Senior Backend Engineer, Vision

About Sarvam Sarvam is building the bedrock of Sovereign AI for India. The company is developing India's full-stack sovereign AI platform, building across research, models, infrastructure and applications with a singular focus on making AI genuinely work for India. Sarvam works with leading enterprises and public institutions and is backed by Lightspeed, Peak XV, and Khosla Ventures. Sarvam partners with India's leading brands, including Tata Capital, SBI Life, CRED, IDFC, and LIC. About the Team Sarvam's research teams build our own vision-language models for OCR and structured extraction. This team builds everything around them — the serving harness that turns a 3B or 30B in-house model into a production document intelligence platform. The bet is specific: with the right harness — routing, decomposition, retries, verification, ensembling, layout awareness, confidence calibration — a small sovereign model should match or beat what teams today get from frontier hosted models like Gemini Flash, at a fraction of the cost and fully within India. Closing that gap is an engineering problem, and it is this team's problem. We run against the full messiness of Indian documents at population scale: PAN and Aadhaar, bank statements, GST filings, insurance and medical reports, 60-page contracts, legal filings and RFPs — across languages, scan quality, and layouts that were never designed to be machine-read. Stack: Go, Python, Temporal, REST, Kubernetes, PostgreSQL, Redis, object storage, OpenTelemetry-based observability. About the Role You will own the architecture of the serving harness for Sarvam's vision models — the system that has to deliver frontier-grade extraction quality out of 3B and 30B in-house models, at national scale, with cost and latency budgets that actually close. This means owning the hard trade-off surface directly: accuracy versus latency versus rupees per page. Multi-pass inference, model routing and cascades, self-consistency and verification passes, confidence-driven escalation, batching and caching strategy, GPU utilisation. These are the levers that decide whether the product works, and you will be the person deciding how to pull them. You will also set the reliability bar. These pipelines process documents that customers cannot afford to lose — KYC, loan underwriting, claims, contracts. Durability, idempotency, backpressure and graceful degradation are the baseline, not the roadmap. The architecture you set will be inherited by everything the team builds after you. What You'll Do Own the end-to-end architecture of the OCR and extraction serving harness: API layer, orchestration, inference layer, post-processing, delivery Design the accuracy harness — multi-pass extraction, ensembling, cross verification, schema-constrained decoding, confidence calibration, targeted re runs — and prove its gains against held-out evaluation sets Architect durable, resumable document workflows in Temporal: fan-out across pages, partial failure recovery, exactly-once side effects, long-running jobs measured in minutes to hours Own the inference serving layer alongside infra: batching strategy, GPU pool management, autoscaling on real signals, queue depth and admission control, multi-model routing Drive latency, throughput and unit economics down deliberately — profile, measure, and defend cost-per-page targets as volume scales Build the observability substrate: distributed tracing across the pipeline, per-stage cost and quality metrics, SLOs, alerting, and post-incident rigour Design for multi-tenancy, tenant isolation, rate limiting and fair scheduling across enterprise customers with very different load shapes Support on-prem and constrained deployments where the whole harness has to run inside a customer's environment Set technical direction and raise the bar through design review and mentorship of SDE 1–2 engineers What We're Looking For - 5–6+ years in backend engineering, with meaningful time spent operating high throughput production systems you were on-call for - Deep proficiency in Go and/or Python, and the judgement to know which belongs where - Strong distributed systems design: queues, workflow orchestration, idempotency, backpressure, retry and timeout semantics, consistency trade-offs, graceful degradation - Production experience with Temporal or an equivalent durable execution engine, on workflows that mattered - Kubernetes in production — autoscaling, resource management, rollouts, debugging under load; GPU workload scheduling is a strong plus - Demonstrable experience serving ML or LLM inference in production: batching, caching, model versioning, A/B rollout, latency budgeting - Rigour about observability and reliability — you have designed SLOs, run incidents, and shipped the fixes - Hard-won cost intuition: you have made a system materially cheaper without giving up quality Bonus Points Direct experience with OCR, IDP, or document AI systems — Textract, Document AI, Azure DI, or something you built yourself GPU inference stacks: vLLM, TensorRT-LLM, Triton, SGLang, Ray Serve Evaluation infrastructure for ML systems — golden sets, regression gates, human in-the-loop review loops Experience with BFSI, healthcare, or public-sector compliance and data-residency constraints On-prem or air-gapped deployment experience Note We are looking for people who can own the outcomes described here, not people who match every line of this specification. If this problem excites you and you believe you can do this work, we want to hear from you. Why Sarvam? Sarvam is a fast-moving, high talent-density team building full-stack AI for India, working on problems that push the frontiers ofAI with real population-scale impact. Work alongside researchers, engineers, builders, and business leaders who move fast and hold each other to a very high bar High ownership and high impact, from day one Everything we do is AI-first, from the way we build an

View the original posting ↗