free text written by a human in a hurry, and the GPS trace is noisy where it matters most — urban canyons, basements, tunnels, devices asleep to save battery, and the occasional spoofed fix. Map-matching with an HMM over the road graph and explicit gap handling is baseline. Any system that treats a raw ping as truth will confidently mis-sequence a route.
Inference cost is a line item on every parcel
When contribution per delivery is measured in tens of rupees, a frontier-model call per event is a pricing decision, not an implementation detail. The workable shape is a cascade: deterministic rules first, a small fine-tuned model next, a large model only for the residual. Cost per decision belongs on the dashboard beside latency.
The driver’s phone is the weakest node and the most important one
Entry-level Android, patchy 4G, a battery that must survive a ten-hour shift, and a user with one hand free. The app has to work offline, queue events durably, reconcile on reconnect without duplicating scans, and never block on the network. Any model that must run there has to fit in a quantised ONNX or TFLite budget, not a GPU.
Real time at scale means skew, not throughput
Aggregate volume is rarely the problem. Peak-hour bursts on a handful of hot pincodes are. Partition keys chosen for convenience create hotspots, out-of-order pings break naive windowing, and replay after an outage either duplicates notifications or drops them. Watermarks, idempotency keys and backpressure need designing in — retrofitting them is a rewrite.
Every party sees a different slice, and that is contractual
Shipper, 3PL, carrier, driver and consignee sit inside different trust boundaries. Rate cards are commercial confidences held by contract; consignee phone numbers and delivery photographs are personal data, carrying consent, purpose-limitation and retention duties under the DPDP Act. Authorisation is domain logic, not middleware. Designing it late means either a leak or a lattice of special cases nobody can reason about.
Every one of these is discoverable in week one. Most programmes discover them in month three, after the architecture is already load-bearing.
Outcomes
What we design toward.
Targets and typical ranges for this class of system. Each one carries its basis, because a number without one is decoration.
- Intra-city ETA error we design toward
- P80 within ±10 minIntra-city ETA error we design towardDesign target for a first production cut, set against a held-out replay of the operator’s own historical trips before any rollout — not a measured client result.
- Freight documents cleared without a human touch
- 70–90%Freight documents cleared without a human touchTypical straight-through range for this class of structured Indian freight document once the confidence gate is tuned; the remainder is routed to review deliberately.
- Assignment decision latency at p99
- < 200 msAssignment decision latency at p99Engineering budget the dispatch loop is built against, including feature lookup and solver time; stated as a design constraint, not an achieved benchmark.
- Inference cost gap a cascade closes
- 10–20xInference cost gap a cascade closesArithmetic on published 2026 list pricing for a small hosted model versus a frontier model on the same per-event payload — a cost model, not a delivered saving.
These are design targets and published benchmark ranges, not delivered client results. We do not publish results attributed to clients.
Solutions
What we bring to this sector.
The practices that do most of the work here. Each one is the same discipline applied against the constraints above.
Blueprints
Reference architectures behind this work.
How the systems are actually assembled, with the engineering decisions and their costs stated. Reference designs, not case studies.
Real-Time Decision Platform
A platform that turns database changes into fresh features within seconds, serves a model decision inside a live request, and writes an immutable record of every decision — the inputs, the model version and the policy that applied — so it can be explained, replayed and rolled back months later.
Read the blueprintDocument Intelligence Pipeline
A high-volume extraction pipeline that turns invoices, claims, KYC packs and trade documents into validated structured records, attaches a calibrated confidence to every field, and sends only the fields that genuinely need a person into an exception queue sized to the real error rate.
Read the blueprintQuestions
The awkward questions.
The ones a technology committee actually asks, answered the way we would answer them in the room.
Next step
Bring us the constraint you keep working around.
A 45-minute working session. We'll tell you what we'd build, what we'd not build, and roughly what it costs. No deck.