Susmita Singh

Susmita Singh

Susmita is a Tech Lead at Google Store, where she drives core AI infra, system architecture, and technical strategy for the Google Store AI Infrastructure team. Operating at the intersection of applied AI and core software engineering, she leads cross-functional initiatives spanning high-throughput AI infrastructure, scalable model serving, server-driven architectures, and large-scale distributed systems. Prior to her work on Google Store, Susmita built a strong foundation in software engineering, developing highly scalable distributed architectures and resilient consumer-facing applications.

Susmita is a Fellow of the Institution of Engineering and Technology (FIET), a Distinguished Fellow of the Soft Computing Research Society (DFSCRS), and an IEEE Senior Member. She actively serves on program committees and judging panels for premier international technology conferences, including SIGCSE, CAV, ASE, and ACSOS. Her academic contributions include reviewing for peer-reviewed journals such as the Journal of Open Research Software and Engineering Applications of Artificial Intelligence.

Authored Publications
Sort By
  • Title
  • Title, descending
  • Year
  • Year, descending
Preview abstract Context: The cost of frontier large language model inference has fallen by two orders of magnitude since 2023, yet the techno-economic forces governing AI value capture remain poorly understood. No existing work provides a unified, multi-layer framework connecting hardware physics to commercial pricing to actuarial constraints. Objectives: This survey aims to establish that Generative AI (GenAI) monetization is structurally bound by five interdependent techno-economic layers: (1) the physical constraints of memory bandwidth and compute, (2) deflationary architectural innovations, (3) the algorithmic economics of inference-time compute, (4) the verification economics governing outcome-based pricing, and (5) the macro-legal realities of enterprise liability. Methods: We conduct a Multivocal Literature Review (MLR) adapting the PRISMA protocol, synthesizing peer-reviewed and grey literature sources—vendor documentation, SLAs, and API pricing data (2022–2026). Two reviewers independently screened all records (Cohen’s κ ≥ 0.81 across all decision stages). Results: We contribute four primary artifacts. First, the Viability Inequality, an analytical model formalizing the conditions under which outcome-based AI pricing is economically sustainable. Second, the Billing Fallacy: aggregate cost growth is driven by Agentic Recursion, not quadratic attention complexity. Third, the Verifiability Bifurcation: objective task domains enable outcome pricing, while subjective domains depend on proxy-based models. Fourth, the Multi-Layer Techno-Economic Taxonomy (M-TET), a unified five-layer framework mapping the full monetization stack from silicon-anchored token pricing through actuarial risk ceilings. Conclusion: GenAI monetization is not a commercial pricing exercise but a dynamic negotiation across hardware, algorithmic, economic, and actuarial layers. In subjective and hybrid task domains, the binding constraint on outcome-based pricing is the cost of verification, not generation. AI value capture depends on engineering low-cost, high-fidelity Verification Engines. View details
Preview abstract Config-driven user interface (CDUI), also called server-driven user interface, moves user interface (UI) layout out of the app binary into server responses so teams can update screens without an app store release. Because one schema serves many installed app versions, edits to it can break older apps. We statically compared parsed schema declarations across four open-source runtime CDUI engines (Yandex DivKit, Cash App Redwood, EnsembleUI, and archived Spotify HubFramework) and one inactive comparator compiled at build time (Airbnb Lona). Out of 1447 commits touching schema paths, 730 modify a runtime engine’s schema: 61.4% only add declarations, 38.1% remove or retype declarations, and 0.5% mark deprecations (93% parser accuracy against two-rater human labels). Standalone schema files had a higher share of additive commits than schemas written in code, though across four runtime engines this difference is not statistically significant. By Fisher’s exact test across commit halves, breaking edits fell significantly only in Redwood (73.3% to 25.0%, 𝑝=0.012). Finally, we compare systems across modularity, centralization, and strictness, showing when additions still break older clients (including Redwood throwing by default). View details
Preview abstract Subscription software must answer one question on every request: may this user do this, now? Production systems usually treat the answer as authoritative state, read at request time, and assume that stronger consistency is always safer. We systematize what is publicly known about entitlement correct- ness when billing truth is owned externally, and ask whether the consistency- token protocols of internal authorization transfer to it. Method. A multivocal review of 85 sources, two propositions, an error-cost model, TLC model checking, and a seeded simulation of 20,000 subscribers replicated across ten seeds. No Zanzibar-style token can exist for entitlement, because external billers share neither a write path nor a clock; staleness bounded by the recon- ciliation interval is achievable instead. In simulation the projection dominates: applying revocations at their effective time rather than on arrival cuts expected cost 27-fold, more than every freshness parameter combined. Under partition, serving the last snapshot dominates; failing open beats failing closed for sub- scribers whenever the cost ratio ρ of false denial to over-service exceeds about 0.35, but the break-even reaches 26 when 5% of users pay. A revocation hold- over pays only against spurious revocations, never beyond the reconciliation interval, and daily reconciliation makes error nearly insensitive to notification loss. Entitlement should be engineered as an eventually consistent derived view with an effective-time projection, per-class staleness bounds, and reconciliation as its source of truth. We contribute a vocabulary, six invariants, a reference architecture, a model-checked specification, and an open simulator. View details
×