Founding AI Architect – Spatiotemporal Intelligence

Location: Pune, India On-site

Type: Full-time Founding Leadership Team

Reports to: CPTO

WHY THIS ROLE EXISTS

Dispatch is building a real-time decision layer for last-mile logistics in India the system that turns demand signals, fleet telemetry, geospatial context, and operational constraints into live routing, matching, and network-management decisions across multiple cities.

You will own that decision layer end to end. Architecture, models, infrastructure, and the team that scales it.

This is a founding systems role. Not a forecasting head on a vendor stack a ground-up engine for a problem the published playbooks don't fully solve.

TEH PROBLEM

Most network orchestration architectures in the literature assume a closed-loop platform: one operator controls supply, demand, pricing, and promotion. The problem we work on is structurally different. The shape of demand is influenced by external actors we don't control; the fleet is ours to manage. The lever that moved the metric is rarely a lever we pulled.

That changes what good modelling looks like.

• Stationary assumptions break.
• Off-the-shelf supervised learning fails at the counterfactual we only ever observe the outcome of the decision we made, never the alternatives.
• The system has to learn continuously from operational reality, separate its own influence from external dynamics, and know when it is outside the conditions it was trained on.

The architecture that satisfies all of this is original work. We don't expect to find it pre-built in a paper.

WHAT YOU'LL OWN

Architecture & research

The representation of the city and the fleet, the heads that turn that representation into decisions across the order lifecycle and across the fleet, and the path from per-capability scoring to a unified model over time.

Counterfactual machinery

Exploration scheduling, off-policy evaluation, simulator design the apparatus that lets a learned policy earn the right to take live traffic.

Production systems

Low-latency inference inside the decision path, online and offline feature stores, training pipelines on multi-month event histories, and the infrastructure (lineage, drift, train/serve skew contracts) that decides whether learned models survive contact with production.

Team

Hire and lead ML scientists and engineers. Set the bar for experimentation, evaluation, and review. Build the internal tooling diagnostics, decision-log replay, simulator UX that the team needs to move fast without losing rigor.

Strategy

Translate operational reality into modelling decisions. Partner with product and ops to deploy intelligence into live workflows. Communicate tradeoffs to founders, board, and partners.

WHAT WE'RE LOOKING FOR

Required:

• 6+ years building and shipping production ML at least one system you owned from architecture through rollout, with measurable real-world impact.
• Direct experience with temporal, spatial, or spatiotemporal modelling depth in at least one, taste for the rest.
• Comfort with the counterfactual / off-policy regime: bandits, offline RL, causal inference, or rigorous uplift work at scale. If you've only operated where ground truth is fully observed, this role will surprise you.
• Strong Python and data systems fundamentals. You'll read and write production code, not just decks.
• Production MLOps fluency: experiment tracking, lineage, drift, staged rollouts.

Strong signal:

• Built or contributed to forecasting, routing, or marketplace-decision systems with consequential outcomes money moving, vehicles moving, livelihoods on the line.
• Worked on open-loop or partially-observable marketplaces anyone who has lived the "we don't control the lever that moved the metric" problem.
• Spatial modelling depth, graph methods on network-shaped data, simulator-gated policy rollout, large-scale telemetry, distributed training, or low-latency inference at city scale.

Preferred:

• Logistics, mobility, ride-hailing, quick-commerce, or transportation systems.
• Simulation or digital-twin environments building them, not just consuming.
• Operations research foundation with ML on top.

WHAT YOU SHOULD KNOW

• Founding leadership. You report to the CPTO and sit on the founding team. Technical strategy is set jointly; execution is yours.
• Pune, on-site. This system has to be debugged at 11 PM in a war room, not over Slack at a +10-hour time zone offset.
• Fast time to impact. The data infrastructure, decision-logging, and shadow-mode serving stack are live. Your first learned model trains on real operational data inside the first month.

HOW TO APPLY

Send us:

1. A pointer to a production ML system you owned what you built, what broke, what you'd do differently.
2. One paragraph on what makes counterfactual learning harder than supervised learning in a real marketplace, and how you'd approach it.
3. Anything else that shows how you think a blog post, a talk, a side project, a code review you're proud of.

Technical conversation within 2 days of applying.